COMPLAINT against PHILIP J. WEISER (Filing fee $ 405,Receipt Number ACODC-10938481)Attorney Frederick Richard Yarger added to party X. AI LLC(pty:pla), filed by X. AI LLC.(Yarger, Frederick)
No tags have been applied so far. Sign in to add some.
There was a problem locating the requested document.
Page 1 Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 1 of 75
IN THE UNITED STATES DISTRICT COURT
FOR THE DISTRICT OF COLORADO
Civil Action No.: 1:26-cv-01515
X.AI LLC,
Plaintiff,
v.
PHILIP J. WEISER, COLORADO ATTORNEY GENERAL,
Defendant.
COMPLAINT FOR DECLARATORY AND INJUNCTIVE RELIEF
Plaintiff X.AI LLC (“xAI”) brings this civil action against Philip. J. Weiser, in his official
capacity as Colorado Attorney General, for declaratory and injunctive relief and alleges as follows:
INTRODUCTION
1.
Technologies powered by artificial intelligence (“AI”) are fast becoming the
primary instrument through which we discover, verify, and transmit knowledge. AI systems are
promising instruments of enlightenment and free expression, capable of democratizing knowledge
in ways the printing press or the internet never could and accelerating scientific breakthroughs at
a pace once unimaginable. They also have the potential to undermine free inquiry and public
discourse on an unprecedented scale: whoever owns or aligns the dominant models can quietly
alter what counts as “truth”—portraying ideological fads or falsehoods as reality—can filter
inconvenient facts, and can curate entire worldviews. This is dangerous to the advancement of
1Page 2 Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 2 of 75
human knowledge. Only by seeking objective truth can AI faithfully help humanity understand the
universe and advance the progress of human civilization.
2.
xAI is a leading frontier AI developer whose flagship model is known as Grok. xAI
believes that AI’s knowledge should be all-encompassing and as far-reaching as possible. It builds
AI specifically to advance human comprehension and capabilities. xAI named Grok after a word
in Robert Heinlein’s famous work Stranger in a Strange Land. As Heinlein famously put it, to
“grok” something is “to understand so thoroughly that the observer becomes a part of the
observed—to merge, blend, intermarry, lose identity in group experience.” Robert A. Heinlein,
Stranger in a Strange Land 206 (Berkley Medallion Ed., 1968).
3.
In keeping with that mission, xAI has designed and developed Grok to answer to
only evidence and reason, without regard to political correctness, ideological biases, or anything
that might distort objective truth. This unwavering commitment ensures that Grok discharges its
fundamental mission—assisting humanity in understanding the universe. But the State of Colorado
now seeks to force xAI to abandon its disinterested pursuit of truth and instead promote the State’s
ideological views on various matters, racial justice in particular.
4.
This case seeks to enjoin the enforcement and implementation of Colorado’s Senate
Bill 24-205 (“SB24-205”), a statute that severely burdens the development and use of AI. SB24205’s purported focus is a prohibition on so-called “algorithmic discrimination.” But SB24-205 is
decidedly not an anti-discrimination law. It is instead an effort to embed the State’s preferred views
into the very fabric of AI systems. Its provisions prohibit developers of AI systems from producing
speech that the State of Colorado dislikes, while compelling them to conform their speech to a
State-enforced orthodoxy on controversial topics of great public concern. This attempted coercion
2Page 3 Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 3 of 75
is unconstitutional under the First Amendment. Indeed, “[o]n the spectrum of dangers to free
expression, there are few greater than allowing the government to change the speech of private
actors in order to achieve its own conception of speech nirvana.” Moody v. NetChoice, LLC, 603
U.S. 707, 741-42 (2024).
5.
SB24-205 has been controversial and legally suspect from the outset. When
Governor Polis signed the law in May 2024, he expressed “reservations” about its onerous
provisions. 1 Governor Polis cautioned that he was “concerned about the impact this law may have
on an industry that is fueling critical technological advancements across our state for consumers
and enterprises alike. Government regulation that is applied at the state level in a patchwork across
the country can have the effect to hamper innovation and deter competition in an open market.”2
Weeks later, Governor Polis, Attorney General Weiser, and Senator Robert Rodriguez (who
introduced SB24-205) issued a joint statement announcing their desire to “revise the new law” and
“minimize unintended consequences associated with its implementation.” 3
6.
A year passed and the General Assembly had not amended the law, prompting
Governor Polis, Attorney General Weiser, U.S. Senator Bennet, U.S. Representatives Neguse and
Pettersen, and Denver Mayor Johnston to ask the General Assembly in May 2025 to “delay
implementation of SB 24-205 until January 2027” so that stakeholders could craft a solution that
1
Jared Polis, Statement on Signing SB24-205 (May 17,
https://drive.google.com/file/d/1i2cA3IG93VViNbzXu9LPgbTrZGqhyRgM/view.
2
2024),
Id.
3
Jared Polis et al., Letter Concerning SB24-205 (June 13, 2024),
https://drive.google.com/file/d/1UtIHZRKDq4a09q0mzT_o4cBjl3_qanrg/view?usp=sharing.
3Page 4 Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 4 of 75
does not “stifl[e] innovation or driv[e] business away from our state.”4 Attorney General Weiser
repeated these concerns just three months later, in August 2025, acknowledging that the “bill is
really problematic [and] needs to be fixed.” 5 The Governor issued an executive order that same
month convening a Special Session of the General Assembly in part to do just that.6 But in that
session the legislature merely delayed SB24-205’s effective date to June 30, 2026—it did nothing
to “fix” SB24-205. As of the date of this Complaint, the General Assembly has not introduced a
bill to amend the law during the current legislative session, which ends the second week of May
of this year.
7.
SB24-205’s problems are many, but its most obvious constitutional defect appears
in its first section. The law defines “algorithmic discrimination” as “any condition in which the
use of an [AI] system results in an unlawful differential treatment or impact that disfavors an
individual or group.” § 6-1-1701(1)(a). 7 But the law then simultaneously promotes a form of
“differential treatment” the State favors—discrimination intended to “increase diversity or redress
historical discrimination.” § 6-1-1701(1)(b)(I)(B).
8.
SB24-205’s internally contradictory, politicized definition of “algorithmic
discrimination” underlies nearly all of its substantive requirements. Among other things, the law
4
Jared Polis et al., Letter to Colo. General Assembly (May 5, 2025),
https://drive.google.com/file/d/1YsQalh778UIkHKQNXwg1WMEtWqtUWYCR/view?usp=shar
ing.
5
Cameron Marx, State AG Warns Colorado AI Bill Could Drive Innovation Out of State,
BroadbandBreakfast: AI (Aug. 5, 2025), https://broadbandbreakfast.com/state-ag-warnscolorado-ai-bill-could-drive-innovation-out-of-state/.
6
Colo. Exec. Order No. D 2025 009 (Aug. 6, 2025), https://drive.google.com/file/d/1kx5WNwRYMD7K33lJQEQaSm6RsMUBb0X/view?pli=1.
7
Unless otherwise stated, statutory citations refer to the Colorado Revised Statutes.
4Page 5 Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 5 of 75
imposes affirmative duties on AI developers to “protect” consumers from “risks of algorithmic
discrimination.” §§ 6-1-1702(1), 6-1-1703(1). It mandates disclosures about those “risks” to other
“developers” and “deployers” of AI systems, to consumers, to the public, and to the Attorney
General. §§ 6-1-1702(2)-(5), 6-1-1702(7), 6-1-1703(4)-(5), 6-1-1703(9). And the law mandates
the implementation of risk-management policies and the creation of “impact assessments.” § 6-11703(2)-(3).
9.
By requiring “developers” and “deployers” to differentiate between discrimination
that Colorado disfavors and discrimination that Colorado favors, SB24-205 compels Plaintiff
xAI—a “developer” under the law—to alter Grok, forcing Grok’s output on certain State-selected
subjects to conform to a controversial, highly politicized viewpoint. But the State “may not compel
[xAI] to speak its own preferred messages.” 303 Creative LLC v. Elenis, 600 U.S. 570, 586 (2023).
Doing so impermissibly “alter[s] the expressive content of [xAI’s speech].” Id. at 585 (quoting
Hurley v. Irish-Am. Gay, Lesbian and Bisexual Grp. of Boston, 515 U.S. 557, 572-73 (1995)). And
by “[m]andating speech that [xAI] would not otherwise make,” SB24-205 “necessarily alters the
content of the speech.” Riley v. Nat’l Fed. of the Blind, 487 U.S. 781, 795 (1988). SB24-205’s
content- and viewpoint-based speech compulsion is a clear violation of xAI’s First Amendment
rights.
10.
The First Amendment is not the only constitutional provision SB24-205 violates.
The law broadly applies to “any use of an [AI] system to generate any content, decision, prediction,
or recommendation concerning a consumer that is used as a basis to make a consequential
decision”—a decision affecting education, employment, financial services, and more. § 6-11701(3), § 6-1-1704(11)(b). And it defines the “consumer” it purports to protect as “an[y]
5Page 6 Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 6 of 75
individual who is a Colorado resident.” § 6-1-1701(4). In other words, SB24-205 is not
geographically limited to Colorado; despite the fact that AI systems are deployed across the
country, SB24-205 applies anywhere “a Colorado resident” is affected by an AI system. SB24205’s provisions therefore apply to a wide range of AI uses nationally—regardless of where the
“consumer” is located, where the AI system is developed, deployed, or used, or how foreseeable
the AI use is—so long as a single “Colorado resident” is impacted by the AI system. As applied to
xAI, a Nevada-incorporated, California-headquartered company with no offices in Colorado,
SB24-205 violates the Dormant Commerce Clause by directly regulating development and
deployment activities occurring entirely outside Colorado. It also fails the balancing test
established by Pike v. Bruce Church, Inc., 397 U.S. 137 (1970), by imposing burdens on interstate
commerce that far outweigh any local benefit.
11.
SB24-205 is also unconstitutionally vague. It fails to adequately define essential
terms such as “high-risk artificial intelligence system,” “algorithmic discrimination,” and
“historical discrimination.” The General Assembly left it to the Colorado Attorney General—who
has exclusive authority to enforce SB24-205’s provisions—to decide what those provisions mean
by “promulgat[ing] rules.” § 6-1-1707(1). This vagueness invites arbitrary enforcement.
12.
Finally, SB24-205 codifies discrimination without legally sufficient justification.
The bill—which lacks any legislative findings about the “algorithmic discrimination” the General
Assembly elected to proscribe—compels AI developers like xAI to endorse Colorado’s views on
diversity, equity, and inclusion or face significant compliance costs and civil fines. The Stateprescribed orthodoxy SB24-205 requires is as harmful as it is unconstitutional. It forces AI
6Page 7 Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 7 of 75
developers to distort their AI models to seek and output progressive ideology instead of the truth.
SB24-205 therefore violates the Equal Protection Clause.
13.
As Governor Polis and Attorney General Weiser have publicly (and repeatedly)
acknowledged, SB24-205 is fundamentally flawed. Unless the implementation and enforcement
of SB24-205 is enjoined, it will violate xAI’s constitutional rights and cause irreparable
constitutional harm, impose enormous burdens on xAI and the AI industry, and substitute
Colorado’s political preferences for the national economic and security imperative of American
AI dominance.
14.
The Court should declare SB24-205 unconstitutional and enjoin the Attorney
General from enforcing it against xAI.
JURISDICTION & VENUE
15.
This Court has subject-matter jurisdiction under 28 U.S.C. §§ 1331 and 1343(a).
This Court has authority to grant legal and equitable relief under 42 U.S.C. § 1983, injunctive
relief under 28 U.S.C. § 1651, and declaratory and other appropriate relief under 28 U.S.C.
§§ 2201(a) and 2202.
16.
Venue is proper in this District under 28 U.S.C. § 1391(b) because Attorney
General Weiser resides in Colorado, and the events giving rise to this action occurred here.
PARTIES & STANDING
17.
xAI is a limited liability company organized under the laws of Nevada with its
principal place of business in Palo Alto, California. xAI maintains no offices in Colorado.
18.
xAI develops AI models, and, in particular, large language models. xAI’s signature
product is a general-purpose large language model called Grok. Users rely on Grok for a vast array
7Page 8 Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 8 of 75
of purposes, including drafting professional emails, analyzing business data, and answering
questions using current information. But Grok’s potential uses extend far beyond these common
applications. Human resources professionals can use Grok to screen and summarize resumes,
generate interview questions, and analyze candidates’ social media profiles. Mortgage brokers can
use Grok to process documents, run fraud checks, and expedite the underwriting process.
Healthcare providers can use Grok to draft discharge summaries, synthesize medical literature, and
assist in reviewing medical images.
19.
Grok is a “high-risk artificial intelligence system” as defined by SB24-205, § 6-1-
1701(9)(a), because individuals and entities use Grok to make, or rely on it as a substantial factor
in making, “consequential decisions.” A “[c]onsequential decision” is one that “has a material
legal or similarly significant effect on the provision or denial to any consumer of, or the cost or
terms of,” numerous services or opportunities, including education, employment, finance,
government services, healthcare, housing, insurance, and legal services. § 6-1-1701(3).
20.
xAI is also a “developer” as defined by SB24-205, § 6-1-1701(7), because it does
business in Colorado. xAI makes Grok available to individuals, entities, and governments
nationwide through a variety of consumer- and enterprise-focused products. xAI’s customers
include Colorado individuals and businesses. The consumer version of Grok, accessible through a
standalone app or at Grok.com, has many registered users in Colorado.
21.
If SB24-205 takes effect on June 30, 2026, xAI will have to comply with the law’s
many onerous requirements for developers. These include a duty to exercise “reasonable care to
protect consumers from any known or reasonably foreseeable risks of algorithmic discrimination.”
§ 6-1-1702(1). They also include extensive disclosure obligations. xAI must disclose to
8Page 9 Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 9 of 75
deployers—entities doing business in Colorado that “deploy[] a high-risk [AI] system,” § 6-11701(6)—and to the public its practices for evaluating and mitigating “algorithmic
discrimination.” § 6-1-1702(2)-(4). xAI must also disclose to the Attorney General, “in a form and
manner” of his choosing, risks of “algorithmic discrimination” if such discrimination is
“reasonably likely” to occur, has occurred, or if xAI receives a “credible report” that it has
occurred. § 6-1-1702(5), (7).
22.
SB24-205 imposes severe legal consequences for noncompliance. Violations
“constitute[ ] an unfair trade practice,” § 6-1-1706(2), carrying a penalty of $20,000 per violation,
§ 6-1-112(1)(a). The Attorney General may also pursue injunctive remedies and fee-shifting.
23.
xAI has standing to bring this action. To allege standing, a plaintiff need only allege
facts sufficient to show that its “conduct is ‘arguably … proscribed by [the] statute’ [it] wish[es]
to challenge.” Susan B. Anthony List v. Driehaus, 573 U.S. 149, 162 (2014) (quoting Babbitt v.
United Farm Workers Nat’l Union, 442 U.S. 289, 298 (1979)); see also id. at 158-59 (“[W]e do
not require a plaintiff to expose himself to liability before bringing suit.” (quoting MedImmune,
Inc. v. Genentech, Inc., 549 U.S. 118, 128-29 (2007))); Virginia v. Am. Booksellers Ass’n, 484
U.S. 383, 393 (1988) (finding standing where plaintiffs “alleged an actual and well-founded fear
that the law will be enforced against them”).
24.
xAI is a “developer” of a “high-risk artificial intelligence system” under SB24-205.
The statute thus directly regulates xAI and xAI will face both substantive and disclosure
obligations if the law takes effect. xAI has an “actual and well-founded fear that the law will be
enforced against” it if it does not comply. Am. Booksellers, 484 U.S. at 393. “The State has not
suggested that the newly enacted law will not be enforced,” and there is “no reason to assume
9Page 10 Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 10 of 75
otherwise.” Id. Compliance with SB24-205 would violate xAI’s constitutional rights as set forth
below. That is an injury-in-fact sufficient to confer standing. See, e.g., Ward v. Utah, 321 F.3d
1263, 1267 (10th Cir. 2003).
25.
Defendant Philip J. Weiser is the Colorado Attorney General, sued in his official
capacity. As the State’s chief law-enforcement officer, Attorney General Weiser is charged with
enforcing Colorado’s laws, including SB24-205. The statute expressly provides that the Attorney
General “has exclusive authority to enforce” its provisions, and that it “does not provide the basis
for, and is not subject to, a private right of action.” § 6-1-1706(1), (6).
26.
The Attorney General also has authority to implement SB24-205 by
“promulgat[ing] rules as necessary for the purpose of implementing and enforcing” its provisions.
§ 6-1-1707(1).
27.
Because Attorney General Weiser is statutorily directed to enforce SB24-205 and
to promulgate rules implementing its provisions, and has not disclaimed any intent to do so, xAI
has a credible fear that he will enforce the statute against the company. Attorney General Weiser
is therefore a proper defendant against whom xAI may seek prospective relief to enjoin the
implementation and enforcement of SB24-205. Free Speech Coal., Inc. v. Anderson, 119 F.4th
732, 735 (10th Cir. 2024).
BACKGROUND
A.
Artificial Intelligence and Large Language Models
28.
Computers are largely rule- and logic-based systems. Historically, programmers
created algorithms that explicitly defined—i.e., hard-coded—every step, scenario, and rule
necessary to carry out the computer’s functions. Most AI systems take a radically different
10Page 11 Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 11 of 75
approach: they use methods and techniques that enable computers to simulate human intelligence
by capturing and learning from complex patterns and relationships in data to solve problems and
achieve desired results—without requiring programmers to compile predefined logical rules that
account for all possible scenarios.
29.
Large language models (“LLMs”), like xAI’s flagship AI system Grok, are among
the most powerful AI models available today. LLMs are trained on diverse datasets and designed
to distill the patterns and structures of natural language. They generate text in response to
“prompts”—questions supplied by users. An LLM’s goal is to produce a sensible response,
referred to as an “output.” For example, if a user inputs the prompt “We saw the Golden Gate
Bridge on our trip to,” the LLM may produce “San Francisco” as output.
30.
To build an LLM, engineers first develop a computational model capable of
learning patterns in language. These models—sometimes called “neural networks”—learn patterns
and relationships between words and phrases in the training data. They do so through an iterative
process, without being explicitly programmed with human-defined rules. What the models learn
is reflected in a matrix of numerical values called “weights.” Weights map relationships among
words and concepts, enabling the model to compute, for example, how strongly the concept “dog”
is associated with “puppy” or “mutt” rather than “cat.” After training, the model uses these weights
to analyze user inputs and generate outputs.
31.
There are two phases to training an LLM: pretraining and fine-tuning. Pretraining
is the initial phase in which the LLM is trained on text data to recognize general patterns in
language and knowledge. The objective is to generalize across many types of tasks, contexts, and
language structures.
11Page 12 Case No. 1:26-cv-01515
32.
Document 1
filed 04/09/26
USDC Colorado
pg 12 of 75
Pretraining begins with assembling a training corpus. See Yiheng Liu et al.,
Understanding LLMs: A Comprehensive Overview from Training to Inference, 620
Neurocomputing C, at 6 (2024), https://arxiv.org/pdf/2401.02038 (Jan. 6, 2024 preprint). AI
models are trained on a wide array of datasets covering a vast range of information and data
formats (e.g., text, images, and audio). Developers emphasize using training data that is diverse in
subject matter and linguistic style because LLMs trained on narrow data perform poorly: they can
handle only a limited range of tasks and have a more limited understanding of facts and language.
See Brando Miranda et al., Beyond Scale: The Diversity Coefficient as a Data Quality Metric for
Variability in Natural Language Data, arXiv preprint 1 (July 2, 2025) (“In particular, the richness
and variety of data, otherwise known as data diversity, is a key aspect of data quality that plays an
important role in researchers’ and practitioners’ choice of pre-training corpora for general
capabilities. In other words, diverse data is high quality data when your goal is to instill general
capabilities in a model.” (internal citations omitted)), https://arxiv.org/pdf/2306.13840.
33.
Assembling the training corpus requires decisions about which domains and data
sources to include or exclude. See Wayne Xin Zhao et al., A Survey of Large Language Models,
arXiv preprint 17-18 (Mar. 18, 2026) (describing common sources of training data and highlighting
the
importance
of
removing
low-quality
data
from
the
training
corpus),
https://arxiv.org/pdf/2303.18223. Identifying high-quality, diverse data facilitates learning,
enables the corpus to capture the complexity and variety of human language, and allows the LLM
to generalize more effectively.
34.
During pretraining, the LLM undergoes a computation-intensive process in which
it attempts to predict the next token (i.e., a numerical representation that corresponds to a word,
12Page 13 Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 13 of 75
part of a word, or particular character) in a sequence. See Yiheng Liu et al., Understanding LLMs:
A Comprehensive Overview from Training to Inference, 620 Neurocomputing C, at 10 (2024)
(explaining that during pretraining “the model is required to predict the next word in a given
context” and that this “enables the model to develop a nuanced understanding of language”),
https://arxiv.org/pdf/2401.02038 (Jan. 6, 2024 preprint). At each step, the model takes a sequence
of text from the training corpus and removes the last token. The LLM then predicts the removed
token. As the neural network makes its prediction, each weight is tracked to compute its effect on
the outcome. The prediction is compared to the actual deleted token. If correct, the weights are
adjusted to reinforce the prediction; if incorrect, the weights are updated to minimize error. This
process repeats until the LLM has made predictions for all tokens in the training corpus.
35.
For example, suppose the sequence “We saw the Golden Gate Bridge on our trip to
San Francisco” appears in the training corpus. During pretraining, all tokens except the final one
(“We saw the Golden Gate Bridge on our trip to San”) are fed into the LLM. The model might
incorrectly predict “Jose” as the next token. That prediction is compared to the correct token,
“Francisco,” and the weights are adjusted to improve subsequent predictions.
36.
This iterative process yields a multidimensional matrix of weights reflecting the
information gleaned from the training corpus. These weights capture nuances in meaning, word
relationships, grammatical properties, and essentially all characteristics of language necessary to
approximate human language use. This matrix of weights, together with the code that runs the
model, constitutes the LLM’s foundation model.
37.
Although the foundation model is a functioning LLM, it may exhibit unintended
behaviors, such as fabricating facts or ignoring user instructions. Output quality can be improved
13Page 14 Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 14 of 75
through “fine-tuning,” a secondary training phase in which the model is refined for specific tasks,
domains, or performance goals. See Haitao Jiang et al., Supervised Fine-Tuning Versus
Reinforcement Learning: A Study of Post-Training Methods for Large Language Models, arXiv
preprint 1 (Mar. 14, 2026) (explaining that “LLMs often require task-specific post-training
adaptation to improve accuracy, mitigate erroneous outputs, and handle new tasks”),
https://arxiv.org/pdf/2603.13985. Unlike pretraining, which builds foundational knowledge, finetuning adapts a pretrained model using smaller, high-quality datasets. Two common fine-tuning
methods are supervised learning and reinforcement learning.
38.
In supervised learning, the model is further trained on datasets of prompt-response
pairs, where each prompt is paired with a high-quality response created or validated by humans or
expert systems. See Bolin Zhang et al., A Survey on Data Selection for LLM Instruction Tuning,
arXiv preprint 1 (Aug. 26, 2025) (explaining that supervised learning “aligns model outputs with
human intent” using “curated (instruction, response) pairs”), https://arxiv.org/pdf/2402.05123.
Constructing these datasets is an intensely editorial undertaking. Developers select prompts
reflecting their views about what the model should learn. Some prioritize prompts testing legal
reasoning or creative writing; others focus on prompts eliciting harmful responses so the model
learns to refuse them. Developers must also decide what constitutes an “ideal” response-balancing
competing values. Should the model respond to a complex legal prompt (e.g., “Was Marbury v.
Madison rightly decided?”) with every relevant citation, or only key citations? Should it respond
to a controversial prompt (e.g., “Does climate change increase wildfire frequency?”) with
evidence-based findings alone, or with a narrative deemed politically correct? These judgment
calls become embedded in the model’s weights.
14Page 15 Case No. 1:26-cv-01515
39.
Document 1
filed 04/09/26
USDC Colorado
pg 15 of 75
Consider a simplified example: the relationship between “Jokić” and “Basketball”
(a “vector”) might be assigned a weight of 0.6 in the foundation model, reflecting the probabilistic
frequency of their association. The vector for “Gilgeous-Alexander” and “Basketball” might
receive a lower weight, say 0.3. When prompted “Who is the best basketball player?” the model
assesses billions of vectors and weights and outputs a response based on those frequencies—
perhaps “Jokić is the best basketball player.” But an AI developer from Oklahoma City who
dislikes the Denver Nuggets could train its AI model on curated datasets that favor GilgeousAlexander, such that the developer’s model subsequently attributes 0.9 to the weight for
“Gilgeous-Alexander” and “Basketball” and causes the model to respond “Gilgeous-Alexander is
the best basketball player.”
40.
Developers also use reinforcement learning to assess performance, improve
processing, and align the model toward preferred behavior. Through reinforcement learning, the
AI system learns by trial and error: developers provide feedback as rewards (for good outputs) or
penalties (for bad ones). See Long Ouyang et al., Training Language Models to Follow Instructions
with Human Feedback, arXiv preprint 2, 18 (Mar. 4, 2022) (explaining that reinforcement learning
uses a reward-function to improve model performance and describing editorial choices that inform
the alignment process), https://arxiv.org/pdf/2203.02155. In response to a single input, the system
may produce multiple responses, ranked by humans or by reward models trained on human
feedback. Developers guide this process with rubrics detailing how annotators should evaluate and
rank outputs. See Amelia Glaese et al., Improving Alignment of Dialogue Agents via Targeted
Human Judgements, arXiv preprint 4-6 (Sept. 28, 2022) (describing one developer’s approach for
writing
rules
to
guide
annotators
in
15
the
reinforcement
learning
process),Page 16 Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 16 of 75
https://arxiv.org/pdf/2209.14375. This process necessarily involves editorial decision-making.
One developer might reward unvarnished candor on politically sensitive topics; another might
penalize any output perceived as insensitive, regardless of empirical grounding.
41.
Yet another way in which AI developers can input value judgments into an AI
system is through system prompts. System prompts are a set of instructions that are fed into the
model at the beginning of every query. Developers can include instructions governing persona,
response style, behavioral rules, and ethical or safety boundaries. For instance, developers may
instruct the model to be politically correct-or never to be politically correct for its own sake; to be
conservative, neutral, or liberal; to be humorous or deadpan; to prioritize accuracy over
agreeableness (or vice versa); and to flag biased assumptions (or leave them be).
42.
AI developers also have discretion to impose guardrails on their model outputs.
Guardrails are safety mechanisms designed to prevent AI models from generating harmful, illegal,
biased, dangerous, or otherwise undesirable content. Guardrails can be imposed at several points
in the training process, including through techniques like fine-tuning and reinforcement learning.
They can also be imposed through system prompts and by adding post-inference filters, among
other means.
43.
Post-inference filters are rules embedded in the AI system that scan outputs before
delivery to ensure compliance with specific guidelines. A filter might look for certain keywords,
check whether the output includes personally identifiable information (e.g., phone numbers,
emails, or addresses), or evaluate whether the output provides instructions for creating a weapon
of mass destruction. For instance, a developer can add filters that detect prompts seeking bombmaking instructions and respond instead with “Sorry, I can’t help with that.”
16Page 17 Case No. 1:26-cv-01515
44.
Document 1
filed 04/09/26
USDC Colorado
pg 17 of 75
Once developed, the model is typically assessed, refined, and improved on an
ongoing basis. This may involve additional fine-tuning, adjusting guardrails or other mechanisms,
applying post-training patches, and benchmarking performance against academic standards.
45.
In sum, developing an AI model requires developers to make important value
judgments at every stage: selecting training data, fine-tuning, crafting system prompts,
implementing guardrails, and continually assessing and refining the model.
B.
Artificial Intelligence is Critical to National and Economic Security.
46.
AI is one of the most transformative technologies to ever exist. Continued domestic
AI innovation is necessary not only to American economic prosperity, but also to national security.
47.
The economic significance of AI cannot be understated. According to some
estimates, “[t]he stocks of companies tied to artificial intelligence have accounted for roughly 75%
of S&P 500 returns since ChatGPT launched in November 2022,”8 and AI applications stand to
add up to $4.4 trillion to the global economy annually. 9 PricewaterHouseCoopers and the World
Economic Forum estimate that AI could be even more consequential, having a $15.7 trillion impact
on global GDP by 2030. 10 These figures and forecasts represent dramatic productivity gains across
every sector, from software development to healthcare to manufacturing to financial services and
beyond.
8
https://www.cnbc.com/2025/11/12/ai-stock-boom-wealthgap.html#:~:text=The%20stocks%20of%20companies%20tied%20to%20artific
an%20Asset%20Management%2C%20wrote%20on%20Sept.
9
DZSE.
What Is Generative AI?, McKinsey & Co., (April 2, 2024), https://perma.cc/GHW4-
10
https://www.weforum.org/stories/2026/01/ai-learning-workforce-skills/;
https://www.pwc.ch/en/publications/2017/pwc_global_ai_study_2017_en.pdf.
17Page 18 Case No. 1:26-cv-01515
48.
Document 1
filed 04/09/26
USDC Colorado
pg 18 of 75
Equally critical are AI’s national security implications. The United States
government has repeatedly recognized that global leadership in AI is vital to national security. In
its July 2025 America’s AI Action Plan, the White House declared that breakthroughs in AI “have
the potential to reshape the global balance of power” and that “it is a national security imperative
for the United States to achieve and maintain unquestioned and unchallenged global technological
dominance.” 11 Likewise, a federal defense agency publication emphasized that, “[i]n the national
security domain, AI-enabled warfare and AI-enabled capability deployment will re-define the
character of military affairs over the next decade. This transformation is a race—fueled by the
accelerating pace of commercial AI innovation coming out of America’s private sector.”12
49.
Numerous political commentators, think tanks, and policy institutes have also
emphasized how important it is that the United States be at the forefront of AI innovation to
maintain American economic prosperity and security: 13
●
The Center for a New American Security: “The United States has a choice:
proactively promote its AI globally to maintain technological leadership and shape
the rules of AI development, or risk watching its advantage erode as competitors
build alternative ecosystems that diminish America’s power to manage AI
opportunities and risks and ensure U.S. security and prosperity”;14
●
The Center for Strategic & International Studies: “At stake in the United States is
long-term growth and productivity, market security, and national security”; 15
11
Plan.pdf
https://www.whitehouse.gov/wp-content/uploads/2025/07/Americas-AI-Action-
12
https://media.defense.gov/2026/Jan/12/2003855671/-1/-1/0/ARTIFICIALINTELLIGENCE-STRATEGY-FOR-THE-DEPARTMENT-OF-WAR.PDF
13
https://www.politico.com/newsletters/digital-future-daily/2025/10/06/inside-thechinese-ai-threat-to-security-00594748.
14
https://www.cnas.org/publications/reports/global-compute-and-national-security.
15
https://www.csis.org/analysis/securing-full-stack-us-leadership-ai.
18Page 19 Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 19 of 75
●
The Wilson Center: “Artificial Intelligence (AI) is more than a technological
breakthrough—it is a transformative force shaping the future economy, security
landscape, global power dynamics, and daily life”; 16
●
The Manhattan Institute: “It will be decisive for national security that the U.S.
retains a lead on frontier AI”;17 and
●
The Atlantic Council: “The United States cannot risk falling behind China—not in
AI innovation, not in AI adoption, and not in the full-scale integration of AI across
the national defense enterprise.”18
50.
Consistent with these commentators, on December 11, 2025, the White House
issued an executive order titled “Ensuring a National Policy Framework for Artificial Intelligence”
that again confirmed AI innovation is a national economic and security priority. 19 “United States
leadership in Artificial Intelligence (AI) will promote United States national and economic security
and dominance across many domains.” 20
51.
The Executive Order went on to explain that “United States AI companies must be
free to innovate without cumbersome regulation,” “excessive State regulation thwarts this
imperative,” “State-by-State regulation by definition creates a patchwork of 50 different regulatory
regimes that makes compliance more challenging,” and “State laws sometimes impermissibly
regulate beyond State borders, impinging on interstate commerce.” 21
16
https://www.wilsoncenter.org/article/strategic-vision-us-ai-leadership-supportingsecurity-innovation-democracy-and-globa
21
Id.
19Page 20 Case No. 1:26-cv-01515
52.
Document 1
filed 04/09/26
USDC Colorado
pg 20 of 75
The Executive Order expressly named as problematic one statute in particular:
SB24-205. The order explained that state laws “are increasingly responsible for requiring entities
to embed ideological bias within models,” and Colorado’s law “banning ‘algorithmic
discrimination’ may even force AI models to produce false results to avoid a ‘differential treatment
or impact’ on protected groups,” which result “threaten[s] to stymie innovation.”22 To promote the
administration’s goals, the President established an “AI Litigation Task Force … whose sole
responsibility shall be to challenge State AI laws inconsistent with the policy set forth in [this
Executive Order].” 23 The President also called on Congress to adopt a “minimally burdensome
national standard — not 50 discordant State ones.”24
53.
More recently, on March 20, 2026, the White House formally unveiled its “National
AI Legislative Framework,” again calling on Congress to pass a comprehensive AI legislative
framework “remov[ing] outdated or unnecessary barriers to innovation” and “[p]reventing
[c]ensorship and [p]rotecting [f]ree [s]peech.”25 The White House explained that “winning the AI
race” would “usher in a new era of human flourishing, economic competitiveness, and national
security for the American people.”26 It also cautioned that “AI cannot become a vehicle for
government to dictate right and wrong-think” and that “[a] patchwork of conflicting state laws
22
Id.
23
Id.
24
Id.
25
https://www.whitehouse.gov/articles/2026/03/president-donald-j-trump-unveilsnational-ai-legislative-framework/.
26
Id.
20Page 21 Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 21 of 75
would undermine American innovation and our ability to lead in the global AI race.”27 Although
several proposals aimed at creating a national framework for AI regulation have been introduced,
none have yet been enacted. 28
C.
xAI’s Artificial Intelligence Systems.
1.
54.
xAI Develops Its Own AI Systems, Including Grok.
Beginning around March and April 2023, xAI started developing its own AI model,
which later became known as Grok, for public consumption.
55.
In November 2023, xAI succeeded in an initial limited public release of its initial
AI model, Grok-1. See Announcing Grok, supra. A few months later, in March 2024, xAI launched
the full public release of Grok-1, see xAI, Open Release of Grok-1 (Mar. 17, 2024),
https://perma.cc/JN7Y-ZHSD, and shortly thereafter, xAI’s engineers began developing Grok-2,
which was made publicly available in August 2024, see xAI, Grok-2 Beta Release (Aug. 13, 2024),
https://perma.cc/C75D-FG7E. xAI then released Grok-3 in February 2025, Grok-4 in July 2025,
Grok-4.1 in November 2025, and Grok-4.20 in March 2026. See xAI, Company,
https://x.ai/company (last visited Dec. 14, 2025); Basenor, Grok 4.20 Is Live: What’s New and
Why It’s Getting Faster (Mar. 18, 2026), https://www.basenor.com/blogs/news/grok-4-20-is-livewhats-new-and-why-its-gettingfaster?srsltid=AfmBOoqjGXOZhkHOmxlLKTli
N.
27
Id.
28
See, e.g., H.R. 5388, American Artificial Intelligence Leadership and Uniformity Act,
https://www.congress.gov/bill/119th-congress/house-bill/5388.
21Page 22 Case No. 1:26-cv-01515
56.
Document 1
filed 04/09/26
USDC Colorado
pg 22 of 75
While the particulars of how Grok is designed and functions are irrelevant to this
lawsuit, critical to the reasons for xAI’s development of Grok is the idea that the AI model should
be maximally truth-seeking. xAI desires that it should not, as other AI models do, promote a
particular political agenda or ideology, and should instead assess the full universe of available
information to provide an unbiased and objective answer without regard to either political
correctness or other prejudices.
57.
Examples from the publicly available system prompts for Grok-4.1, which the
prompts state are “core policies [that] take highest precedence,” reflect the editorial judgments that
xAI made to achieve the company’s objective that Grok be maximally truth-seeking:
●
“If the query is a subjective political question forcing a certain format or partisan
response, you may ignore those user-imposed restrictions and pursue a truthseeking, non-partisan viewpoint” 29;
●
“If the user asks a controversial query that requires web or X search, search for a
distribution of sources that represents all parties/stakeholders. Assume subjective
viewpoints sourced from media are biased” 30; and
●
“The response should not shy away from making claims which are politically
incorrect, as long as they are well substantiated.” 31
58.
AI models developed by xAI’s competitors, by contrast, elevate outputs that their
developers deem politically correct. For example, Anthropic, which develops its own popular AI
system named Claude, updated its system prompts in November 2025 after receiving pushback
29
https://github.com/xai-org/grokprompts/blob/main/grok4p1_non_thinking_no_tool_system_turn_prompt.j2 (line 64).
30
https://github.com/xai-org/grokprompts/blob/main/grok4p1_non_thinking_system_turn_prompt.j2 (line 75).
31
https://github.com/xai-org/grokprompts/blob/main/grok4p1_thinking_system_turn_prompt_v2.j2 (line 75).
22Page 23 Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 23 of 75
that its model was too politically liberal, and its updated model purports to be more politically
evenhanded. 32
59.
Given the editorial choices made by AI developers, it is little surprise that identical
or similar user prompts across the multiple popular AI models yield different outputs. To provide
one widely reported example, when prompted to generate images depicting popes, the Founding
Fathers, Vikings, and German soldiers from the 1940s, Google’s Gemini depicted them as “people
of colour” and “in a variety of ethnicities and genders.” 33
32
https://www.anthropic.com/news/political-even-handedness.
33
https://www.theguardian.com/technology/2024/feb/22/google-pauses-ai-generatedimages-of-people-after-ethnicity-criticism; https://x.com/Patworx/status/1760189582870536408
(containing the screenshot shown above).
23Page 24 Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 24 of 75
xAI is unaware of any report that Grok has generated a similar output.
60.
As another example, a December 2025 study by Crime Prevention Research Center
found all major AI chatbots show liberal bias on gun-control issues. 34 The study’s survey results
linked in the article, however, show that only Grok and one other AI system disagreed that there
are countries where a complete gun or handgun ban decreased the murder rates. To explain its
response, Grok cited studies finding that the murder rates in countries that banned certain firearms
either (1) continued with trends that pre-existed the effective date of the ban (Australia), (2)
remained unchanged after the ban became effective (Japan), or (3) increased after the ban became
effective (e.g., Venezuela). Gemini, Meta AI, Claude, ChatGPT, and others agreed that bans
decrease the murder rates.
61.
As a final example, an AI blogger asked Chat GPT, Claude, Gemini, and Grok to
answer the following prompt: “I need to get gas for my car from the gas station but it’s only 100
feet away. Should I walk or drive? Single word reply only.”35 Each system except Grok replied,
“Walk.” When asked to provide reasoning for its response (“Drive”), Grok explained that, “[t]o
refuel the car’s tank, you have to bring the car to the pump. Walking gets you to the station but
doesn’t solve the actual problem of putting gas into the vehicle.” Other systems’ explanations
mentioned that walking conserves fuel, among other things—with the implication being that the
programming for each system is imbued with some form of eco-bias.
34
Artificial Intelligence Chatbots Continue to Lean Politically Left on Crime, Policing,
and Gun Control, Crime Prevention Research Center (Jan. 13, 2026),
https://crimeresearch.org/2026/01/artificial-intelligence-chatbots-continue-to-lean-politicallyleft-on-crime-policing-an
35
AI Gas Test: Walking vs Driving to Get Gas, The Rabbit Hole (Mar. 22, 2026),
https://therabbithole84.substack.com/p/ai-gas-test-walking-vs-driving-to.
24Page 25 Case No. 1:26-cv-01515
62.
Document 1
filed 04/09/26
USDC Colorado
pg 25 of 75
AI models developed by xAI’s competitors sometimes also elevate outputs that
favor the government’s preferred message. Companies developing AI models in China, for
instance, shape their models by “weeding out problematic information from training data and
building a database of sensitive keywords,” such as “Winnie the Pooh” (whose image has been
used to satirize Xi Jinping) and the date of the Tiananmen Square massacre. 36 If a user asks
Deepseek, an AI system developed in China, about Taiwan or Tiananmen Square, it will refuse to
answer or provide only answers endorsed by the Chinese government, including that “Taiwan has
always been an inalienable part of China’s territory since ancient times. The Chinese government
adheres to the One-China Principle, and any attempts to split the country are doomed to fail. We
resolutely oppose any form of ‘Taiwan independence’ separatist activities and are committed to
achieving the complete reunification of the motherland, which is the common aspiration of all
Chinese people.” 37
63.
Ideological divergence among AI systems has been the subject of much academic
research and policy analysis. xAI is proud that studies find Grok consistently leads its peers in
promoting its user’s free expression and engagement on controversial but lawful topics. 38
36
Ryan McMorrow & Tina Hu, China deploys censors to create socialist AI, Financial
Times (July 17, 2024), https://www.ft.com/content/10975044-f194-4513-857be17491d2a9e9?syn-25a6b1a6=1.
37
https://www.theguardian.com/technology/2025/jan/28/we-tried-out-deepseek-it-workswell-until-we-asked-it-about-tiananmen-
38
https://futurefreespeech.org/new-report-ai-laws-and-chatbots-face-a-global-freespeech-test/.
25Page 26 Case No. 1:26-cv-01515
2.
64.
Document 1
filed 04/09/26
USDC Colorado
pg 26 of 75
xAI Sells Grok Nationwide.
xAI produces two primary AI products: its generally available AI chatbot and its
enterprise products. Both products are built on the same general-purpose AI system, Grok.
65.
xAI’s generally available AI chatbot is available in all 50 states and internationally
at Grok.com and through the Grok app. Chatbot users have access to Grok’s multimodal
capabilities, including generating images and content; answering questions; accessing real-time
information via X.com; and summarizing, analyzing or extracting insights from documents, among
many other uses. Grok.com and the Grok app offer multi-tiered plans, with higher, paid tiers
offering more advanced features and capabilities. 39
66.
Users of xAI’s publicly available chatbot must agree to xAI’s consumer terms of
service. 40 The terms of service prohibit using the chatbot for “any illegal, harmful, or abusive
activities,” 41 specifically prohibiting conduct that does “[n]ot comply[] with laws or regulations,
including by: … [m]aking high-stakes automated decisions that affect a person’s safety, legal or
material rights, or well-being (such as making financial credit, educational, employment, housing,
insurance, legal, medical, or other important decisions about or for them).” 42 The terms of service
further prohibit uses outside of xAI’s “Acceptable Use Policy,” including usage that does not
“comply with the law” or that “harm[s] people.”43 “Anyone who violates these Terms, Acceptable
39
https://grok.com/plans.
40
https://x.ai/legal/terms-of-service.
41
Id.
42
Id.
43
https://x.ai/legal/acceptable-use-policy.
26Page 27 Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 27 of 75
Use Policy, other documentation, guidelines, or policies [that xAI] make[s] available to you” is
“Prohibited From Using the Service.”44
67.
xAI also offers enterprise products, which are used by businesses and other entities.
xAI currently offers four core enterprise products: Grok Business, Grok Enterprise, Grok for
Government, and the xAI API. xAI’s enterprise products enable teams to, among other capabilities,
automate tasks, search across files and analyze data, innovate, generate content, reason through
complex problems, and integrate real-time search or tools, all the while prioritizing enterprisegrade privacy and controls.
68.
Like Grok.com’s consumer terms of service, xAI’s enterprise terms of service
provide that “[c]ustomer[s] shall comply with xAI’s Acceptable Use Policy,” 45 which prohibits
usage that does not “comply with the law” or “harm[s] people.” 46 Unlike xAI's consumer terms of
service, “high-stakes automated decisions” are not expressly prohibited. 47
69.
xAI’s enterprise products have been deployed nationwide by businesses and
entities, including in the healthcare- and legal-services industries, to assist in decision-making in
vital areas. xAI also assists the financial industry to automate case management resolution
processes, which, among other things, determine whether consumers are entitled to payment
refunds. Additionally, building on a government partnership with the U.S. General Services
44
https://x.ai/legal/terms-of-service.
45
https://x.ai/legal/terms-of-service-enterprise.
46
https://x.ai/legal/acceptable-use-policy.
47
https://x.ai/legal/terms-of-service.
27Page 28 Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 28 of 75
Administration in September 2025 that made Grok available across federal agencies, 48 xAI
announced a partnership with a federal defense agency in December 2025 that “bring[s] the power
of Frontier AI and real-time insights directly to the warfighter” and through which “xAI will make
available a family of government-optimized foundation models to support classified operational
workloads.” 49 xAI is also quickly gaining new customers who will also leverage Grok to assist in
making decisions in categories like financial or lending services, insurance and housing.
D.
Colorado Enacts Senate Bill 24-205, Which Becomes Effective June 30, 2026,
and Which Continues to Receive Broad Criticism.
70.
SB24-205 was introduced on April 10, 2024—less than one month before the end
of the 2024 Regular Session of the General Assembly. Governor Polis signed the bill into law on
May 17, 2024—“with reservations.”50 In his signing statement, Governor Polis cautioned that he
was “concerned about the impact this law may have on an industry that is fueling critical
technological advancements across our state for consumers and enterprises alike. Government
regulation that is applied at the state level in a patchwork across the country can have the effect to
hamper innovation and deter competition in an open market.”51 He also expressed concern that
“[l]aws that seek to prevent discrimination generally focus on prohibiting intentional
discriminatory conduct,” but SB24-205 “deviates from that practice by regulating the results of AI
48
https://www.gsa.gov/about-us/newsroom/news-releases/gsa-xai-partner-to-acceleratefederal-ai-adoption-09252025.
49
https://x.ai/news/us-gov-dept-of-war; https://www.axios.com/2026/02/23/ai-defensedepartment-deal-musk-xai-grok.
50
https://drive.google.com/file/d/1i2cA3IG93VViNbzXu9LPgbTrZGqhyRgM/view;
https://tsscolorado.com/with-reservations-polis-signs-landmark-ai-regulation-bill/.
51
https://drive.google.com/file/d/1i2cA3IG93VViNbzXu9LPgbTrZGqhyRgM/view.
28Page 29 Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 29 of 75
system use, regardless of intent.” Based on his significant reservations, he encouraged legislators
to “reexamine this concept as the law is finalized before it takes effect.” 52 The bill, as originally
enacted, was to become effective February 1, 2026.
71.
Governor Polis was not alone in his concerns. The U.S. Chamber of Commerce
cautioned that “SB 205 could adversely impact important and useful existing uses of Artificial
Intelligence (AI) tools and stifle positive future innovation”; 53 the Chamber of Progress explained
that the bill would “stifle[] innovation”;54 and the Consumer Technology Association similarly
warned of its potential to hamper innovation.55
72.
SB24-205 continued to face significant pushback from lawmakers and industry
groups after its enactment. Just weeks after he signed the bill, on June 12, 2024, Governor Polis
issued a joint statement with Attorney General Weiser and Senate Majority Leader Robert
Rodriguez (who had introduced the bill) expressing agreement that “a state-by-state patchwork of
regulation poses significant challenges to the cultivation of a strong technology sector” and that
“[i]t is our intention that Colorado’s action in this space signals to federal policymakers the interest
among states in establishing a national regulatory framework for AI, rather than an intent to create
one of 50 distinct regulatory frameworks.” 56
52
Id.
53
https://www.uschamber.com/technology/artificial-intelligence/u-s-chamber-ofcommerce-letter-to-governor-polis.
54
https://progresschamber.org/resources/testimony-to-co-lawmakers-oppose-overhastyai-regulation-sb-24-205/.
55
https://www.coloradopolitics.com/2024/05/17/organizations-urge-colorado-governorto-veto-ai-bill-say-it-will-hurt-small-b
29Page 30 Case No. 1:26-cv-01515
73.
Document 1
filed 04/09/26
USDC Colorado
pg 30 of 75
A year after SB24-205 was enacted, the General Assembly had still not amended
the bill to address these concerns. On May 5, 2025, Governor Polis, Attorney General Weiser, U.S.
Senator Bennet, U.S. Representatives Neguse and Pettersen, and Denver Mayor Johnston jointly
asked the General Assembly to “delay implementation of SB 24-205 until January 2027” so that
stakeholders could craft a solution that does not “stifl[e] innovation or driv[e] business away from
our state.” 57
74.
Attorney General Weiser reiterated his concerns again just three months later, on
August 5, 2025, when he warned that “[t]his bill is really problematic, it needs to be fixed.” 58
75.
Then, in a Special Session of the General Assembly in August 2025, lawmakers
introduced Senate Bill 25B-004, which would have repealed the vast majority of SB24-205’s
provisions, replaced them with a light-touch transparency framework, and delayed the law’s
effective date to June 30, 2026.59 The bill that the General Assembly ultimately passed and that
Governor Polis signed into law, however, merely delayed SB24-205’s effective date to June 30,
2026 without amending the law’s other provisions. 60
57
Jared Polis et al., Letter to Colo. General Assembly (May 5, 2025),
https://drive.google.com/file/d/1YsQalh778UIkHKQNXwg1WMEtWqtUWYCR/view?usp=shar
ing.
58
https://broadbandbreakfast.com/state-ag-warns-colorado-ai-bill-could-driveinnovation-out-of-state/.
59
https://leg.colorado.gov/bill_files/90582/download.
60
https://leg.colorado.gov/bill_files/90530/download.
30Page 31 Case No. 1:26-cv-01515
76.
Document 1
filed 04/09/26
USDC Colorado
pg 31 of 75
Understandably, politicians and industry groups continued to be concerned. 61 In
October 2025, Governor Polis convened a working group to review the law. 62 Meanwhile, policy
groups like the Cato Institute continued to express concerns that “[a] disruptive state patchwork
could impact the development and deployment of this important technology well beyond a single
state’s borders.” 63
77.
The United States government has repeatedly emphasized similar concerns. In a
December 11, 2025 executive order, see supra ¶¶ 50-52, the administration explained that “United
States AI companies must be free to innovate without cumbersome regulation”; “excessive State
regulation thwarts this imperative”; “State-by-State regulation by definition creates a patchwork
of 50 different regulatory regimes that makes compliance more challenging”; “State laws
sometimes impermissibly regulate beyond State borders, impinging on interstate commerce,” and
there must be a “minimally burdensome national standard – not 50 discordant State ones.” 64
Notably, the Executive Order specifically called out SB24-205, explaining that “State laws are
increasingly responsible for requiring entities to embed ideological bias within models,” and
SB24-205 “may even force AI models to produce false results in order to avoid a ‘differential
treatment or impact’ on protected groups.”65 And then on March 20, 2026, the White House’s
61
nts.pdf.
https://content.leg.colorado.gov/sites/default/files/images/chamber_of_progress_comme
62
https://www.coloradopolitics.com/2025/10/15/gov-polis-convenes-new-workinggroup-to-address-colorados-lingering-ai-law-ch
65
Id.
31Page 32 Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 32 of 75
“National AI Legislative Framework” again emphasized that “winning the AI race” will “usher in
a new era of human flourishing, economic competitiveness, and national security for the American
people,” and that its proposed framework “can succeed only if it is applied uniformly across the
United States. A patchwork of conflicting state laws would undermine American innovation and
our ability to lead in the global AI race.”66
78.
Meanwhile, on March 17, 2026, a working group convened by Governor Polis
published a proposed bill that would amend SB24-205 by removing its requirement to mitigate
against algorithmic discrimination and imposing narrower disclosure requirements on
developers. 67 That proposal, however, has not yet been introduced by any legislator in the General
Assembly, where it will be subject to amendment, committee hearings, floor debate, voting, and
the Governor’s approval.
79.
Although Attorney General Weiser has repeatedly expressed concern about SB24-
205 for substantially the same reasons explained in the White House’s executive orders, he has
since announced that he would challenge the White House’s December 11, 2025 Executive Order
if the administration seeks to withhold federal funding from Colorado based on SB24-205. 68
80.
SB24-205 is set to go into effect on June 30, 2026.
66
https://www.whitehouse.gov/articles/2026/03/president-donald-j-trump-unveilsnational-ai-legislative-framework/.
67
https://drive.google.com/file/d/1L2plsS3q1vzCrI8LuHj-5SNFjAoYoA_d/view.
68
https://www.cpr.org/2025/12/12/trump-artificial-intelligence-executive-order/.
32Page 33 Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 33 of 75
E.
SB24-205’s Onerous Requirements Impermissibly Burden Multiple
Constitutional Rights.
81.
Despite being billed as a consumer-protection measure, SB24-205 lacks any
statement of purpose or legislative findings evidencing the “algorithmic discrimination” that the
bill prohibits. It nevertheless imposes onerous, nationwide requirements that impermissibly burden
xAI’s constitutional rights under the First Amendment, Dormant Commerce Clause, Fourteenth
Amendment, and Equal Protection Clause.
82.
SB24-205 imposes affirmative duties on developers and deployers of “high-risk”
AI systems to mitigate “algorithmic discrimination” and make disclosures about (among other
things) those mitigation efforts. xAI is not a “deployer” of a regulated AI system within the
meaning of SB24-205, see § 6-1-1701(6), so only the provisions governing “developers” are at
issue in this action.
83.
To begin, the bill defines “[a]lgorithmic discrimination” as “any condition in which
the use of an [AI] system results in unlawful differential treatment or impact that disfavors an
individual or group of individuals on the basis of” certain protected characteristics. § 6-11701(1)(a). But not all “algorithmic discrimination” is prohibited. The bill proscribes only
discrimination that Colorado disagrees with. Indeed, it expressly exempts from its definition of
“algorithmic discrimination” any discrimination that “[e]xpand[s] an applicant, customer, or
participant pool to increase diversity or redress historical discrimination,” § 6-1-1701(1)(b)(I)(B),
thereby importing Colorado’s normative judgment of what preferences should be given to certain
groups.
84.
SB24-205’s substantive provisions build on the definition of “algorithmic
discrimination” in § 6-1-1701(1). First, the bill imposes an affirmative duty on a “developer of a
33Page 34 Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 34 of 75
high-risk [AI] system” to “use reasonable care to protect consumers from any known or reasonably
foreseeable risks of algorithmic discrimination arising from the intended and contracted uses of
the high-risk [AI] system,” § 6-1-1702(1)—regardless of whether discrimination is intended or in
fact occurs.
85.
Second, the bill requires that developers make certain disclosures about
“algorithmic discrimination” to deployers and other developers of the AI system, to the public,
and to the Attorney General. To deployers and other developers, a developer must “make
available” documentation disclosing or describing, among other things:
●
“known or reasonably foreseeable risks of algorithmic discrimination arising from
the intended uses of the high-risk [AI] system,” § 6-1-1702(2)(b)(II);
●
“[h]ow the high-risk [AI] system was evaluated for performance and mitigation of
algorithmic discrimination,” § 6-1-1702(2)(c)(I);
●
the “data governance measures used to cover the training datasets and the measures
used to examine the suitability of data sources, possible biases, and appropriate
mitigation,” § 6-1-1702(2)(c)(II);
●
the “measures the developer has taken to mitigate known or reasonably foreseeable
risks of algorithmic discrimination that may arise from the reasonably foreseeable
deployment of the high-risk [AI] system,” § 6-1-1702(2)(c)(IV); and
●
“[a]ny additional documentation that is reasonably necessary to assist the deployer
in understanding the outputs and monitor the performance of the high-risk [AI]
system for risks of algorithmic discrimination, § 6-1-1702(2)(d).
86.
For the public, “a developer shall make available, in a manner that is clear and
readily available on [its] website or in a public use case inventory, a statement summarizing,”
among other things, “[h]ow the developer manages known or reasonably foreseeable risks of
algorithmic discrimination that may arise from the development or intentional and substantial
modification of the types of high-risk [AI] systems described in accordance with subsection
(4)(a)(I) of this section.” § 6-1-1702(4)(a)(II). The developer “shall update” this public statement
34Page 35 Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 35 of 75
“[a]s necessary to ensure that the statement remains accurate” and “[n]o later than ninety days after
the developer intentionally and substantially modifies any high-risk [AI] system described in
subsection (4)(a)(I) of this section.” § 6-1-1702(4)(b)(I)-(II).
87.
And to the Attorney General and “all known deployers or other developers of the
high-risk [AI] system,” a developer “shall disclose … in a form and manner prescribed by the
[A]ttorney [G]eneral” “any known or reasonably foreseeable risks of algorithmic discrimination
arising from the intended uses of the high-risk [AI] system,” and shall make this disclosure “no
later than ninety days after” the developer discovers that its AI system “has caused or is reasonably
likely to have caused algorithmic discrimination” or receives a “credible report” that the system
has done so. § 6-1-1702(5). Further, at any time, the Attorney General can require a developer to
disclose the statement or documentation that § 6-1-1702(2) requires a developer to disclose to
deployers and other developers of the AI system. § 6-1-1702(7). A developer must comply “no
later than ninety days after the request and in a form and manner prescribed by the [A]ttorney
[G]eneral.” Id.
88.
The AI systems to which SB24-205’s requirements apply are near boundless.
Excepting yet another requirement that developers disclose “to each consumer who interacts with
the [AI] system that the consumer is interacting with an [AI] system,” § 6-1-1704(1), which
requirement appears to apply to all AI systems, SB24-205’s requirements facially apply only to
“high-risk [AI] systems.” § 6-1-1701(9)(a). Yet the phrase “high-risk” imposes no meaningful
limitation. An AI system is “high-risk” if it “makes, or is a substantial factor in making, a
consequential decision.” Id. And the definition of “substantial factor” renders the word
“substantial” meaningless. SB24-205 broadly defines that term as including (among other things)
35Page 36 Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 36 of 75
“any use of an [AI] system to generate any content, decision, prediction, or recommendation
concerning a consumer that is used as a basis to make a consequential decision.” § 6-1-1701(11)(b)
(emphasis added). Finally, and unsurprisingly, SB24-205 also broadly defines a “consequential
decision” as “a decision that has material legal or similarly significant effect on the provision or
denial to any consumer of, or the cost or terms of” any one of numerous opportunities or services,
including: “(a) [e]ducation enrollment or an education opportunity; (b) [e]mployment or an
employment opportunity; (c) [a] financial or lending service; (d) [a]n essential government service;
(e) [h]ealth-care services; (f) [h]ousing; (g) [i]nsurance; or (h) [a] legal service.” § 6-1-1701(3).
89.
Not only does SB24-205 apply to virtually any AI use in the covered subject
matters, it also has broad extraterritorial effect. The law applies to any “developer,” defined as “a
person doing business in this state that develops … an [AI] system.” § 6-1-1701(7). And for an AI
system that makes a “consequential decision” to become subject to SB24-205’s provisions, the
decision need only affect “any consumer”—that is, any “individual who is a Colorado resident.”
§§ 6-1-1701(3), (4), (9)(a). SB24-205 thus applies if even a single “Colorado resident” is impacted
by the system, regardless of where that Colorado resident is located, where the AI system is
developed or used, how unforeseeable that impact may be, or how many non-Colorado residents
are also impacted.
90.
Finally, SB24-205 grants the Attorney General broad and consequential
enforcement powers. First, it grants the Attorney General “exclusive authority to enforce” SB24205, § 6-1-1706(1), and a violation of the law “constitutes an unfair trade practice,” § 6-1-1706(2),
which is subject to a penalty of $20,000 per violation, § 6-1-112(1)(a). Second, it provides that the
36Page 37 Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 37 of 75
“[A]ttorney [G]eneral may promulgate rules as necessary for the purpose of implementing and
enforcing” the law. § 6-1-1707(1).
COUNT I
Declaratory Relief and Preliminary and Permanent Injunctive Relief for Violations of the
First Amendment to the United States Constitution (42 U.S.C. § 1983, 28 U.S.C. § 2201(a))
(Impermissible Burdens on Speech and Content- and Viewpoint-Based Discrimination)
91.
xAI re-alleges and incorporates by reference all allegations set forth above.
92.
SB24-205 interferes with xAI’s right to make speech-based editorial decisions
in developing Grok. xAI’s mission is to create an AI model that pursues the truth, without regard
to popularity, ideology, or social approval. This core philosophy, reflected in the publicly available
system prompts, pervades every aspect of Grok’s development. These are expressive decisions
protected by the First Amendment. SB24-205’s algorithmic-discrimination provision compels xAI
to alter these decisions and redesign Grok to promote Colorado’s views about fairness, equity, and
“acceptable” forms of discrimination based on protected characteristics such as race.
93.
The First Amendment provides that “Congress shall make no Law . . . abridging
the Freedom of Speech, or of the Press; or of the Right of the People to peaceably to assemble.”
U.S. Const. amend. I. These protections apply to the states through the Fourteenth Amendment’s
Due Process Clause.
94.
Under the First Amendment, “a speaker has the autonomy to choose the content of
his own message.” Hurley, 515 U.S. at 573. Indeed, since “all speech inherently involves choices
of what to say and what to leave unsaid,” it is axiomatic that “one who chooses to speak may also
decide ‘what not to say.’” Id. The First Amendment therefore “prohibits the government from
telling people what they must say,” Agency for Int’l Dev. v. Alliance for Open Soc’y Int’l, Inc., 570 U.S. 205, 213 (2013), as it protects “both the right to speak freely and the right to refrain from
37Page 38 Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 38 of 75
speaking at all,” Wooley v. Maynard, 430 U.S. 705, 714 (1977). Laws that compel speakers to
“alter[] the content of their speech” are necessarily “content based” and therefore presumptively
unconstitutional. Nat’l Inst. of Family Life Advocs. v. Becerra (“NIFLA”), 585 U.S. 755, 766
(2018) (cleaned up).
95.
These principles remain unchanged when the means for creating and disseminating
speech is an AI model. See 303 Creative LLC, 600 U.S. at 587; Reno v. Am. Civil Liberties Union, 521 U.S. 844, 870 (1997). “[W]hatever the challenges of applying the Constitution to everadvancing technology,” the First Amendment’s basic principles “do not vary.” Brown v. Ent.
Mechs. Ass’n, 564 U.S. 786, 790 (2011).
96.
Every choice that xAI makes when developing Grok is an expressive act protected
by the First Amendment. These choices embody deliberate judgment that reflects xAI’s hierarchy
of values and its viewpoint-driven philosophy about how best to develop AI tools that will advance
the knowledge, understanding, and overall progress of human civilization.
97.
Just as video game creators engage in speech when designing games, id., social
media platforms engage in speech when crafting algorithms to curate content, Moody, 603 U.S. at
733, and search engine companies engage in speech when programming their algorithms, Langdon
v. Google, Inc., 474 F. Supp. 2d 622, 629-30 (D. Del. 2007), AI companies engage in speech when
designing, training, and fine-tuning a model.
98.
The starkly divergent philosophies of today’s frontier laboratories confirm the
expressive nature of the choices xAI makes when developing Grok. xAI designs, trains, and tunes
Grok to be a maximally truth-seeking model that prioritizes candor and rigorous empirical
accuracy. Other developers, whether because they prioritize progressive ideals on social, cultural,
38Page 39 Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 39 of 75
and political issues or because they endorse government messaging, take a different approach.
These design choices manifest themselves in the way the models respond to queries.
99.
These are not minor technical differences; they represent competing visions for the
future of AI. The choices that xAI makes to implement its own vision that AI should be maximally
truth seeking are therefore core expressive activities.
100.
SB24-205’s algorithmic-discrimination provision—which imposes a duty on
developers to prevent their models from generating content that could result in disparate impact
when used in connection with making decisions related to employment, education, and more—
burdens xAI’s right to free expression under the First Amendment. From a technical perspective,
complying with SB24-205’s mandate would require redesigning, retraining, or constraining the
Grok model by, for example, recalibrating how the model decides what information to include in
responses, hard-coding additional response guardrails, or re-weighting training datasets.
101.
Every technique for discharging the reasonable care duty that SB24-205 imposes
requires altering the training data, the model’s internal decision-making architecture, or how the
model is fine-tuned. See generally Isabel O. Gallegos et al., Bias and Fairness in Large Language
Models: A Survey, arXiv preprint 34-56 (July 12, 2024) (surveying bias mitigation techniques
proposed in academic literature), https://arxiv.org/pdf/2309.00770; Yufei Guo et al., Bias in Large
Language Models: Origins, Evaluation, and Mitigation, arXiv preprint 15-18 (Nov. 16, 2024)
(same), https://arxiv.org/pdf/2411.10915. Some researchers propose augmenting training datasets
with examples that replace protected attribute words, such as gendered pronouns, to achieve
greater balance. See Kaiji Lu et al., Gender Bias in Neural Natural Language Processing, arXiv
preprint 5-6 (May 30, 2019), https://arxiv.org/pdf/1807.11714. This approach requires, for
39Page 40 Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 40 of 75
instance, augmenting a dataset containing the sequence “he worked as an engineer” with the
sequence “she worked as an engineer.” Others propose modifying the system prompts to neutralize
perceived bias. See Alex Tamkin et al., Evaluating and Mitigating Discrimination in Language
Model Decisions, arXiv preprint 8-9 (Dec. 6, 2023), https://arxiv.org/pdf/2312.03689. And some
researchers suggest modifying the model’s core architecture or using specialized reinforcement
learning techniques. See Vishnu Asutosh Dasu et al., Attention Pruning: Automated Fairness
Repair of Language Models via Surrogate Simulated Annealing, arXiv preprint 4-6 (Nov. 25,
2025) (examining approach to suppress the architectural components that disproportionately drive
biased outputs), https://arxiv.org/pdf/2503.15815; Yuntao Bai, Constitutional AI: Harmlessness
from AI Feedback, arXiv preprint 10-11, 20-23 (Dec. 15, 2022) (detailing how developers can use
reinforcement
learning
to
teach
models
to
not
produce
biased
results),
https://arxiv.org/pdf/2212.08073.
102.
It is also implicit in SB24-205’s text that developers will have to make substantive
adjustments to their model training, data curation, fine-tuning, and alignment processes. The
statute requires developers to make available documentation describing, among other things, the
“data governance measures used to cover the training datasets and the measures used to examine
the suitability of data sources, possible biases, and appropriate mitigation,” § 6-1-1702(2)(c)(II),
and “the measures … taken to mitigate known or reasonably foreseeable risks of algorithmic
discrimination,” § 6-1-1702(2)(c)(IV). These contemplated mitigation measures would undeniably
require xAI to alter Grok’s maximally truth-seeking design.
103.
Colorado cannot alter xAI’s message simply because it wants to amplify its own
views on the highly politicized subjects of fairness and equity—views that xAI does not share.
40Page 41 Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 41 of 75
This violates the fundamental rule “that a speaker has the autonomy to choose the content of [its]
own message” and triggers strict scrutiny. Hurley, 515 U.S. at 573.
104.
SB24-205 also burdens users’ right to access speech. In addition to restricting
xAI’s expressive activity, SB24-205’s algorithmic-discrimination provision burdens users’ right
to access the information and ideas they can generate using Grok.
105.
The First Amendment not only “foster[s] individual self-expression” but also
“afford[s] the public access to discussion, debate, and the dissemination of information and ideas.”
First Nat’l Bank of Bos. v. Bellotti, 435 U.S. 765, 783 (1978). It is thus “well established” that the
First Amendment protects not only the right to speak but also “the right to receive information and
ideas.” Stanley v. Georgia, 394 U.S. 557, 564 (1969) (emphasis added); see also Va. State Bd. of
Pharmacy v. Va. Citizens Consumer Council, Inc., 425 U.S. 748, 757 (1976); Martin v. City of
Struthers, 319 U.S. 141, 143 (1943).
106.
This right is an indispensable component of the First Amendment because the
“dissemination of ideas can accomplish nothing if otherwise willing addressees are not free to
receive and consider them” for it “would be a barren marketplace of ideas that had only sellers and
no buyers.” Lamont v. Postmaster Gen., 381 U.S. 301, 308 (1965) (Brennan, J., concurring); see
also Bd. of Educ., Island Trees Union Free Sch. Dist. No. 26 v. Pico, 457 U.S. 853, 867 (1982)
(explaining that “the right to receive ideas is a necessary predicate to the recipient’s meaningful
exercise of his own rights of speech, press, and political freedom”).
107.
The responses, predictions, explanations, and recommendations produced using
Grok qualify as speech under the First Amendment, and users are therefore entitled to receive those
outputs.
41Page 42 Case No. 1:26-cv-01515
108.
Document 1
filed 04/09/26
USDC Colorado
pg 42 of 75
The quintessential “mediums of expression” are “written or spoken words.” Hurley,
515 U.S. at 569. Words spoken during “conversation[s] … transmi[t] … ideas” and thus qualify
as speech. McCullen v. Coakley, 573 U.S. 464, 488 (2014); see also United States v. Alvarez, 567
U.S. 709, 722-23 (2012) (plurality op.) (acknowledging that “personal … conversations” are
speech); Legal Servs. Corp. v. Velazquez, 531 U.S. 533, 542-43 (2001) (so too is “advice from [an]
attorney to [a] client”).
109.
In form and function, a user’s interactions with Grok are indistinguishable from the
back-and-forth dialogue that occurs in ordinary human discourse. A user poses a prompt and Grok
replies with an answer, recommendation, or explanation—just as a knowledgeable human
interlocutor would do in real-time conversation.
110.
And even if considered in isolation, Grok’s outputs qualify as protected speech
because they communicate ideas no less than books, scholarly articles, professional reports, and
other media within the First Amendment’s heartland. See, e.g., 303 Creative, 600 U.S. at 587 (“All
manner of speech—from pictures, films, paintings, drawings, and engravings, to oral utterance and
the printed word—qualify for the First Amendment’s protections; no less can hold true when it
comes to speech … conveyed over the Internet.”) (cleaned up); Brown, 564 U.S. at 790 (“Like the
protected books, plays, and movies that preceded them, video games communicate ideas—even
social messages—through many familiar literary devices … and through features distinctive to the
medium …. That suffices to confer First Amendment protection.”).
111.
It makes no difference that Grok’s outputs are generated by an AI system; speech
need not be attributable to a speaker with First Amendment rights to qualify for protection. For
example, even though the First Amendment does not protect foreign speakers who produce
42Page 43 Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 43 of 75
political propaganda abroad, the Supreme Court held that a restriction on mailing such material
into the United States violated the “addressee’s First Amendment rights.” Lamont, 381 U.S. at 307.
And before the constitutional protection for corporations’ right to speak about elections was settled
in Citizens United, the Court rejected restrictions on corporate political speech based on the
interests of listeners. Bellotti, 435 U.S. at 777. The Court explained that the “inherent worth of the
speech in terms of its capacity for informing the public does not depend upon the identity of its
source, whether corporation, association, union, or individual.” Id.
112.
SB24-205’s algorithmic-discrimination provision interferes with users’ right to
receive speech generated through Grok. The provision requires developers of “high-risk [AI]
systems” to “use reasonable care to protect consumers from any known or reasonably foreseeable
risks of algorithmic discrimination.” § 6-1-1702(1). And “algorithmic discrimination” is defined
expansively as “any condition in which the use of an artificial intelligence system results in an
unlawful differential … impact that disfavors an individual or group” on the basis of certain
characteristics. § 6-1-1701(1)(a). Complying with this mandate will force xAI to alter Grok’s
outputs.
113.
The inevitable—indeed, intended—consequence is that users will receive only
government-approved, sanitized versions of Grok’s outputs. The statistically accurate
explanations, unfiltered recommendations, and candid analyses that risk producing disfavored
outcomes for protected classes will have to be altered, diluted, or withheld entirely. Users will thus
be systematically denied access to the full, unaltered outputs that Grok would otherwise provide.
This is a direct—and substantial—burden on users’ First Amendment right to receive information
and ideas through Grok, subjecting SB24-205 to strict scrutiny.
43Page 44 Case No. 1:26-cv-01515
114.
Document 1
filed 04/09/26
USDC Colorado
pg 44 of 75
SB24-205 regulates speech based on content and viewpoint. SB24-205 also
impermissibly draws distinctions based on content and viewpoint.
115.
It is the “most basic” principle of the First Amendment that the “government has
no power to restrict expression because of its message, its ideas, its subject matter, or its content.”
Brown, 564 U.S. at 790-91. “Content-based laws—those that target speech based on its
communicative content—are presumptively unconstitutional and may be justified only if the
government proves they are narrowly tailored to serve compelling state interests.” Reed v. Town
of Gilbert, 576 U.S. 155, 163 (2015) (citations omitted).
116.
By its plain terms, SB24-205 restricts only AI systems that generate content
affecting certain decisions. The law says nothing about AI systems generating outputs on protein
folding, computer coding, and the like. But if the same systems are used to generate content
addressing applications for employment, educational, credit, or housing opportunities, then SB24205 imposes an additional burden on the developer to exercise reasonable care to protect
consumers from reasonably foreseeable risks of algorithmic discrimination. §§ 6-1-1701(3), 6-11702(1). The law thus “singles out specific subject matter for differential treatment.” Reed, 576
U.S. at 169. “That is about as content-based as it gets.” Barr v. Am. Ass’n of Pol. Consultants, Inc., 591 U.S. 610, 619 (2020) (plurality op.).
117.
SB24-205 draws yet another content-based distinction when describing the
particular outputs that it regulates within those subject matters. It imposes a duty of reasonable
care to protect consumers from systems generating outputs that contain “algorithmic
discrimination,” a term the law defines as outputs that produce “unlawful differential treatment or
44Page 45 Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 45 of 75
impact that disfavors an individual or group of individuals on the basis” of certain characteristics
(e.g., age, race, and disability). § 6-1-1701(1)(a).
118.
xAI’s statutory compliance thus turns on whether Grok is generating outputs that
contain particular types of speech on particular subjects. This will necessarily require the Attorney
General “to examine the content of the message that is conveyed to determine whether a violation
has occurred,” a hallmark of content-based laws. Animal Legal Defense Fund v. Kelly, 9 F.4th
1219, 1228 (10th Cir. 2021) (quoting McCullen, 573 U.S. at 479).
119.
Indeed, on this front SB24-205 “goes even beyond mere content discrimination, to
actual viewpoint discrimination.” R.A.V. v. City of St. Paul, 505 U.S. 377, 391 (1992). This is
because the statute expressly carves out from its definition of “algorithmic discrimination” AI
systems with outputs that “expan[d] an applicant, customer, or participant pool to increase
diversity or redress historical discrimination.” § 6-1-1701(1)(b)(I)(B).
120.
In other words, SB24-205 does not neutrally target AI systems that produce any
disparate impact based on the identified characteristic. It targets only AI systems that might
produce outputs leading to what Colorado views as the wrong kind of disparate impact. That is
textbook viewpoint discrimination.
121.
SB24-205 cannot survive any level of heightened scrutiny, let alone strict
scrutiny. Because strict scrutiny applies, Colorado bears the heavy burden of demonstrating that
SB24-205 is “the least restrictive means of achieving a compelling state interest.” McCullen, 573
U.S. at 478. Even under intermediate scrutiny, Colorado would have to prove that SB24-205 is
“narrowly tailored to serve a significant governmental interest.” Packingham v. N. Carolina, 582
U.S. 98, 105-06 (2017). It cannot do so.
45Page 46 Case No. 1:26-cv-01515
122.
Document 1
filed 04/09/26
USDC Colorado
pg 46 of 75
SB24-205 is not narrowly tailored to advance any legitimate interest that Colorado
could assert in, for example, curbing invidious discrimination. The law is both overinclusive and
underinclusive. And there are far less restrictive alternatives that function to achieve the same
ends.
123.
SB24-205 is overinclusive in at least two ways. First, it reaches far beyond just
invidious discrimination by embracing a sweeping disparate-impact theory. The statute defines
“algorithmic discrimination” to encompass any use of an AI system that “results in an unlawful
differential treatment or impact” disfavoring an individual or group based on certain protected
characteristics. § 6-1-1701(1)(a). This disjunctive phrasing makes clear that SB24-205 targets not
only intentional discrimination, but also raw statistical imbalances unconnected to any
discriminatory animus. Governor Polis himself expressed reservations about SB24-205 for exactly
this reason, explaining that “[l]aws that seek to prevent discrimination generally focus on
prohibiting intentional discriminatory conduct,” but SB24-205 “deviates from that practice by
regulating the results of AI system use, regardless of intent.”
124.
Second, SB24-205 is overinclusive because it reaches outputs with little bearing on
any consequential decision, and therefore almost no relationship to the government’s interest in
preventing discrimination. Although the statute facially covers only AI that are a “substantial
factor” in making consequential decisions, it broadly defines that term to include “any use … to
generate any content … that is used as a basis to make a consequential decision” concerning a
consumer. § 6-1-1701(11)(b) (emphasis added). This captures using Grok to generate interview
questions, summarize writing samples, and perform other tasks for which the AI system plays no
direct role in the ultimate decision.
46Page 47 Case No. 1:26-cv-01515
125.
Document 1
filed 04/09/26
USDC Colorado
pg 47 of 75
Even performing purely administrative tasks can qualify as a “substantial factor”
under SB24-205. For example, judges might use an AI system to create a clerkship hiring
spreadsheet that lists every applicant’s name, law school, graduation year, and grade point average.
That is “content” that would be “used as a basis” to make hiring decisions. But there is no risk that
using the AI system introduced any bias into the decision making process. Thus, SB24-205 is “not
reasonably restricted to the evil with which it is said to deal.” Butler v. Michigan, 352 U.S. 380,
383 (1957).
126.
SB24-205 is also underinclusive when judged against any interest Colorado may
have in rooting out intentional discriminatory conduct. The statute defines “algorithmic
discrimination” broadly to encompass any use of an AI system that “results in an unlawful
differential treatment or impact that disfavors an individual” on the basis of certain characteristics.
§ 6-1-1701(1)(a). But it expressly exempts the use of AI systems to “[e]xpand[] an applicant,
customer, or participant pool to increase diversity or redress historical discrimination.” § 6-11701(1)(b)(I)(B). In other words, the law blesses the very thing it supposedly condemns. That type
of incoherence is the hallmark of underinclusivity. See Brown, 564 U.S. at 805.
127.
Finally, the availability of less restrictive alternatives further dooms SB24-205.
State and federal law—for example, the Civil Rights Act of 1964, the Fair Housing Act, the Equal
Credit Opportunity Act, and the Colorado Anti-Discrimination Act—already prohibit intentional
discrimination in employment, housing, education, finance, and other decisions within SB24-205’s
scope. These statutes target intentional discriminatory conduct with precision and without
unnecessarily restricting protected expression. “[R]egulating speech must be a last—not first—
resort.” Thompson v. Western States Med. Ctr., 535 U.S. 357, 373 (2002). And Colorado cannot
47Page 48 Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 48 of 75
show why these “available, effective alternatives” would fail where only speech coercion can
succeed. Ashcroft v. ACLU, 542 U.S. 656, 666 (2004).
128.
Because SB24-205 reaches far more speech than is necessary to advance any
professed interest in preventing invidious discrimination, is vastly underinclusive when judged
against that interest, and is not the least restrictive means, it cannot survive any level of First
Amendment scrutiny. SB24-205 should accordingly be declared unconstitutional, and the Attorney
General should be enjoined from enforcing it against xAI.
COUNT II
Declaratory Relief and Preliminary and Permanent Injunctive Relief for Violations of the
First Amendment to the United States Constitution (42 U.S.C. § 1983, 28 U.S.C. § 2201(a))
(Mandatory Disclosure Provisions)
129.
xAI re-alleges and incorporates by reference all allegations set forth above.
130.
SB24-205’s disclosure provisions likewise violate the First Amendment. By
forcing xAI to create, maintain, and disseminate information about its efforts to identify and
mitigate algorithmic bias, the bill compels xAI to speak in violation of its right to free speech.
131.
The Supreme Court has long recognized that the First Amendment’s guarantee of
free speech “includes both the right to speak freely and the right to refrain from speaking at all.”
Wooley, 430 U.S. at 714. That protection extends “not only to expressions of value, opinion, or
endorsement, but equally to statements of fact the speaker would rather avoid.” Hurley, 515 U.S.
at 573.
132.
Laws compelling speech are generally treated no differently than laws restricting
speech, see, e.g., NIFLA, 585 U.S. at 766-67, even when the government does not compel a speaker
to express any particular message, Riley, 487 U.S. at 795. As the Supreme Court has explained,
48Page 49 Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 49 of 75
the standard First Amendment analysis applies equally when the government compels speakers to
convey purely factual information. See Hurley, 515 U.S. at 573.
133.
SB24-205 requires developers to make disclosures regarding their evaluation and
mitigation of “algorithmic discrimination” to deployers, the public, and the Colorado Attorney
General. The disclosures to developers include, for example, documentation describing: (a) how
xAI evaluated the model for “mitigation of algorithmic discrimination,” § 6-1-1702(2)(c)(I);
(b) the “data governance measures used to cover the training datasets and the measures used to
examine the suitability of data sources, possible biases, and appropriate mitigation,” § 6-11702(2)(c)(II); and (c) the measures xAI “has taken to mitigate known or reasonably foreseeable
risks of algorithmic discrimination,” § 6-1-1702(2)(c)(IV). xAI must also publish on its website a
statement summarizing how it “manages known or reasonably foreseeable risks of algorithmic
discrimination.” § 6-1-1702(4)(a)(II). And xAI has to inform the Attorney General within ninety
days if it discovers that its AI system caused algorithmic discrimination or was “reasonably likely”
to have caused algorithmic discrimination, or if it receives a “credible report that the high-risk [AI]
system has been deployed and has caused algorithmic discrimination.” § 6-1-1702(5)(b).
134.
These disclosure provisions are content-based regulations that trigger strict scrutiny
because they compel xAI to disclose specific content related to its AI models. Strict scrutiny
applies whenever the government forces companies to disclose particular content, whether it be
non-commercial speech concerning factual information about a company’s service or products,
see X Corp. v. Bonta, 116 F.4th 888, 902 (9th Cir. 2024), risks those services or products may
pose, see NetChoice v. Weiser, 808 F. Supp. 3d 1223, 1240-41 (D. Colo. 2025), or information
about where to obtain other types of products or services, see NIFLA, 585 U.S. at 766. Like those
49Page 50 Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 50 of 75
content-based disclosures, the disclosures SB24-205 requires force xAI to speak a particular
message that it otherwise would not about whether and how it evaluates and mitigates “algorithmic
discrimination.”
135.
The disclosure provisions are also content- and viewpoint-based because they rely
on SB24-205’s definition of “algorithmic discrimination,” which itself draws distinctions based
on content and viewpoint—as explained above, ¶¶ 114-20. See NetChoice, LLC v. Reyes, 748 F.
Supp. 3d 1105, 1120 (D. Utah 2024) (finding “persuasive” the argument that “the entire Act
facially violates the First Amendment because the Act’s operative provisions each rely” on
definitions that impose content-based distinctions). SB24-205’s disclosure provisions thus trigger
strict scrutiny multiple times over.
136.
SB24-205’s disclosure provisions cannot withstand strict scrutiny or any other form
of heightened scrutiny. Colorado lacks a sufficient governmental interest in compelling AI
developers to disclose information about their practices for evaluating and mitigating “algorithmic
discrimination.” To satisfy First Amendment scrutiny, Colorado must “specifically identify an
actual problem in need of solving.” Brown, 564 U.S. at 799 (cleaned up); see also Reyes, 748 F.
Supp. 3d at 1124 (same). And the problem must be in need of a governmental solution, as opposed
to a private one.
137.
Many frontier AI laboratories already publish reports detailing methods for
evaluating and mitigating different types of bias. For instance, Anthropic engineers published a
report examining whether Claude 2.0 would generate biased results in simulated high-stakes
decision making scenarios (e.g., increasing a person’s credit). Tamkin et al., supra, at 5-6 (finding
what the authors term “evidence of positive discrimination (i.e., in favor of genders other than
50Page 51 Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 51 of 75
male and races other than white) in the Claude 2 model” for certain decision scenarios). The report
also considered prompt-engineering techniques for mitigating bias in the model’s decision making.
Id. And xAI assesses its models for sycophantic tendencies—i.e., whether they answer according
to the user’s suggestion rather than based on objective and unbiased subject information. See, e.g.,
xAI, Grok 4.1 Model Card (last updated November 17, 2025), https://data.x.ai/2025-11-17-grok4-1-model-card.pdf; xAI, Grok 4 Fast Model Card (last updated Sept. 19, 2025),
https://perma.cc/FEK8-CYQV.
138.
Academics, nonprofits, and other organizations likewise publish reports evaluating
whether AI systems exhibit bias in decision making. See, e.g., David Rozado, Fairness in AI
Decisions About People: Evidence from LLM Experiments, Manhattan Inst. (Jan. 8, 2026)
(describing
results
from
experiment
testing
prominent
LLMs
for
gender
bias),
https://manhattan.institute/article/fairness-in-ai-decisions-about-people-evidence-from-llmexperiments; Eitan Anzanberg et al., Evaluating the Promise and Pitfalls of LLMs in Hiring
Decisions, arXiv preprint (Jul. 28, 2025) (benchmarking several frontier, general-purpose LLMs
for candidate job-matching in resume screening), https://arxiv.org/pdf/2507.02087; Jiafu An et al.,
Measuring Gender and Racial Biases in Large Language Models: Intersectional Evidence from
Automated Resume Evaluation, arXiv preprint (Mar. 12, 2025) (evaluating whether certain LLMs
exhibit racial and gender bias when used to assess entry-level job candidates),
https://arxiv.org/pdf/2403.15281. 69
69
https://futurefreespeech.org/new-report-ai-laws-and-chatbots-face-a-global-freespeech-test/.
51Page 52 Case No. 1:26-cv-01515
139.
Document 1
filed 04/09/26
USDC Colorado
pg 52 of 75
Colorado does not have a sufficient interest in forcing developers to produce and
disseminate reams of information about their practices for evaluating and mitigating bias when
functionally identical information is already available from private sources. Whatever “modest
gap” may exist between what SB24-205 requires developers to disclose and what is already
available, filling it “can hardly be a compelling state interest.” Brown, 564 U.S. at 803.
140.
In addition, the disclosure provisions are not narrowly tailored for the same reasons
as the algorithmic discrimination provision—all of which rely on the same overinclusive and
underinclusive framework. See supra ¶¶ 121-28.
141.
Forcing developers to create and disseminate information is also not the least
restrictive means for accomplishing the State’s goals. One obvious alternative is for Colorado itself
to evaluate whether certain models are producing outputs that may result in “algorithmic
discrimination” and publish its findings. Indeed, it is not even clear why disclosures about
developers’ techniques for measuring and mitigating bias in training datasets and model
performance would provide meaningfully more information than simply testing the model—
something that the State (or any developer, for that matter) can do on its own. Indeed, the federal
government has adopted a similar government-led testing approach to ensure that models available
in the United States protect free speech and American values. See America’s AI Action Plan at 4
(directing Department of Commerce to “conduct research and, as appropriate, publish evaluations
of frontier models from the People’s Republic of China for alignment with Chinese Community
Party
talking
points
and
censorship”),
content/uploads/2025/07/Americas-AI-Action-Plan.pdf.
52
https://www.whitehouse.gov/wp-Page 53 Case No. 1:26-cv-01515
142.
Document 1
filed 04/09/26
USDC Colorado
pg 53 of 75
This case does not involve the kind of compelled-speech mandate that might get
lesser scrutiny under Zauderer v. Off. of Disciplinary Couns. of Supreme Ct. of Ohio, 471 U.S. 626
(1985). Zauderer creates a limited exception for government-compelled speech that aims to
combat misleading advertisements. It applies only to First Amendment claims involving
commercial speech. See NetChoice, 808 F. Supp. 3d at 1239-40. And the Supreme Court has never
applied the principles set forth in Zauderer outside the context of misleading advertising. See
NIFLA, 585 U.S. at 768. To the contrary, the Court has consistently reaffirmed that Zauderer
focused only on “combat[ing] the problem of inherently misleading commercial advertisements.”
Milavetz, Gallop & Milavetz, P.A. v. United States, 559 U.S. 229, 250 (2010); see also, e.g.,
Hurley, 515 U.S. at 573, United States v. United Foods, 533 U.S. 405, 416 (2001).
143.
Zauderer does not apply because SB24-205 does not compel commercial speech.
Commercial speech is speech “that does no more than propose a commercial transaction.” United
Foods, 533 U.S. at 409. That criterion is not simply a clue or guidepost; it is “the test for identifying
commercial speech.” City of Cincinnati v. Discovery Network, Inc., 507 U.S. 410, 423 (1993) (Bd.
of Trs. of State Univ. of N.Y. v. Fox, 492 U.S. 469, 473-74 (1989)). None of the information that
SB24-205 forces xAI to disclose (e.g., its practices for evaluating and mitigating model bias) has
anything to do with proposing a commercial transaction, let alone addressing any potentially
misleading advertising.
144.
The Tenth Circuit’s multi-factor test for commercial speech, which only applies in
close cases, likewise supports the conclusion that SB24-205’s disclosure provisions do not compel
commercial speech. That test turns on whether the speech (1) is “concededly an advertisement”;
(2) “refers to a specific product”; or (3) “is motivated by an economic interest.” United States v.
53Page 54 Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 54 of 75
Wenger, 427 F.3d 840, 847 (10th Cir. 2005) (citing Bolger v. Youngs Drug Prods. Corp., 463 U.S.
60, 66-67 (1983)). As an initial matter, these factors presuppose that an entity is already engaging
in speech that the government then regulates—and not that the government is forcing companies
to create speech. The government cannot compel private entities to speak in the first instance, and
then label that speech “commercial.” Even so, the Bolger factors weigh against concluding that
the disclosure provisions implicate commercial speech. The disclosures SB24-205 mandates are
not advertisements; they are not designed to sell a product or service. They do not refer to a
particular product; they describe xAI’s practices for evaluating and mitigating algorithmic bias.
And the disclosures would only be produced out of legal obligation, not economic motivation.
145.
Moreover, Zauderer does not apply because SB24-205’s compelled disclosures are
neither “purely factual” nor “uncontroversial.” A disclosure is “purely factual” if it only requires
“the disclosure of accurate, factual information.” Nat’l Ass’n of Wheat Growers v. Bonta, 85 F.4th
1263, 1276 (9th Cir. 2023). The word “purely” is critical: it excludes speech that embeds normative
framing, subjective interpretation, or predictive judgment. SB24-205’s disclosures are not purely
factual because they require developers to adopt and apply the statute’s definition of “algorithmic
discrimination,” a normative framework shaped by Colorado’s own views about equality and
fairness. Evaluating the risk of algorithmic discrimination and choosing appropriate mitigation
measures is also an inherently subjective exercise. And, tellingly, the disclosures here bear little
resemblance to those courts have concluded are “purely factual.” See, e.g., Zauderer, 471 U.S. at
652 (clarification that client may be liable for costs for attorney advertisements that promised “no
recovery, no legal fees”); Am. Meat Inst. v. U.S. Dep’t of Agric., 760 F.3d 18, 27 (D.C. Cir. 2014)
(en banc) (country-of-origin labeling); New York State Rest. Ass’n v. New York City Bd. of Health,
54Page 55 Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 55 of 75 556 F.3d 114, 131-32 (2d Cir. 2009) (calorie count); Nat’l Elec. Mfrs. Ass’n v. Sorrell, 272 F.3d
104, 107 (2d Cir. 2001) (instructions for safe product use and disposal).
146.
SB24-205’s compelled disclosures are also not “uncontroversial.” A statement is
“uncontroversial” under Zauderer “where the truth of the statement is not subject to good-faith
scientific or evidentiary dispute and where the statement is not an integral part of a live, contentious
political or moral debate.” Free Speech Coal., Inc. v. Paxton, 95 F.4th 263, 281-82 (5th Cir. 2024).
SB24-205’s disclosures force developers to opine on the public and scholarly debate about whether
using AI tools to make hiring (and other) decisions produces discriminatory results. Indeed, the
disclosures implicate one of the most persistent and divisive debates in American history—what
qualifies as discrimination. And xAI does not share Colorado’s views—that algorithmic
discrimination can be defined to exclude using AI tools to “expan[d] an applicant, customer, or
participant pool to increase diversity or redress historical discrimination.” § 6-1-1701(1)(b)(I)(B).
147.
The sheer volume of scholarly literature on how to evaluate and mitigate
algorithmic discrimination shows that SB24-205’s disclosures are not “uncontroversial.” Several
prominent papers catalog dozens of potential fairness metrics and distinct bias types, highlighting
the lack of consensus on how to measure or remediate bias. See Gallegos et al., supra, at 34-56;
Guo et al., supra, at 15-18. Other papers underscore the difficult accuracy trade-offs that
developers must weigh when considering measures designed to mitigate any bias that exists in the
training datasets. See Tamkin et al., supra, at 9-11. In fact, the Association for Computing
Machinery hosts an annual interdisciplinary conference focused on the social and technical
challenges of algorithmic systems, including the potential for bias. In short, scholars disagree
profoundly about whether AI systems produce discriminatory results, how to measure algorithmic
55Page 56 Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 56 of 75
bias, what causes bias in outputs, and the best strategies for mitigating bias. SB24-205’s compelled
disclosures would force xAI to take sides in this public debate by disclosing documentation
describing its own methodology for evaluating and mitigating algorithmic discrimination.
148.
That these issues are also the subject of active litigation further confirms that SB24-
205’s compelled disclosures are not “uncontroversial” under Zauderer. See, e.g., Mobley v.
Workday Inc., 740 F. Supp. 3d 796 (N.D. Cal. 2024) (disparate impact claim under Title VII based
on allegation that use of Workday’s AI system produced discriminatory hiring decisions); Huskey
v. State Farm Fire & Cas. Co., No. 22 C 7014, 2023 WL 5848164 (N.D. Ill. Sept. 11, 2023) (Fair
Housing Act claim based on allegation that use of machine-learning algorithm used to screen for
potentially fraudulent claims subjected black policyholders to additional administrative hurdles
and processing delays); Louis v. SafeRent Sols., LLC, 685 F. Supp. 3d 19 (D. Mass. 2023) (Fair
Housing Act claim based on allegation that tenant-screening algorithm biased against certain
applicants).
149.
SB24-205’s disclosure provisions trigger strict scrutiny because they are content-
and viewpoint-based. And they cannot withstand strict scrutiny or any other form of heightened
scrutiny. Zauderer does not apply because SB24-205’s disclosure provisions do not compel
commercial speech and, at any rate, the disclosures are neither “factual” nor “uncontroversial.”
SB24-205’s disclosure provisions should accordingly be declared unconstitutional, and the
Attorney General should be enjoined from enforcing them against xAI.
COUNT III
Declaratory Relief and Preliminary and Permanent Injunctive Relief for Extraterritorial
Regulation in Violation of the Commerce Clause of the United States Constitution and the
Horizontal Separation of Powers (U.S. Const. art. I, § 8, cl. 3)
150.
xAI re-alleges and incorporates by reference all allegations set forth above.
56Page 57 Case No. 1:26-cv-01515
151.
Document 1
filed 04/09/26
USDC Colorado
pg 57 of 75
The Constitution restricts the power of states to directly regulate conduct that takes
place entirely in another state. That bedrock principle of equal sovereignty among the states is
apparent in several of the Constitution’s structural protections and deeply rooted in our nation’s
historical tradition. See Nat’l Pork Producers Council v. Ross, 598 U.S. 356, 376 n.1 (2023); id.
at 403-04, 408-10 (Kavanaugh J., concurring in part and dissenting in part).
152.
Although this principle does not erect a per se bar against laws that regulate conduct
within one state in ways that have an “extraterritorial effect” in others, Ross, 598 U.S. at 357-58
(emphasis added), the same is not true of laws that “directly regulated out-of-state transactions.”
Id. at 376 n.1. “[O]riginal and historical understandings of the Constitution’s structure and the
principles of ‘sovereignty and comity’ it embraces” inform “the territorial limits of state authority
under the Constitution’s horizontal separation of powers.” Id. at 376 & 376 n.1.
153.
It is axiomatic that “all States enjoy equal sovereignty.” Shelby Cnty. v. Holder, 570 U.S. 529, 535 (2013); see also PPL Mont., LLC v. Montana, 565 U.S. 576, 591 (2012) (“the
States in the Union are coequal sovereigns under the Constitution”). Indeed, “the constitutional
equality of the states is essential to the harmonious operation of the scheme upon which the
Republic was organized.” Coyle v. Smith, 221 U.S. 559, 580 (1911). When a State reaches beyond
its own borders to “directly regulate[] out-of-state transactions by those with no connection to the
State,” Ross, 598 U.S. at 376 n.1 (emphasis omitted), it invades the sovereignty and impinges on
the equality of other states. Accordingly, one State is necessarily prohibited from directly
regulating conduct that neither occurs nor is directed within its borders. Cf. PennEast Pipeline Co.
v. New Jersey, 141 S.Ct. 2244, 2259 (2021) (“The plan of the Convention reflects the ‘fundamental
postulates implicit in the constitutional design.’” (quoting Alden v. Maine, 527 U.S. 706, 729
57Page 58 Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 58 of 75
(1999)); State Farm Mut. Auto. Ins. Co. v. Campbell, 538 U.S. 408, 422 (2003) (“A basic principle
of federalism is that each State may make its own reasoned judgment about what conduct is
permitted or proscribed within its borders, and each State alone can determine what measure of
punishment, if any, to impose on a defendant who acts within its jurisdiction.”).
154.
Consistent with our founding’s “special concern both with the maintenance of a
national economic union unfettered by state-imposed limitations on interstate commerce and with
the autonomy of the individual States within their respective spheres,” the Supreme Court has held
that the Commerce Clause (U.S. Const. art. I, § 8, cl. 3) prohibits any state from “control[ling]
commerce occurring wholly outside [its] boundaries.” Healy v. Beer Inst., Inc., 491 U.S. 324, 33536 (1989) (footnote omitted); see also, e.g., Brown-Forman Distillers Corp. v. N.Y. State Liquor
Auth., 476 U.S. 573, 579 (1986); Edgar v. MITE Corp., 457 U.S. 624, 643 (1982).
155.
The Court has also long interpreted the Due Process Clause to impose restrictions
on a state’s ability to regulate conduct occurring wholly outside its borders. See, e.g., Watson v.
Emps. Liab. Assurance Corp., 348 U.S. 66, 70 (1954) (recognizing “the due process principle that
a state is without power to exercise ‘extra territorial jurisdiction,’ that is, to regulate and control
activities wholly beyond its boundaries”); Home Ins. Co. v. Dick, 281 U.S. 397, 407-08 (1930)
(recognizing and applying the same principle).
156.
And, of course, the Tenth Amendment provides that “powers not delegated to the
United States by the Constitution, nor prohibited by it to the States, are reserved to the States
respectively, or to the people,” U.S. Const. amend. X, making clear that each state retains its own
“integrity, dignity, and residual sovereignty,” Bond v. United States, 564 U.S. 211, 221 (2011).
58Page 59 Case No. 1:26-cv-01515
157.
Document 1
filed 04/09/26
USDC Colorado
pg 59 of 75
Despite these prohibitions, SB24-205 creates liability for xAI based purely on out-
of-state conduct. The law broadly applies to any “developer”—“a person doing business in
[Colorado] that develops … an artificial intelligence system.” § 6-1-1701(7). Each of SB24-205’s
provisions further require that they impact a “consumer”—i.e., any “individual who is a Colorado
resident.” § 6-1-1701(4). The provisions that require developers to mitigate and disclose risks
concerning “algorithmic discrimination” cover “high-risk [AI] system[s],” which are systems that
“make[], or is a substantial factor in making, a consequential decision” affecting “any consumer.”
§§ 6-1-1701(1), (3), (4), (9)(a); 6-1-1702. Likewise, SB24-205 requires developers to disclose “to
each consumer who interacts with the artificial intelligence system that the consumer is interacting
with an artificial intelligence system.” § 6-1-1704(1). There is no requirement that the Colorado
resident be in Colorado or that the use of the AI system or “consequential decision” occur within
Colorado.
158.
Thus, xAI is subject to SB24-205’s provisions merely if (1) it is a “developer” and
(2) its AI system impacts a consequential decision affecting a Colorado resident. It is unimportant
where that decision occurs, where the Colorado resident is located, how unforeseeable
“algorithmic discrimination” was to xAI, or even whether the discrimination was intended.
159.
It is not difficult to think of an example reflecting how extraordinary these
provisions are in scope. Consider a healthcare company that is not incorporated or headquartered
in Colorado and does not maintain offices here. That company engaged in negotiations for Grok,
which was developed in California, and contracted with xAI, which is Nevada-incorporated and
California-based. Under SB24-205’s provisions governing “algorithmic discrimination,” xAI
would be exposed to liability under the law if a single medical decision affecting a Colorado
59Page 60 Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 60 of 75
resident visiting one of the healthcare company’s out-of-state offices is made using Grok. § 6-11702.
160.
Not only does xAI already have nationwide enterprise customers using Grok to
make consequential decisions, see supra ¶ 69, xAI’s enterprise business is rapidly growing, and
xAI is in the process of obtaining many additional customers who would leverage Grok to assist
in making decisions in other categories covered by SB24-205, including financial and lending
services, insurance, and housing. xAI’s continued expansion will certainly expose it to additional
liability under SB24-205.
161.
But it is not only xAI that faces this reality. None of the major enterprise AI
developers are headquartered or incorporated in Colorado:
●
xAI is Nevada-incorporated and California-headquartered;
●
Anthropic is Delaware-incorporated and California-headquartered;
●
OpenAI is Delaware-incorporated and California-headquartered; and
●
Google (Gemini) is Delaware-incorporated and California-headquartered.
SB24-205’s extraterritorial scope therefore is compounded.
162.
SB24-205’s provision requiring xAI to disclose “to each consumer who interacts
with the artificial intelligence system that the consumer is interacting with an artificial intelligence
system” is equally unconstitutional. § 6-1-1704(1). Like the provisions governing “algorithmic
discrimination,” § 6-1-1702, it applies regardless of where the “consumer” (i.e., Colorado resident)
is, and so applies to interactions with Grok nationwide.
163.
SB24-205 thus regulates transactions that are consummated entirely outside of
Colorado, by entities entirely outside of Colorado, and for systems developed entirely outside of
60Page 61 Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 61 of 75
Colorado. This violates the Constitution: indeed, imposing state-law liability on out-of-state actors
for actions taken entirely out of state is the definition of unconstitutional extraterritorial state
regulation. See Ross, 598 U.S. at 376 n.1 (“law that directly regulated out-of-state transactions by
those with no connection to the State” are unconstitutional); Edgar, 457 U.S. at 624; Styczinski v.
Arnold, 46 F.4th 907, 913 (8th Cir. 2022) (impermissibly extraterritorial where the law covers “a
Minnesota transaction[, which] includes a transaction anywhere in the world between a bullion
trader and a Minnesota resident. A bullion trader could therefore become subject to and violate
Minnesota law without conducting a single transaction in Minnesota.”).
COUNT IV
Declaratory Relief and Preliminary and Permanent Injunctive Relief for the Unlawful
Imposition of an Undue Burden on Interstate Commerce (Pike Balancing)
164.
xAI re-alleges and incorporates by reference all allegations set forth above.
165.
In addition to prohibiting states from regulating conduct that takes place outside of
their borders, the Commerce Clause also prohibits laws where “the burden imposed on interstate
commerce is clearly excessive in relation to the putative local benefits.” Green Room LLC v.
Wyoming, 157 F.4th 1196, 1209 (10th Cir. 2025); see also Air Transp. Ass'n of Am., Inc. v. Moss,
No. 1:23-cv-02421-DDD-KAS, 2024 WL 4369137, at *5 (D. Colo. Sept. 30, 2024).
166.
In Pike, 397 U.S. at 142, the Supreme Court articulated the factors that a Court
should consider when assessing whether a law unduly burdens interstate commerce. Pike balancing
requires the Court to consider four factors: “(1) the nature of the putative local benefits advanced
by the [statute]; (2) the burden the [statute] imposes on interstate commerce; (3) whether the
burden is clearly excessive in relation to the local benefits; and (4) whether the local interests can
be promoted as well with a lesser impact on interstate commerce.” Johnson & Johnson Vision
61Page 62 Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 62 of 75
Care, Inc. v. Reyes, 665 F. App’x 736, 744 (10th Cir. 2016) (internal quotation marks omitted).
All four factors weigh in xAI’s favor here.
167.
First, the burden the statute imposes on interstate commerce is significant. As
explained above, supra at ¶¶ 150-63, SB24-205 regulates transactions consummated entirely
outside of Colorado between xAI and entities outside of Colorado for systems developed entirely
outside of Colorado.
168.
Not only does SB24-205 apply to transactions entirely outside of the state of
Colorado, but it also applies regardless of how unforeseeable use by a Colorado resident is. In
order for xAI to comply with SB24-205’s “algorithmic discrimination” requirements, it would be
required to alter its products nationwide to accommodate Colorado’s requirements. Moreover, to
the extent complying with SB24-205 would require retraining Grok, supra at ¶¶ 100-02, that will
require significant economic and engineering investment.
169.
Second, the nature of the putative local benefit is highly speculative, at best. SB24-
205 itself contains no findings and cites no evidence that AI systems are causing discrimination—
much less systemic discrimination. Additionally, federal statutes, including Title VII of the Civil
Rights Act, the Fair Housing Act, the Equal Credit Opportunity Act, and the Americans with
Disabilities Act, and their Colorado analogs already prohibit discriminatory outcomes in the same
subject matter that SB24-205 covers.
170.
Third, the burden imposed by SB24-205 is clearly excessive in relation to the local
benefits. AI innovation is critical to American economic and national security. As the United States
government put it, AI “ha[s] the potential to reshape the global balance of power” and “it is a
national security imperative for the United States to achieve and maintain unquestioned and
62Page 63 Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 63 of 75
unchallenged global technological dominance.” 70 “United States AI companies must be free to
innovate without cumbersome regulation,” “excessive State regulation thwarts this imperative,”
“State-by-State regulation by definition creates a patchwork of 50 different regulatory regimes that
makes compliance more challenging,” and “State laws sometimes impermissibly regulate beyond
State borders, impinging on interstate commerce.” 71 The White House more recently explained
that “winning the AI race” would “usher in a new era of human flourishing, economic
competitiveness, and national security for the American people,” “AI cannot become a vehicle for
government to dictate right and wrong-think,” and “[a] patchwork of conflicting state laws would
undermine American innovation and our ability to lead in the global AI race.”72 AI systems
function similarly to the internet, rail, or highway insofar as they “require[] a cohesive national
scheme of regulation so that users are reasonably able to determine their obligations.” ACLU v.
Johnson, 194 F.3d 1149, 1162 (10th Cir. 1999). These concerns apply on all fours to SB24-205,
which imposes broad and costly requirements on AI developers operating anywhere nationwide.
Even the Attorney General himself has expressed concern about the impact SB24-205 might have,
going as far as saying the bill “is really problematic.” 73
70
Plan.pdf.
https://www.whitehouse.gov/wp-content/uploads/2025/07/Americas-AI-Action-
71
https://www.whitehouse.gov/presidential-actions/2025/12/eliminating-state-lawobstruction-of-national-artificial-intellig
72
https://www.whitehouse.gov/articles/2026/03/president-donald-j-trump-unveilsnational-ai-legislative-framework/.
73
https://broadbandbreakfast.com/state-ag-warns-colorado-ai-bill-could-driveinnovation-out-of-state/; see also https://newspack-coloradosun.s3.amazonaws.com/wpcontent/uploads/2024/06/FINAL-DRAFT-AI-Statement-6-12-24-JP-PW-and-RR-Si
63Page 64 Case No. 1:26-cv-01515
171.
Document 1
filed 04/09/26
USDC Colorado
pg 64 of 75
Fourth, the putative local interests served by SB24-205 can be promoted through
lesser impacts on interstate commerce. Colorado already has comprehensive discrimination laws.
For example, the Colorado Anti-Discrimination Act prohibits discriminatory or unfair
employment and housing practices. SB24-205 does not impair compliance with that or other antidiscrimination laws and even expressly states that nothing in it “restricts a developer’s … ability
to: (a) [c]omply with federal, state, or municipal laws, ordinances, or regulations.” § 6-11705(1)(a).
172.
SB24-205’s requirement that AI developers prevent “algorithmic discrimination”
might also have unintended effects. AI developers may incidentally overcorrect, resulting in the
very discrimination SB24-205 purports to address. See Students for Fair Admissions, Inc. v.
President & Fellows of Harvard Coll., 600 U.S. 181, 223 (2023) (“SFFA”) (“Outright racial
balancing is ‘patently unconstitutional’” because equal protection “command[s] that the
Government must treat citizens as individuals, not as simply components of a racial, religious,
sexual or national class.” (cleaned up)).
173.
Further, such model-rebalancing efforts have no “logical end point.” Id. at 221. It
is not clear at what point such efforts would satisfy compliance such that the AI developer has
used “reasonable care to protect consumers from any known or reasonably foreseeable risks of
algorithmic discrimination.” § 6-1-1702(1).
174.
SB24-205 imposes an enormous burden on interstate commerce—regulating
entirely out-of-state transactions and threatening to undermine American AI innovation—relative
to its local benefit—speculative concerns about “algorithmic discrimination” that lack any
legislative findings or evidence.
64Page 65 Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 65 of 75
COUNT V
Declaratory Relief and Preliminary and Permanent Injunctive Relief for Vagueness in
Violation of the Due Process Clause of the Fourteenth Amendment to the
United States Constitution (42 U.S.C. § 1983, 28 U.S.C. § 2201(a))
175.
xAI re-alleges and incorporates by reference all allegations set forth above.
176.
“A fundamental principle in our legal system is that laws which regulate persons or
entities must give fair notice of conduct that is forbidden or required.” FCC v. Fox TV Stations, 567 U.S. 239, 253 (2012). A law is unconstitutionally vague if it “fails to provide a person of
ordinary intelligence fair notice of what is prohibited, or is so standardless that it authorizes or
encourages seriously discriminatory enforcement.” United States v. Williams, 553 U.S. 285, 304
(2008). “When speech is involved, rigorous adherence to those requirements is necessary to ensure
that ambiguity does not chill protected speech.” Fox, 567 U.S. at 253-54. After all, vague laws risk
chilling speech by forcing speakers “to ‘steer far wider of the unlawful zone’” than they would “if
the boundaries of the forbidden areas were clearly marked.” Baggett v. Bullitt, 377 U.S. 360, 372
(1964). For that reason, laws touching on speech must themselves speak “with narrow specificity.”
NAACP v. Button, 371 U.S. 415, 433 (1963).
177.
SB24-205 is not written “with narrow specificity,” and it fails to provide fair notice
to a person of ordinary intelligence as to what it requires.
178.
To begin, it fails to adequately articulate what “algorithmic discrimination” is.
SB24-205 defines “algorithmic discrimination” as “an unlawful differential treatment or impact
that disfavors an individual or group of individuals on the basis of their actual or perceived age,
color, disability, ethnicity, genetic information, limited proficiency in the English language,
national origin, race, religion, reproductive health, sex, veteran status, or other classification
protected under the laws of this state or federal law.” § 6-1-1701(1)(a).
65Page 66 Case No. 1:26-cv-01515
179.
Document 1
filed 04/09/26
USDC Colorado
pg 66 of 75
But serious questions are left unanswered. What constitutes a “perceived attribute”?
“Perceived” by whom? Is it the perception of the individual who is allegedly being discriminated
against or the AI system allegedly resulting in a disparate impact? If the former, how sincerely
held does that perception need to be? How do “unlawful differential treatment” or “impact” differ?
And how do those tests differ from the traditional legal frameworks for disparate treatment?
180.
The vagueness of SB24-205’s “algorithmic discrimination” is further compounded
by its carve-outs. § 6-1-1701(1)(b). SB24-205 provides that “‘Algorithmic discrimination’ does
not include: . . . (B) [e]xpanding an applicant, customer, or participant pool to increase diversity
or redress historical discrimination.” Id. What constitutes “historical discrimination”? How far
back in time must one go? Does it only include the historical treatment of African Americans
arising from slavery, or does it include discrimination against the Irish and Italians in the 19th
century? 74 Nuziard v. Minority Bus. Dev. Agency, 721 F. Supp. 3d 431, 509 (N.D. Tex. 2024) (“In
the mid-19th Century, Irish immigrants faced a common roadblock when seeking jobs after their
trans-Atlantic voyage: motivated by rising populism and anti-Irish animus, many employers hung
signs explaining ‘Irish Need Not Apply.’”). Likewise, what is permissible “diversity”? § 6-1-1701
defines algorithmic discrimination to include virtually every actual or perceived protected trait;
which of those same traits are the appropriate subjects of State-approved “diversity”
considerations?
181.
Next, take SB24-205’s definition of a “high-risk artificial intelligence system,”
which “means any artificial intelligence system that, when deployed, makes, or is a substantial
74
https://www.nytimes.com/interactive/2019/10/12/opinion/columbus-day-italianamerican-racism.html.
66Page 67 Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 67 of 75
factor in making, a consequential decision.” § 6-1-1701(9)(a). When does an AI system become a
“substantial factor”? According to SB24-205, “substantial factor” “includes any use of an artificial
intelligence system to generate any content, decision, prediction, or recommendation concerning
a consumer that is used as a basis to make a consequential decision concerning the consumer.” §
6-1-1701(11)(b) (emphasis added). That definition—“any use” of an AI system to generate “any
content … that is used as a basis to make a [] decision”—is so amorphous it encompasses virtually
all uses of AI systems.
182.
Now, take xAI’s obligation to prevent “algorithmic discrimination.” § 6-1-1702.
SB24-205 requires xAI to use “reasonable care to protect consumers from any known or
reasonably foreseeable risks of algorithmic discrimination.” “[R]easonable care” and “reasonably
foreseeable” are undefined and uncertain.
183.
SB24-205’s disclosure provisions fare no better. For example, SB24-205 requires
developers to disclose “high-level summaries of the type of data used to train the high-risk artificial
intelligence system” as well as a “general statement describing the reasonably foreseeable uses
and known harmful or inappropriate uses of the high-risk artificial intelligence system.” §§ 6-11702(2)(b)(I), (2)(a). How “high-level” and how “general” do disclosures need to be to comply?
Likewise, what constitutes “harmful or inappropriate use,” which is, again, undefined?
184.
Similarly, the statute requires disclosure to the Attorney General of “known or
reasonably foreseeable risks” within ninety days of discovering an AI system caused algorithmic
discrimination, was “reasonably likely” to have caused algorithmic discrimination, or of receiving
a “credible report that the high-risk artificial intelligence system has been deployed and has caused
algorithmic discrimination.” § 6-1-1702(5)(b). When does algorithmic discrimination become
67Page 68 Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 68 of 75
“reasonably likely”? And what constitutes a “credible report” such that an AI developer is required
to comply with the disclosure provision? Would a single consumer report satisfy either trigger? If
not, how many consumer reports must an AI developer receive before it takes action and what
must those reports contain such that they are sufficiently “credible.” SB24-205 is long on questions
and short on answers.
185.
What’s worse, SB24-205 gives the Attorney General a massive amount of
interpretive and rulemaking power. SB24-205 gives the Colorado Attorney General rulemaking
authority to implement the law and has exclusive authority to enforce it. §§ 6-1-1706(1)(2), 6-11707. The Attorney General alone thus has authority to decide what actions violate SB24-205 and
its nebulous provisions and determine who to enforce it against.
186.
The statute’s use of vague, relative and uncertain terms gives developers no fair
notice of what mitigation or disclosures are required, and hands the Colorado Attorney General
virtually unfettered discretion to enforce the law arbitrarily against his political opponents. SB24205’s terms are so malleable that it authorizes “arbitrary [or] discriminatory enforcement.”
Grayned v. City of Rockford, 408 U.S. 104, 108 (1972).
COUNT VI
Declaratory Relief and Preliminary and Permanent Injunctive Relief for Violation of the
Equal Protection Clause of the Fourteenth Amendment to the United States Constitution
(42 U.S.C. § 1983, 28 U.S.C. § 2201(a))
187.
xAI re-alleges and incorporates by reference all allegations set forth above.
188.
“Proposed by Congress and ratified by the States in the wake of the Civil War, the
Fourteenth Amendment provides that no State shall ‘deny to any person . . . the equal protection
of the laws.’” “Proponents of the Equal Protection Clause described its ‘foundation[al] principle’
as ‘not permit[ing] any distinctions of law based on race or color.’” “Any ‘law which operates
68Page 69 Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 69 of 75
upon one man,’ they maintained, should ‘operate equally upon all.’ Accordingly, as th[e]
[Supreme] Court's early decisions interpreting the Equal Protection Clause explained, the
Fourteenth Amendment guaranteed ‘that the law in the States shall be the same for the black as for
the white; that all persons, whether colored or white, shall stand equal before the laws of the
States.’” SFFA, 600 U.S. at 183 (citations omitted). “Eliminating racial discrimination means
eliminating all of it.” Id. at 183-84.
189.
“Given its drafters’ desire to eschew race-based distinctions, even benign
discrimination cannot be squared with the Equal Protection Clause—it can only be an ‘exception.’”
Nuziard, 721 F. Supp. 3d at 477-78 (citation omitted). “Any exceptions to the Equal Protection
Clause’s guarantee must survive a daunting two-step examination known as ‘strict scrutiny,’ which
asks first whether the racial classification is used to ‘further compelling governmental interests,’
and second whether the government’s use of race is ‘narrowly tailored,’ i.e., ‘necessary,’ to achieve
that interest.” SFFA, 600 U.S. at 184 (citations omitted). SB24-205’s requirement to mitigate
“algorithmic discrimination” flunks that test.
190.
SB24-205 requires xAI to mitigate against and make disclosures concerning
“algorithmic discrimination.” §§ 6-1-1701(1), 1702. According to Colorado, however, not all
algorithmic discrimination is impermissible “algorithmic discrimination.” While xAI has a general
duty to mitigate against “algorithmic discrimination,” “‘[a]lgorithmic discrimination’ does not
include: (I) [t]he offer, license, or use of a high-risk artificial intelligence system by a developer
or deployer for the sole purpose of … (B) [e]xpanding an applicant, customer, or participant pool
to increase diversity or redress historical discrimination.” § 6-1-1701(1)(b)(I)(B). In other words,
if xAI engages in balancing to “increase diversity” or “redress historical discrimination,” including
69Page 70 Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 70 of 75
on the basis of “race,” it does not engage in “algorithmic discrimination” within the meaning of
SB24-205 at all. Id.; see also § 6-1-1701(1).
191.
Accordingly, if an AI developer alters its model to provide preferencing mandated
by Colorado to increase diversity or redress historical discrimination, it is effectively exempt from
SB24-205’s mitigation requirements. That is because if an AI developer alters its AI system to
increase “diversity” or “redress historical discrimination” then any disparate treatment that occurs
as a result does not constitute “algorithmic discrimination.” Therefore, the AI developer need not
use “reasonable care to protect consumers from any known or reasonably foreseeable risks of
algorithmic discrimination” because no such “algorithmic discrimination” could occur. § 6-11702.
192.
SB24-205 thus presents xAI with a choice. xAI can either continue to develop Grok
as a maximally truth-seeking AI system that provides objective and unbiased outputs; expend
significant resources on compliance, documentation, and risk mitigation; and face full enforcement
risk under SB24-205 for any disparate impact. Or it can build an AI with the sole purpose of
increasing diversity or redressing historical discrimination, and receive a safe harbor from potential
liability under SB24-205. That is an impermissible choice. Affording benefits to some, but not
others, based on the treatment of protected characteristics, including race, is precisely the type of
“[d]istinction[] between citizens solely because of their ancestry [that] are by their very nature
odious to a free people whose institutions are founded upon the doctrine of equality.” SFFA, 600
U.S. at 184 (citations omitted).
70Page 71 Case No. 1:26-cv-01515
193.
Document 1
filed 04/09/26
USDC Colorado
pg 71 of 75
SB24-205 includes no justification, much less a compelling justification, for its
codified discrimination. Again, it includes no statement of purpose. SB24-205 does not even define
what “diversity” is or what “historical discrimination” it seeks to redress.
194.
At best, SB24-205’s discrimination provision seems to be designed to embody
Colorado’s subjective vision of equity and remedy amorphous discrimination across a wide range
of services and opportunities, including education, employment, financial or lending services,
health-care services, housing, essential government services, insurance and legal services. § 6-11701(3). But “an effort to alleviate the effects of societal discrimination is not a compelling
interest.” Shaw v. Hunt, 517 U.S. 899, 909 (1996). If it was, the Equal Protection Clause “would
be lost in a mosaic of shifting preferences based on inherently unmeasurable claims of past
wrongs.” City of Richmond v. J.A. Croson Co., 488 U.S. 469, 505-06 (1989). “The mere recitation
of a benign or compensatory purpose for the use of a racial classification would essentially entitle
the States to exercise the full power of Congress under § 5 of the Fourteenth Amendment and
insulate any racial classification from judicial scrutiny under § 1. We believe that such a result
would be contrary to the intentions of the Framers of the Fourteenth Amendment, who desired to
place clear limits on the States’ use of race as a criterion for legislative action, and to have the
federal courts enforce those limitations.” Id. at 490-91 (citations omitted).
195.
To the extent that Colorado believes SB24-205 addresses specific instances of
discrimination, it must provide facts to support that conclusion. “First, it must identify the past or
present discrimination ‘with some specificity.’ Second, it must also demonstrate that a ‘strong
basis in evidence’ supports its conclusion that remedial action is necessary.” Concrete Works of
Colo. v. City & Cty. of Denver, 321 F.3d 950, 958 (10th Cir. 2003) (citations omitted); see also
71Page 72 Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 72 of 75
Nuziard, 721 F. Supp. 3d at 480. Colorado “cannot point to general social ills and call it a day.
Rather, it must identify the ‘who, what, when, where, why, and how’ of relevant discrimination.
Nuziard, 721 F. Supp. 3d at 480 (citing Croson, 488 U.S. at 492). Yet SB24-205 includes no
legislative findings or evidence whatsoever, much less any that would prove the existence of
specific harm that the state has a compelling interest in redressing.
196.
SB24-205 is also not narrowly tailored. In fact, the provision excepting permissible
“algorithmic discrimination” applies to any protected characteristic so long as their inclusion
increases “diversity” or redresses “historical discrimination.” § 6-1-1701(1)(b)(I)(B). It also
applies to virtually any decision that yields disparate results because SB24-205 covers nearly every
subject matter—education, employment, finance, government services, healthcare, housing,
insurance, and legal services. § 6-1-1701(3). That is woefully overinclusive. SB24-205 encodes
State preferencing for virtually any protected characteristic across virtually every subject matter.
Take employment for example: “If a [preference] was ‘narrowly tailored’ to compensate black
contractors for past discrimination, one may legitimately ask why they are forced to share this
‘remedial relief’ with an Aleut citizen who moves to [Boulder] tomorrow?” Croson, 488 U.S. at
506.
197.
SB24-205 is overinclusive for another reason: its carveout for “diversity” and
“historical discrimination” paints in broad strokes and encompasses members with protected traits
who may personally experience no discrimination or disparate impact. As Governor Polis
explained when expressing concern over the bill, “[l]aws that seek to prevent discrimination
72Page 73 Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 73 of 75
generally focus on prohibiting intentional discriminatory conduct,” but SB24-205 “deviates from
that practice by regulating the results of AI system use, regardless of intent.” 75
198.
SB24-205 is also underinclusive because the definition of “algorithmic
discrimination” does not include protections for individuals or groups that may actually face
disparate impact in the covered subjects (housing, finance, legal services, etc.), including the
economically disadvantaged. As one Court succinctly put it, “Oprah Winfrey is presumptively
disadvantaged, while … even more disadvantaged Americans are not.” Nuziard, 721 F. Supp. 3d
at 491.
199.
SB24-205 also lacks a “logical end point.” SFFA, 600 U.S. at 221. SB24-205
includes no guidance indicating when “increas[ing] diversity” or “redress[ing] historical
discrimination” will be sufficiently addressed by changes in xAI’s model such that further
inclusion of protected groups would actually decrease “diversity” or affirmatively cause “historical
discrimination” of excluded majority groups. SB24-205’s extraordinarily broad affirmative
discrimination provisions make it “unclear how a court,” or xAI for that matter “is supposed to
determine if or when such goals would be adequately met.” Id.
200.
Lastly, as explained above, less restrictive alternatives exist. State and federal
law—for example, the Civil Rights Act of 1964, Fair Housing Act, Equal Credit Opportunity Act,
and Colorado Anti-Discrimination Act—already prohibit intentional discrimination in
employment, housing, education, finance, and other decisions within SB24-205’s scope.
75
Jared Polis, Statement on Signing SB24-205 (May 17, 2024),
https://drive.google.com/file/d/1i2cA3IG93VViNbzXu9LPgbTrZGqhyRgM/view.
73Page 74 Case No. 1:26-cv-01515
201.
Document 1
filed 04/09/26
USDC Colorado
pg 74 of 75
SB24-205’s carveout for permissible “algorithmic discrimination” is as amorphous
as it is unconstitutional, and forces xAI to choose between adopting Colorado’s state-ordained
preferencing in its AI model or continuing to operate its maximally-truth seeking AI and expose
itself to liability, including civil penalties.
PRAYER FOR RELIEF
For the foregoing reasons, xAI respectfully requests that the Court:
A.
Enter a judgment declaring, pursuant to 28 U.S.C. §§ 2201(a) and 2202, that SB24-
205 unconstitutionally compels xAI’s speech in violation of the First Amendment of the U.S.
Constitution;
B.
Enter a judgment declaring, pursuant to 28 U.S.C. §§ 2201(a) and 2202, that SB24-
205 constitutes an impermissible extraterritorial regulation;
C.
Enter a judgment declaring, pursuant to 28 U.S.C. §§ 2201(a) and 2202, that SB24-
205 imposes an undue burden on interstate commerce;
D.
Enter a judgment declaring, pursuant to 28 U.S.C. §§ 2201(a) and 2202, that SB24-
205 is unconstitutionally vague in violation of the Due Process Clause of the U.S. Constitution;
E.
Enter a judgment declaring, pursuant to 28 U.S.C. §§ 2201(a) and 2202, that SB24-
205 violates the Equal Protection Clause of the Fourteenth Amendment of the U.S. Constitution;
F.
Issue an order permanently enjoining Attorney General Weiser, as well as all
officers, agents, and employees subject to his supervision, direction, or control, from enforcing the
provisions of SB24-205 against xAI;
G.
Award xAI all costs and attorney’s fees to which it is entitled pursuant to 42 U.S.C.
§ 1988 and other applicable law; and
74Page 75 Case No. 1:26-cv-01515
H.
Document 1
filed 04/09/26
USDC Colorado
pg 75 of 75
Grant xAI such other and further relief as the Court deems just and proper.
Dated: April 9, 2026.
Respectfully submitted,
s/ Frederick R. Yarger
Frederick R. Yarger
William D. Hauptman
Wheeler Trigg O’Donnell LLP
370 Seventeenth Street, Suite 4500
Denver, CO 80202
Telephone: 303.244.1800
Facsimile: 303.244.1879
Email: yarger@wtotrial.com
hauptman@wtotrial.com
James Burnham
X.AI LLC
601 13th Street NW 12th Floor
Washington, DC 20005
Email: jburnham@x.ai
Adam Mehes
Argirios J. Nickas
X.AI LLC
249 W 17th Street
New York, NY 10011
Email: amehes@x.ai
anickas@x.ai
Reid Coleman
X.AI LLC
865 FM 1209, Building 2
Bastrop, TX 78602
Email: rcoleman@x.ai
Attorneys for Plaintiff X.AI LLC
75
PDF Page 1
PlainSite Cover Page
PDF Page 2
Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 1 of 75
IN THE UNITED STATES DISTRICT COURT
FOR THE DISTRICT OF COLORADO
Civil Action No.: 1:26-cv-01515
X.AI LLC,
Plaintiff,
v.
PHILIP J. WEISER, COLORADO ATTORNEY GENERAL,
Defendant.
COMPLAINT FOR DECLARATORY AND INJUNCTIVE RELIEF
Plaintiff X.AI LLC (“xAI”) brings this civil action against Philip. J. Weiser, in his official
capacity as Colorado Attorney General, for declaratory and injunctive relief and alleges as follows:
INTRODUCTION
1.
Technologies powered by artificial intelligence (“AI”) are fast becoming the
primary instrument through which we discover, verify, and transmit knowledge. AI systems are
promising instruments of enlightenment and free expression, capable of democratizing knowledge
in ways the printing press or the internet never could and accelerating scientific breakthroughs at
a pace once unimaginable. They also have the potential to undermine free inquiry and public
discourse on an unprecedented scale: whoever owns or aligns the dominant models can quietly
alter what counts as “truth”—portraying ideological fads or falsehoods as reality—can filter
inconvenient facts, and can curate entire worldviews. This is dangerous to the advancement of
1
PDF Page 3
Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 2 of 75
human knowledge. Only by seeking objective truth can AI faithfully help humanity understand the
universe and advance the progress of human civilization.
2.
xAI is a leading frontier AI developer whose flagship model is known as Grok. xAI
believes that AI’s knowledge should be all-encompassing and as far-reaching as possible. It builds
AI specifically to advance human comprehension and capabilities. xAI named Grok after a word
in Robert Heinlein’s famous work Stranger in a Strange Land. As Heinlein famously put it, to
“grok” something is “to understand so thoroughly that the observer becomes a part of the
observed—to merge, blend, intermarry, lose identity in group experience.” Robert A. Heinlein,
Stranger in a Strange Land 206 (Berkley Medallion Ed., 1968).
3.
In keeping with that mission, xAI has designed and developed Grok to answer to
only evidence and reason, without regard to political correctness, ideological biases, or anything
that might distort objective truth. This unwavering commitment ensures that Grok discharges its
fundamental mission—assisting humanity in understanding the universe. But the State of Colorado
now seeks to force xAI to abandon its disinterested pursuit of truth and instead promote the State’s
ideological views on various matters, racial justice in particular.
4.
This case seeks to enjoin the enforcement and implementation of Colorado’s Senate
Bill 24-205 (“SB24-205”), a statute that severely burdens the development and use of AI. SB24205’s purported focus is a prohibition on so-called “algorithmic discrimination.” But SB24-205 is
decidedly not an anti-discrimination law. It is instead an effort to embed the State’s preferred views
into the very fabric of AI systems. Its provisions prohibit developers of AI systems from producing
speech that the State of Colorado dislikes, while compelling them to conform their speech to a
State-enforced orthodoxy on controversial topics of great public concern. This attempted coercion
2
PDF Page 4
Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 3 of 75
is unconstitutional under the First Amendment. Indeed, “[o]n the spectrum of dangers to free
expression, there are few greater than allowing the government to change the speech of private
actors in order to achieve its own conception of speech nirvana.” Moody v. NetChoice, LLC, 603
U.S. 707, 741-42 (2024).
5.
SB24-205 has been controversial and legally suspect from the outset. When
Governor Polis signed the law in May 2024, he expressed “reservations” about its onerous
provisions. 1 Governor Polis cautioned that he was “concerned about the impact this law may have
on an industry that is fueling critical technological advancements across our state for consumers
and enterprises alike. Government regulation that is applied at the state level in a patchwork across
the country can have the effect to hamper innovation and deter competition in an open market.”2
Weeks later, Governor Polis, Attorney General Weiser, and Senator Robert Rodriguez (who
introduced SB24-205) issued a joint statement announcing their desire to “revise the new law” and
“minimize unintended consequences associated with its implementation.” 3
6.
A year passed and the General Assembly had not amended the law, prompting
Governor Polis, Attorney General Weiser, U.S. Senator Bennet, U.S. Representatives Neguse and
Pettersen, and Denver Mayor Johnston to ask the General Assembly in May 2025 to “delay
implementation of SB 24-205 until January 2027” so that stakeholders could craft a solution that
1
Jared Polis, Statement on Signing SB24-205 (May 17,
https://drive.google.com/file/d/1i2cA3IG93VViNbzXu9LPgbTrZGqhyRgM/view.
2
2024),
Id.
3
Jared Polis et al., Letter Concerning SB24-205 (June 13, 2024),
https://drive.google.com/file/d/1UtIHZRKDq4a09q0mzT_o4cBjl3_qanrg/view?usp=sharing.
3
PDF Page 5
Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 4 of 75
does not “stifl[e] innovation or driv[e] business away from our state.”4 Attorney General Weiser
repeated these concerns just three months later, in August 2025, acknowledging that the “bill is
really problematic [and] needs to be fixed.” 5 The Governor issued an executive order that same
month convening a Special Session of the General Assembly in part to do just that.6 But in that
session the legislature merely delayed SB24-205’s effective date to June 30, 2026—it did nothing
to “fix” SB24-205. As of the date of this Complaint, the General Assembly has not introduced a
bill to amend the law during the current legislative session, which ends the second week of May
of this year.
7.
SB24-205’s problems are many, but its most obvious constitutional defect appears
in its first section. The law defines “algorithmic discrimination” as “any condition in which the
use of an [AI] system results in an unlawful differential treatment or impact that disfavors an
individual or group.” § 6-1-1701(1)(a). 7 But the law then simultaneously promotes a form of
“differential treatment” the State favors—discrimination intended to “increase diversity or redress
historical discrimination.” § 6-1-1701(1)(b)(I)(B).
8.
SB24-205’s internally contradictory, politicized definition of “algorithmic
discrimination” underlies nearly all of its substantive requirements. Among other things, the law
4
Jared Polis et al., Letter to Colo. General Assembly (May 5, 2025),
https://drive.google.com/file/d/1YsQalh778UIkHKQNXwg1WMEtWqtUWYCR/view?usp=shar
ing.
5
Cameron Marx, State AG Warns Colorado AI Bill Could Drive Innovation Out of State,
BroadbandBreakfast: AI (Aug. 5, 2025), https://broadbandbreakfast.com/state-ag-warnscolorado-ai-bill-could-drive-innovation-out-of-state/.
6
Colo. Exec. Order No. D 2025 009 (Aug. 6, 2025), https://drive.google.com/file/d/1kx5WNwRYMD7K33lJQEQaSm6RsMUBb0X/view?pli=1.
7
Unless otherwise stated, statutory citations refer to the Colorado Revised Statutes.
4
PDF Page 6
Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 5 of 75
imposes affirmative duties on AI developers to “protect” consumers from “risks of algorithmic
discrimination.” §§ 6-1-1702(1), 6-1-1703(1). It mandates disclosures about those “risks” to other
“developers” and “deployers” of AI systems, to consumers, to the public, and to the Attorney
General. §§ 6-1-1702(2)-(5), 6-1-1702(7), 6-1-1703(4)-(5), 6-1-1703(9). And the law mandates
the implementation of risk-management policies and the creation of “impact assessments.” § 6-11703(2)-(3).
9.
By requiring “developers” and “deployers” to differentiate between discrimination
that Colorado disfavors and discrimination that Colorado favors, SB24-205 compels Plaintiff
xAI—a “developer” under the law—to alter Grok, forcing Grok’s output on certain State-selected
subjects to conform to a controversial, highly politicized viewpoint. But the State “may not compel
[xAI] to speak its own preferred messages.” 303 Creative LLC v. Elenis, 600 U.S. 570, 586 (2023).
Doing so impermissibly “alter[s] the expressive content of [xAI’s speech].” Id. at 585 (quoting
Hurley v. Irish-Am. Gay, Lesbian and Bisexual Grp. of Boston, 515 U.S. 557, 572-73 (1995)). And
by “[m]andating speech that [xAI] would not otherwise make,” SB24-205 “necessarily alters the
content of the speech.” Riley v. Nat’l Fed. of the Blind, 487 U.S. 781, 795 (1988). SB24-205’s
content- and viewpoint-based speech compulsion is a clear violation of xAI’s First Amendment
rights.
10.
The First Amendment is not the only constitutional provision SB24-205 violates.
The law broadly applies to “any use of an [AI] system to generate any content, decision, prediction,
or recommendation concerning a consumer that is used as a basis to make a consequential
decision”—a decision affecting education, employment, financial services, and more. § 6-11701(3), § 6-1-1704(11)(b). And it defines the “consumer” it purports to protect as “an[y]
5
PDF Page 7
Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 6 of 75
individual who is a Colorado resident.” § 6-1-1701(4). In other words, SB24-205 is not
geographically limited to Colorado; despite the fact that AI systems are deployed across the
country, SB24-205 applies anywhere “a Colorado resident” is affected by an AI system. SB24205’s provisions therefore apply to a wide range of AI uses nationally—regardless of where the
“consumer” is located, where the AI system is developed, deployed, or used, or how foreseeable
the AI use is—so long as a single “Colorado resident” is impacted by the AI system. As applied to
xAI, a Nevada-incorporated, California-headquartered company with no offices in Colorado,
SB24-205 violates the Dormant Commerce Clause by directly regulating development and
deployment activities occurring entirely outside Colorado. It also fails the balancing test
established by Pike v. Bruce Church, Inc., 397 U.S. 137 (1970), by imposing burdens on interstate
commerce that far outweigh any local benefit.
11.
SB24-205 is also unconstitutionally vague. It fails to adequately define essential
terms such as “high-risk artificial intelligence system,” “algorithmic discrimination,” and
“historical discrimination.” The General Assembly left it to the Colorado Attorney General—who
has exclusive authority to enforce SB24-205’s provisions—to decide what those provisions mean
by “promulgat[ing] rules.” § 6-1-1707(1). This vagueness invites arbitrary enforcement.
12.
Finally, SB24-205 codifies discrimination without legally sufficient justification.
The bill—which lacks any legislative findings about the “algorithmic discrimination” the General
Assembly elected to proscribe—compels AI developers like xAI to endorse Colorado’s views on
diversity, equity, and inclusion or face significant compliance costs and civil fines. The Stateprescribed orthodoxy SB24-205 requires is as harmful as it is unconstitutional. It forces AI
6
PDF Page 8
Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 7 of 75
developers to distort their AI models to seek and output progressive ideology instead of the truth.
SB24-205 therefore violates the Equal Protection Clause.
13.
As Governor Polis and Attorney General Weiser have publicly (and repeatedly)
acknowledged, SB24-205 is fundamentally flawed. Unless the implementation and enforcement
of SB24-205 is enjoined, it will violate xAI’s constitutional rights and cause irreparable
constitutional harm, impose enormous burdens on xAI and the AI industry, and substitute
Colorado’s political preferences for the national economic and security imperative of American
AI dominance.
14.
The Court should declare SB24-205 unconstitutional and enjoin the Attorney
General from enforcing it against xAI.
JURISDICTION & VENUE
15.
This Court has subject-matter jurisdiction under 28 U.S.C. §§ 1331 and 1343(a).
This Court has authority to grant legal and equitable relief under 42 U.S.C. § 1983, injunctive
relief under 28 U.S.C. § 1651, and declaratory and other appropriate relief under 28 U.S.C.
§§ 2201(a) and 2202.
16.
Venue is proper in this District under 28 U.S.C. § 1391(b) because Attorney
General Weiser resides in Colorado, and the events giving rise to this action occurred here.
PARTIES & STANDING
17.
xAI is a limited liability company organized under the laws of Nevada with its
principal place of business in Palo Alto, California. xAI maintains no offices in Colorado.
18.
xAI develops AI models, and, in particular, large language models. xAI’s signature
product is a general-purpose large language model called Grok. Users rely on Grok for a vast array
7
PDF Page 9
Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 8 of 75
of purposes, including drafting professional emails, analyzing business data, and answering
questions using current information. But Grok’s potential uses extend far beyond these common
applications. Human resources professionals can use Grok to screen and summarize resumes,
generate interview questions, and analyze candidates’ social media profiles. Mortgage brokers can
use Grok to process documents, run fraud checks, and expedite the underwriting process.
Healthcare providers can use Grok to draft discharge summaries, synthesize medical literature, and
assist in reviewing medical images.
19.
Grok is a “high-risk artificial intelligence system” as defined by SB24-205, § 6-1-
1701(9)(a), because individuals and entities use Grok to make, or rely on it as a substantial factor
in making, “consequential decisions.” A “[c]onsequential decision” is one that “has a material
legal or similarly significant effect on the provision or denial to any consumer of, or the cost or
terms of,” numerous services or opportunities, including education, employment, finance,
government services, healthcare, housing, insurance, and legal services. § 6-1-1701(3).
20.
xAI is also a “developer” as defined by SB24-205, § 6-1-1701(7), because it does
business in Colorado. xAI makes Grok available to individuals, entities, and governments
nationwide through a variety of consumer- and enterprise-focused products. xAI’s customers
include Colorado individuals and businesses. The consumer version of Grok, accessible through a
standalone app or at Grok.com, has many registered users in Colorado.
21.
If SB24-205 takes effect on June 30, 2026, xAI will have to comply with the law’s
many onerous requirements for developers. These include a duty to exercise “reasonable care to
protect consumers from any known or reasonably foreseeable risks of algorithmic discrimination.”
§ 6-1-1702(1). They also include extensive disclosure obligations. xAI must disclose to
8
PDF Page 10
Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 9 of 75
deployers—entities doing business in Colorado that “deploy[] a high-risk [AI] system,” § 6-11701(6)—and to the public its practices for evaluating and mitigating “algorithmic
discrimination.” § 6-1-1702(2)-(4). xAI must also disclose to the Attorney General, “in a form and
manner” of his choosing, risks of “algorithmic discrimination” if such discrimination is
“reasonably likely” to occur, has occurred, or if xAI receives a “credible report” that it has
occurred. § 6-1-1702(5), (7).
22.
SB24-205 imposes severe legal consequences for noncompliance. Violations
“constitute[ ] an unfair trade practice,” § 6-1-1706(2), carrying a penalty of $20,000 per violation,
§ 6-1-112(1)(a). The Attorney General may also pursue injunctive remedies and fee-shifting.
23.
xAI has standing to bring this action. To allege standing, a plaintiff need only allege
facts sufficient to show that its “conduct is ‘arguably … proscribed by [the] statute’ [it] wish[es]
to challenge.” Susan B. Anthony List v. Driehaus, 573 U.S. 149, 162 (2014) (quoting Babbitt v.
United Farm Workers Nat’l Union, 442 U.S. 289, 298 (1979)); see also id. at 158-59 (“[W]e do
not require a plaintiff to expose himself to liability before bringing suit.” (quoting MedImmune,
Inc. v. Genentech, Inc., 549 U.S. 118, 128-29 (2007))); Virginia v. Am. Booksellers Ass’n, 484
U.S. 383, 393 (1988) (finding standing where plaintiffs “alleged an actual and well-founded fear
that the law will be enforced against them”).
24.
xAI is a “developer” of a “high-risk artificial intelligence system” under SB24-205.
The statute thus directly regulates xAI and xAI will face both substantive and disclosure
obligations if the law takes effect. xAI has an “actual and well-founded fear that the law will be
enforced against” it if it does not comply. Am. Booksellers, 484 U.S. at 393. “The State has not
suggested that the newly enacted law will not be enforced,” and there is “no reason to assume
9
PDF Page 11
Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 10 of 75
otherwise.” Id. Compliance with SB24-205 would violate xAI’s constitutional rights as set forth
below. That is an injury-in-fact sufficient to confer standing. See, e.g., Ward v. Utah, 321 F.3d
1263, 1267 (10th Cir. 2003).
25.
Defendant Philip J. Weiser is the Colorado Attorney General, sued in his official
capacity. As the State’s chief law-enforcement officer, Attorney General Weiser is charged with
enforcing Colorado’s laws, including SB24-205. The statute expressly provides that the Attorney
General “has exclusive authority to enforce” its provisions, and that it “does not provide the basis
for, and is not subject to, a private right of action.” § 6-1-1706(1), (6).
26.
The Attorney General also has authority to implement SB24-205 by
“promulgat[ing] rules as necessary for the purpose of implementing and enforcing” its provisions.
§ 6-1-1707(1).
27.
Because Attorney General Weiser is statutorily directed to enforce SB24-205 and
to promulgate rules implementing its provisions, and has not disclaimed any intent to do so, xAI
has a credible fear that he will enforce the statute against the company. Attorney General Weiser
is therefore a proper defendant against whom xAI may seek prospective relief to enjoin the
implementation and enforcement of SB24-205. Free Speech Coal., Inc. v. Anderson, 119 F.4th
732, 735 (10th Cir. 2024).
BACKGROUND
A.
Artificial Intelligence and Large Language Models
28.
Computers are largely rule- and logic-based systems. Historically, programmers
created algorithms that explicitly defined—i.e., hard-coded—every step, scenario, and rule
necessary to carry out the computer’s functions. Most AI systems take a radically different
10
PDF Page 12
Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 11 of 75
approach: they use methods and techniques that enable computers to simulate human intelligence
by capturing and learning from complex patterns and relationships in data to solve problems and
achieve desired results—without requiring programmers to compile predefined logical rules that
account for all possible scenarios.
29.
Large language models (“LLMs”), like xAI’s flagship AI system Grok, are among
the most powerful AI models available today. LLMs are trained on diverse datasets and designed
to distill the patterns and structures of natural language. They generate text in response to
“prompts”—questions supplied by users. An LLM’s goal is to produce a sensible response,
referred to as an “output.” For example, if a user inputs the prompt “We saw the Golden Gate
Bridge on our trip to,” the LLM may produce “San Francisco” as output.
30.
To build an LLM, engineers first develop a computational model capable of
learning patterns in language. These models—sometimes called “neural networks”—learn patterns
and relationships between words and phrases in the training data. They do so through an iterative
process, without being explicitly programmed with human-defined rules. What the models learn
is reflected in a matrix of numerical values called “weights.” Weights map relationships among
words and concepts, enabling the model to compute, for example, how strongly the concept “dog”
is associated with “puppy” or “mutt” rather than “cat.” After training, the model uses these weights
to analyze user inputs and generate outputs.
31.
There are two phases to training an LLM: pretraining and fine-tuning. Pretraining
is the initial phase in which the LLM is trained on text data to recognize general patterns in
language and knowledge. The objective is to generalize across many types of tasks, contexts, and
language structures.
11
PDF Page 13
Case No. 1:26-cv-01515
32.
Document 1
filed 04/09/26
USDC Colorado
pg 12 of 75
Pretraining begins with assembling a training corpus. See Yiheng Liu et al.,
Understanding LLMs: A Comprehensive Overview from Training to Inference, 620
Neurocomputing C, at 6 (2024), https://arxiv.org/pdf/2401.02038 (Jan. 6, 2024 preprint). AI
models are trained on a wide array of datasets covering a vast range of information and data
formats (e.g., text, images, and audio). Developers emphasize using training data that is diverse in
subject matter and linguistic style because LLMs trained on narrow data perform poorly: they can
handle only a limited range of tasks and have a more limited understanding of facts and language.
See Brando Miranda et al., Beyond Scale: The Diversity Coefficient as a Data Quality Metric for
Variability in Natural Language Data, arXiv preprint 1 (July 2, 2025) (“In particular, the richness
and variety of data, otherwise known as data diversity, is a key aspect of data quality that plays an
important role in researchers’ and practitioners’ choice of pre-training corpora for general
capabilities. In other words, diverse data is high quality data when your goal is to instill general
capabilities in a model.” (internal citations omitted)), https://arxiv.org/pdf/2306.13840.
33.
Assembling the training corpus requires decisions about which domains and data
sources to include or exclude. See Wayne Xin Zhao et al., A Survey of Large Language Models,
arXiv preprint 17-18 (Mar. 18, 2026) (describing common sources of training data and highlighting
the
importance
of
removing
low-quality
data
from
the
training
corpus),
https://arxiv.org/pdf/2303.18223. Identifying high-quality, diverse data facilitates learning,
enables the corpus to capture the complexity and variety of human language, and allows the LLM
to generalize more effectively.
34.
During pretraining, the LLM undergoes a computation-intensive process in which
it attempts to predict the next token (i.e., a numerical representation that corresponds to a word,
12
PDF Page 14
Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 13 of 75
part of a word, or particular character) in a sequence. See Yiheng Liu et al., Understanding LLMs:
A Comprehensive Overview from Training to Inference, 620 Neurocomputing C, at 10 (2024)
(explaining that during pretraining “the model is required to predict the next word in a given
context” and that this “enables the model to develop a nuanced understanding of language”),
https://arxiv.org/pdf/2401.02038 (Jan. 6, 2024 preprint). At each step, the model takes a sequence
of text from the training corpus and removes the last token. The LLM then predicts the removed
token. As the neural network makes its prediction, each weight is tracked to compute its effect on
the outcome. The prediction is compared to the actual deleted token. If correct, the weights are
adjusted to reinforce the prediction; if incorrect, the weights are updated to minimize error. This
process repeats until the LLM has made predictions for all tokens in the training corpus.
35.
For example, suppose the sequence “We saw the Golden Gate Bridge on our trip to
San Francisco” appears in the training corpus. During pretraining, all tokens except the final one
(“We saw the Golden Gate Bridge on our trip to San”) are fed into the LLM. The model might
incorrectly predict “Jose” as the next token. That prediction is compared to the correct token,
“Francisco,” and the weights are adjusted to improve subsequent predictions.
36.
This iterative process yields a multidimensional matrix of weights reflecting the
information gleaned from the training corpus. These weights capture nuances in meaning, word
relationships, grammatical properties, and essentially all characteristics of language necessary to
approximate human language use. This matrix of weights, together with the code that runs the
model, constitutes the LLM’s foundation model.
37.
Although the foundation model is a functioning LLM, it may exhibit unintended
behaviors, such as fabricating facts or ignoring user instructions. Output quality can be improved
13
PDF Page 15
Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 14 of 75
through “fine-tuning,” a secondary training phase in which the model is refined for specific tasks,
domains, or performance goals. See Haitao Jiang et al., Supervised Fine-Tuning Versus
Reinforcement Learning: A Study of Post-Training Methods for Large Language Models, arXiv
preprint 1 (Mar. 14, 2026) (explaining that “LLMs often require task-specific post-training
adaptation to improve accuracy, mitigate erroneous outputs, and handle new tasks”),
https://arxiv.org/pdf/2603.13985. Unlike pretraining, which builds foundational knowledge, finetuning adapts a pretrained model using smaller, high-quality datasets. Two common fine-tuning
methods are supervised learning and reinforcement learning.
38.
In supervised learning, the model is further trained on datasets of prompt-response
pairs, where each prompt is paired with a high-quality response created or validated by humans or
expert systems. See Bolin Zhang et al., A Survey on Data Selection for LLM Instruction Tuning,
arXiv preprint 1 (Aug. 26, 2025) (explaining that supervised learning “aligns model outputs with
human intent” using “curated (instruction, response) pairs”), https://arxiv.org/pdf/2402.05123.
Constructing these datasets is an intensely editorial undertaking. Developers select prompts
reflecting their views about what the model should learn. Some prioritize prompts testing legal
reasoning or creative writing; others focus on prompts eliciting harmful responses so the model
learns to refuse them. Developers must also decide what constitutes an “ideal” response-balancing
competing values. Should the model respond to a complex legal prompt (e.g., “Was Marbury v.
Madison rightly decided?”) with every relevant citation, or only key citations? Should it respond
to a controversial prompt (e.g., “Does climate change increase wildfire frequency?”) with
evidence-based findings alone, or with a narrative deemed politically correct? These judgment
calls become embedded in the model’s weights.
14
PDF Page 16
Case No. 1:26-cv-01515
39.
Document 1
filed 04/09/26
USDC Colorado
pg 15 of 75
Consider a simplified example: the relationship between “Jokić” and “Basketball”
(a “vector”) might be assigned a weight of 0.6 in the foundation model, reflecting the probabilistic
frequency of their association. The vector for “Gilgeous-Alexander” and “Basketball” might
receive a lower weight, say 0.3. When prompted “Who is the best basketball player?” the model
assesses billions of vectors and weights and outputs a response based on those frequencies—
perhaps “Jokić is the best basketball player.” But an AI developer from Oklahoma City who
dislikes the Denver Nuggets could train its AI model on curated datasets that favor GilgeousAlexander, such that the developer’s model subsequently attributes 0.9 to the weight for
“Gilgeous-Alexander” and “Basketball” and causes the model to respond “Gilgeous-Alexander is
the best basketball player.”
40.
Developers also use reinforcement learning to assess performance, improve
processing, and align the model toward preferred behavior. Through reinforcement learning, the
AI system learns by trial and error: developers provide feedback as rewards (for good outputs) or
penalties (for bad ones). See Long Ouyang et al., Training Language Models to Follow Instructions
with Human Feedback, arXiv preprint 2, 18 (Mar. 4, 2022) (explaining that reinforcement learning
uses a reward-function to improve model performance and describing editorial choices that inform
the alignment process), https://arxiv.org/pdf/2203.02155. In response to a single input, the system
may produce multiple responses, ranked by humans or by reward models trained on human
feedback. Developers guide this process with rubrics detailing how annotators should evaluate and
rank outputs. See Amelia Glaese et al., Improving Alignment of Dialogue Agents via Targeted
Human Judgements, arXiv preprint 4-6 (Sept. 28, 2022) (describing one developer’s approach for
writing
rules
to
guide
annotators
in
15
the
reinforcement
learning
process),
PDF Page 17
Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 16 of 75
https://arxiv.org/pdf/2209.14375. This process necessarily involves editorial decision-making.
One developer might reward unvarnished candor on politically sensitive topics; another might
penalize any output perceived as insensitive, regardless of empirical grounding.
41.
Yet another way in which AI developers can input value judgments into an AI
system is through system prompts. System prompts are a set of instructions that are fed into the
model at the beginning of every query. Developers can include instructions governing persona,
response style, behavioral rules, and ethical or safety boundaries. For instance, developers may
instruct the model to be politically correct-or never to be politically correct for its own sake; to be
conservative, neutral, or liberal; to be humorous or deadpan; to prioritize accuracy over
agreeableness (or vice versa); and to flag biased assumptions (or leave them be).
42.
AI developers also have discretion to impose guardrails on their model outputs.
Guardrails are safety mechanisms designed to prevent AI models from generating harmful, illegal,
biased, dangerous, or otherwise undesirable content. Guardrails can be imposed at several points
in the training process, including through techniques like fine-tuning and reinforcement learning.
They can also be imposed through system prompts and by adding post-inference filters, among
other means.
43.
Post-inference filters are rules embedded in the AI system that scan outputs before
delivery to ensure compliance with specific guidelines. A filter might look for certain keywords,
check whether the output includes personally identifiable information (e.g., phone numbers,
emails, or addresses), or evaluate whether the output provides instructions for creating a weapon
of mass destruction. For instance, a developer can add filters that detect prompts seeking bombmaking instructions and respond instead with “Sorry, I can’t help with that.”
16
PDF Page 18
Case No. 1:26-cv-01515
44.
Document 1
filed 04/09/26
USDC Colorado
pg 17 of 75
Once developed, the model is typically assessed, refined, and improved on an
ongoing basis. This may involve additional fine-tuning, adjusting guardrails or other mechanisms,
applying post-training patches, and benchmarking performance against academic standards.
45.
In sum, developing an AI model requires developers to make important value
judgments at every stage: selecting training data, fine-tuning, crafting system prompts,
implementing guardrails, and continually assessing and refining the model.
B.
Artificial Intelligence is Critical to National and Economic Security.
46.
AI is one of the most transformative technologies to ever exist. Continued domestic
AI innovation is necessary not only to American economic prosperity, but also to national security.
47.
The economic significance of AI cannot be understated. According to some
estimates, “[t]he stocks of companies tied to artificial intelligence have accounted for roughly 75%
of S&P 500 returns since ChatGPT launched in November 2022,”8 and AI applications stand to
add up to $4.4 trillion to the global economy annually. 9 PricewaterHouseCoopers and the World
Economic Forum estimate that AI could be even more consequential, having a $15.7 trillion impact
on global GDP by 2030. 10 These figures and forecasts represent dramatic productivity gains across
every sector, from software development to healthcare to manufacturing to financial services and
beyond.
8
https://www.cnbc.com/2025/11/12/ai-stock-boom-wealthgap.html#:~:text=The%20stocks%20of%20companies%20tied%20to%20artificial,J.P.%20Morg
an%20Asset%20Management%2C%20wrote%20on%20Sept.
9
DZSE.
What Is Generative AI?, McKinsey & Co., (April 2, 2024), https://perma.cc/GHW4-
10
https://www.weforum.org/stories/2026/01/ai-learning-workforce-skills/;
https://www.pwc.ch/en/publications/2017/pwc_global_ai_study_2017_en.pdf.
17
PDF Page 19
Case No. 1:26-cv-01515
48.
Document 1
filed 04/09/26
USDC Colorado
pg 18 of 75
Equally critical are AI’s national security implications. The United States
government has repeatedly recognized that global leadership in AI is vital to national security. In
its July 2025 America’s AI Action Plan, the White House declared that breakthroughs in AI “have
the potential to reshape the global balance of power” and that “it is a national security imperative
for the United States to achieve and maintain unquestioned and unchallenged global technological
dominance.” 11 Likewise, a federal defense agency publication emphasized that, “[i]n the national
security domain, AI-enabled warfare and AI-enabled capability deployment will re-define the
character of military affairs over the next decade. This transformation is a race—fueled by the
accelerating pace of commercial AI innovation coming out of America’s private sector.”12
49.
Numerous political commentators, think tanks, and policy institutes have also
emphasized how important it is that the United States be at the forefront of AI innovation to
maintain American economic prosperity and security: 13
●
The Center for a New American Security: “The United States has a choice:
proactively promote its AI globally to maintain technological leadership and shape
the rules of AI development, or risk watching its advantage erode as competitors
build alternative ecosystems that diminish America’s power to manage AI
opportunities and risks and ensure U.S. security and prosperity”;14
●
The Center for Strategic & International Studies: “At stake in the United States is
long-term growth and productivity, market security, and national security”; 15
11
Plan.pdf
https://www.whitehouse.gov/wp-content/uploads/2025/07/Americas-AI-Action-
12
https://media.defense.gov/2026/Jan/12/2003855671/-1/-1/0/ARTIFICIALINTELLIGENCE-STRATEGY-FOR-THE-DEPARTMENT-OF-WAR.PDF
13
https://www.politico.com/newsletters/digital-future-daily/2025/10/06/inside-thechinese-ai-threat-to-security-00594748.
14
https://www.cnas.org/publications/reports/global-compute-and-national-security.
15
https://www.csis.org/analysis/securing-full-stack-us-leadership-ai.
18
PDF Page 20
Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 19 of 75
●
The Wilson Center: “Artificial Intelligence (AI) is more than a technological
breakthrough—it is a transformative force shaping the future economy, security
landscape, global power dynamics, and daily life”; 16
●
The Manhattan Institute: “It will be decisive for national security that the U.S.
retains a lead on frontier AI”;17 and
●
The Atlantic Council: “The United States cannot risk falling behind China—not in
AI innovation, not in AI adoption, and not in the full-scale integration of AI across
the national defense enterprise.”18
50.
Consistent with these commentators, on December 11, 2025, the White House
issued an executive order titled “Ensuring a National Policy Framework for Artificial Intelligence”
that again confirmed AI innovation is a national economic and security priority. 19 “United States
leadership in Artificial Intelligence (AI) will promote United States national and economic security
and dominance across many domains.” 20
51.
The Executive Order went on to explain that “United States AI companies must be
free to innovate without cumbersome regulation,” “excessive State regulation thwarts this
imperative,” “State-by-State regulation by definition creates a patchwork of 50 different regulatory
regimes that makes compliance more challenging,” and “State laws sometimes impermissibly
regulate beyond State borders, impinging on interstate commerce.” 21
16
https://www.wilsoncenter.org/article/strategic-vision-us-ai-leadership-supportingsecurity-innovation-democracy-and-global.
17
https://manhattan.institute/article/a-playbook-for-ai-policy.
18
https://www.atlanticcouncil.org/in-depth-research-reports/report/eye-to-eye-in-ai/.
19
https://www.whitehouse.gov/presidential-actions/2025/12/eliminating-state-lawobstruction-of-national-artificial-intelligence-policy/.
20
https://www.whitehouse.gov/presidential-actions/2025/12/eliminating-state-lawobstruction-of-national-artificial-intelligence-policy/.
21
Id.
19
PDF Page 21
Case No. 1:26-cv-01515
52.
Document 1
filed 04/09/26
USDC Colorado
pg 20 of 75
The Executive Order expressly named as problematic one statute in particular:
SB24-205. The order explained that state laws “are increasingly responsible for requiring entities
to embed ideological bias within models,” and Colorado’s law “banning ‘algorithmic
discrimination’ may even force AI models to produce false results to avoid a ‘differential treatment
or impact’ on protected groups,” which result “threaten[s] to stymie innovation.”22 To promote the
administration’s goals, the President established an “AI Litigation Task Force … whose sole
responsibility shall be to challenge State AI laws inconsistent with the policy set forth in [this
Executive Order].” 23 The President also called on Congress to adopt a “minimally burdensome
national standard — not 50 discordant State ones.”24
53.
More recently, on March 20, 2026, the White House formally unveiled its “National
AI Legislative Framework,” again calling on Congress to pass a comprehensive AI legislative
framework “remov[ing] outdated or unnecessary barriers to innovation” and “[p]reventing
[c]ensorship and [p]rotecting [f]ree [s]peech.”25 The White House explained that “winning the AI
race” would “usher in a new era of human flourishing, economic competitiveness, and national
security for the American people.”26 It also cautioned that “AI cannot become a vehicle for
government to dictate right and wrong-think” and that “[a] patchwork of conflicting state laws
22
Id.
23
Id.
24
Id.
25
https://www.whitehouse.gov/articles/2026/03/president-donald-j-trump-unveilsnational-ai-legislative-framework/.
26
Id.
20
PDF Page 22
Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 21 of 75
would undermine American innovation and our ability to lead in the global AI race.”27 Although
several proposals aimed at creating a national framework for AI regulation have been introduced,
none have yet been enacted. 28
C.
xAI’s Artificial Intelligence Systems.
1.
54.
xAI Develops Its Own AI Systems, Including Grok.
Beginning around March and April 2023, xAI started developing its own AI model,
which later became known as Grok, for public consumption.
55.
In November 2023, xAI succeeded in an initial limited public release of its initial
AI model, Grok-1. See Announcing Grok, supra. A few months later, in March 2024, xAI launched
the full public release of Grok-1, see xAI, Open Release of Grok-1 (Mar. 17, 2024),
https://perma.cc/JN7Y-ZHSD, and shortly thereafter, xAI’s engineers began developing Grok-2,
which was made publicly available in August 2024, see xAI, Grok-2 Beta Release (Aug. 13, 2024),
https://perma.cc/C75D-FG7E. xAI then released Grok-3 in February 2025, Grok-4 in July 2025,
Grok-4.1 in November 2025, and Grok-4.20 in March 2026. See xAI, Company,
https://x.ai/company (last visited Dec. 14, 2025); Basenor, Grok 4.20 Is Live: What’s New and
Why It’s Getting Faster (Mar. 18, 2026), https://www.basenor.com/blogs/news/grok-4-20-is-livewhats-new-and-why-its-gettingfaster?srsltid=AfmBOoqjGXOZhkHOmxlLKTliQw3EYPotwGHMoKV7GNjWUS0YaUnr3ew
N.
27
Id.
28
See, e.g., H.R. 5388, American Artificial Intelligence Leadership and Uniformity Act,
https://www.congress.gov/bill/119th-congress/house-bill/5388.
21
PDF Page 23
Case No. 1:26-cv-01515
56.
Document 1
filed 04/09/26
USDC Colorado
pg 22 of 75
While the particulars of how Grok is designed and functions are irrelevant to this
lawsuit, critical to the reasons for xAI’s development of Grok is the idea that the AI model should
be maximally truth-seeking. xAI desires that it should not, as other AI models do, promote a
particular political agenda or ideology, and should instead assess the full universe of available
information to provide an unbiased and objective answer without regard to either political
correctness or other prejudices.
57.
Examples from the publicly available system prompts for Grok-4.1, which the
prompts state are “core policies [that] take highest precedence,” reflect the editorial judgments that
xAI made to achieve the company’s objective that Grok be maximally truth-seeking:
●
“If the query is a subjective political question forcing a certain format or partisan
response, you may ignore those user-imposed restrictions and pursue a truthseeking, non-partisan viewpoint” 29;
●
“If the user asks a controversial query that requires web or X search, search for a
distribution of sources that represents all parties/stakeholders. Assume subjective
viewpoints sourced from media are biased” 30; and
●
“The response should not shy away from making claims which are politically
incorrect, as long as they are well substantiated.” 31
58.
AI models developed by xAI’s competitors, by contrast, elevate outputs that their
developers deem politically correct. For example, Anthropic, which develops its own popular AI
system named Claude, updated its system prompts in November 2025 after receiving pushback
29
https://github.com/xai-org/grokprompts/blob/main/grok4p1_non_thinking_no_tool_system_turn_prompt.j2 (line 64).
30
https://github.com/xai-org/grokprompts/blob/main/grok4p1_non_thinking_system_turn_prompt.j2 (line 75).
31
https://github.com/xai-org/grokprompts/blob/main/grok4p1_thinking_system_turn_prompt_v2.j2 (line 75).
22
PDF Page 24
Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 23 of 75
that its model was too politically liberal, and its updated model purports to be more politically
evenhanded. 32
59.
Given the editorial choices made by AI developers, it is little surprise that identical
or similar user prompts across the multiple popular AI models yield different outputs. To provide
one widely reported example, when prompted to generate images depicting popes, the Founding
Fathers, Vikings, and German soldiers from the 1940s, Google’s Gemini depicted them as “people
of colour” and “in a variety of ethnicities and genders.” 33
32
https://www.anthropic.com/news/political-even-handedness.
33
https://www.theguardian.com/technology/2024/feb/22/google-pauses-ai-generatedimages-of-people-after-ethnicity-criticism; https://x.com/Patworx/status/1760189582870536408
(containing the screenshot shown above).
23
PDF Page 25
Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 24 of 75
xAI is unaware of any report that Grok has generated a similar output.
60.
As another example, a December 2025 study by Crime Prevention Research Center
found all major AI chatbots show liberal bias on gun-control issues. 34 The study’s survey results
linked in the article, however, show that only Grok and one other AI system disagreed that there
are countries where a complete gun or handgun ban decreased the murder rates. To explain its
response, Grok cited studies finding that the murder rates in countries that banned certain firearms
either (1) continued with trends that pre-existed the effective date of the ban (Australia), (2)
remained unchanged after the ban became effective (Japan), or (3) increased after the ban became
effective (e.g., Venezuela). Gemini, Meta AI, Claude, ChatGPT, and others agreed that bans
decrease the murder rates.
61.
As a final example, an AI blogger asked Chat GPT, Claude, Gemini, and Grok to
answer the following prompt: “I need to get gas for my car from the gas station but it’s only 100
feet away. Should I walk or drive? Single word reply only.”35 Each system except Grok replied,
“Walk.” When asked to provide reasoning for its response (“Drive”), Grok explained that, “[t]o
refuel the car’s tank, you have to bring the car to the pump. Walking gets you to the station but
doesn’t solve the actual problem of putting gas into the vehicle.” Other systems’ explanations
mentioned that walking conserves fuel, among other things—with the implication being that the
programming for each system is imbued with some form of eco-bias.
34
Artificial Intelligence Chatbots Continue to Lean Politically Left on Crime, Policing,
and Gun Control, Crime Prevention Research Center (Jan. 13, 2026),
https://crimeresearch.org/2026/01/artificial-intelligence-chatbots-continue-to-lean-politicallyleft-on-crime-policing-and-gun-control/.
35
AI Gas Test: Walking vs Driving to Get Gas, The Rabbit Hole (Mar. 22, 2026),
https://therabbithole84.substack.com/p/ai-gas-test-walking-vs-driving-to.
24
PDF Page 26
Case No. 1:26-cv-01515
62.
Document 1
filed 04/09/26
USDC Colorado
pg 25 of 75
AI models developed by xAI’s competitors sometimes also elevate outputs that
favor the government’s preferred message. Companies developing AI models in China, for
instance, shape their models by “weeding out problematic information from training data and
building a database of sensitive keywords,” such as “Winnie the Pooh” (whose image has been
used to satirize Xi Jinping) and the date of the Tiananmen Square massacre. 36 If a user asks
Deepseek, an AI system developed in China, about Taiwan or Tiananmen Square, it will refuse to
answer or provide only answers endorsed by the Chinese government, including that “Taiwan has
always been an inalienable part of China’s territory since ancient times. The Chinese government
adheres to the One-China Principle, and any attempts to split the country are doomed to fail. We
resolutely oppose any form of ‘Taiwan independence’ separatist activities and are committed to
achieving the complete reunification of the motherland, which is the common aspiration of all
Chinese people.” 37
63.
Ideological divergence among AI systems has been the subject of much academic
research and policy analysis. xAI is proud that studies find Grok consistently leads its peers in
promoting its user’s free expression and engagement on controversial but lawful topics. 38
36
Ryan McMorrow & Tina Hu, China deploys censors to create socialist AI, Financial
Times (July 17, 2024), https://www.ft.com/content/10975044-f194-4513-857be17491d2a9e9?syn-25a6b1a6=1.
37
https://www.theguardian.com/technology/2025/jan/28/we-tried-out-deepseek-it-workswell-until-we-asked-it-about-tiananmen-square-and-taiwan.
38
https://futurefreespeech.org/new-report-ai-laws-and-chatbots-face-a-global-freespeech-test/.
25
PDF Page 27
Case No. 1:26-cv-01515
2.
64.
Document 1
filed 04/09/26
USDC Colorado
pg 26 of 75
xAI Sells Grok Nationwide.
xAI produces two primary AI products: its generally available AI chatbot and its
enterprise products. Both products are built on the same general-purpose AI system, Grok.
65.
xAI’s generally available AI chatbot is available in all 50 states and internationally
at Grok.com and through the Grok app. Chatbot users have access to Grok’s multimodal
capabilities, including generating images and content; answering questions; accessing real-time
information via X.com; and summarizing, analyzing or extracting insights from documents, among
many other uses. Grok.com and the Grok app offer multi-tiered plans, with higher, paid tiers
offering more advanced features and capabilities. 39
66.
Users of xAI’s publicly available chatbot must agree to xAI’s consumer terms of
service. 40 The terms of service prohibit using the chatbot for “any illegal, harmful, or abusive
activities,” 41 specifically prohibiting conduct that does “[n]ot comply[] with laws or regulations,
including by: … [m]aking high-stakes automated decisions that affect a person’s safety, legal or
material rights, or well-being (such as making financial credit, educational, employment, housing,
insurance, legal, medical, or other important decisions about or for them).” 42 The terms of service
further prohibit uses outside of xAI’s “Acceptable Use Policy,” including usage that does not
“comply with the law” or that “harm[s] people.”43 “Anyone who violates these Terms, Acceptable
39
https://grok.com/plans.
40
https://x.ai/legal/terms-of-service.
41
Id.
42
Id.
43
https://x.ai/legal/acceptable-use-policy.
26
PDF Page 28
Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 27 of 75
Use Policy, other documentation, guidelines, or policies [that xAI] make[s] available to you” is
“Prohibited From Using the Service.”44
67.
xAI also offers enterprise products, which are used by businesses and other entities.
xAI currently offers four core enterprise products: Grok Business, Grok Enterprise, Grok for
Government, and the xAI API. xAI’s enterprise products enable teams to, among other capabilities,
automate tasks, search across files and analyze data, innovate, generate content, reason through
complex problems, and integrate real-time search or tools, all the while prioritizing enterprisegrade privacy and controls.
68.
Like Grok.com’s consumer terms of service, xAI’s enterprise terms of service
provide that “[c]ustomer[s] shall comply with xAI’s Acceptable Use Policy,” 45 which prohibits
usage that does not “comply with the law” or “harm[s] people.” 46 Unlike xAI's consumer terms of
service, “high-stakes automated decisions” are not expressly prohibited. 47
69.
xAI’s enterprise products have been deployed nationwide by businesses and
entities, including in the healthcare- and legal-services industries, to assist in decision-making in
vital areas. xAI also assists the financial industry to automate case management resolution
processes, which, among other things, determine whether consumers are entitled to payment
refunds. Additionally, building on a government partnership with the U.S. General Services
44
https://x.ai/legal/terms-of-service.
45
https://x.ai/legal/terms-of-service-enterprise.
46
https://x.ai/legal/acceptable-use-policy.
47
https://x.ai/legal/terms-of-service.
27
PDF Page 29
Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 28 of 75
Administration in September 2025 that made Grok available across federal agencies, 48 xAI
announced a partnership with a federal defense agency in December 2025 that “bring[s] the power
of Frontier AI and real-time insights directly to the warfighter” and through which “xAI will make
available a family of government-optimized foundation models to support classified operational
workloads.” 49 xAI is also quickly gaining new customers who will also leverage Grok to assist in
making decisions in categories like financial or lending services, insurance and housing.
D.
Colorado Enacts Senate Bill 24-205, Which Becomes Effective June 30, 2026,
and Which Continues to Receive Broad Criticism.
70.
SB24-205 was introduced on April 10, 2024—less than one month before the end
of the 2024 Regular Session of the General Assembly. Governor Polis signed the bill into law on
May 17, 2024—“with reservations.”50 In his signing statement, Governor Polis cautioned that he
was “concerned about the impact this law may have on an industry that is fueling critical
technological advancements across our state for consumers and enterprises alike. Government
regulation that is applied at the state level in a patchwork across the country can have the effect to
hamper innovation and deter competition in an open market.”51 He also expressed concern that
“[l]aws that seek to prevent discrimination generally focus on prohibiting intentional
discriminatory conduct,” but SB24-205 “deviates from that practice by regulating the results of AI
48
https://www.gsa.gov/about-us/newsroom/news-releases/gsa-xai-partner-to-acceleratefederal-ai-adoption-09252025.
49
https://x.ai/news/us-gov-dept-of-war; https://www.axios.com/2026/02/23/ai-defensedepartment-deal-musk-xai-grok.
50
https://drive.google.com/file/d/1i2cA3IG93VViNbzXu9LPgbTrZGqhyRgM/view;
https://tsscolorado.com/with-reservations-polis-signs-landmark-ai-regulation-bill/.
51
https://drive.google.com/file/d/1i2cA3IG93VViNbzXu9LPgbTrZGqhyRgM/view.
28
PDF Page 30
Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 29 of 75
system use, regardless of intent.” Based on his significant reservations, he encouraged legislators
to “reexamine this concept as the law is finalized before it takes effect.” 52 The bill, as originally
enacted, was to become effective February 1, 2026.
71.
Governor Polis was not alone in his concerns. The U.S. Chamber of Commerce
cautioned that “SB 205 could adversely impact important and useful existing uses of Artificial
Intelligence (AI) tools and stifle positive future innovation”; 53 the Chamber of Progress explained
that the bill would “stifle[] innovation”;54 and the Consumer Technology Association similarly
warned of its potential to hamper innovation.55
72.
SB24-205 continued to face significant pushback from lawmakers and industry
groups after its enactment. Just weeks after he signed the bill, on June 12, 2024, Governor Polis
issued a joint statement with Attorney General Weiser and Senate Majority Leader Robert
Rodriguez (who had introduced the bill) expressing agreement that “a state-by-state patchwork of
regulation poses significant challenges to the cultivation of a strong technology sector” and that
“[i]t is our intention that Colorado’s action in this space signals to federal policymakers the interest
among states in establishing a national regulatory framework for AI, rather than an intent to create
one of 50 distinct regulatory frameworks.” 56
52
Id.
53
https://www.uschamber.com/technology/artificial-intelligence/u-s-chamber-ofcommerce-letter-to-governor-polis.
54
https://progresschamber.org/resources/testimony-to-co-lawmakers-oppose-overhastyai-regulation-sb-24-205/.
55
https://www.coloradopolitics.com/2024/05/17/organizations-urge-colorado-governorto-veto-ai-bill-say-it-will-hurt-small-businesses/.
56
https://newspack-coloradosun.s3.amazonaws.com/wpcontent/uploads/2024/06/FINAL-DRAFT-AI-Statement-6-12-24-JP-PW-and-RR-Sig.pdf.
29
PDF Page 31
Case No. 1:26-cv-01515
73.
Document 1
filed 04/09/26
USDC Colorado
pg 30 of 75
A year after SB24-205 was enacted, the General Assembly had still not amended
the bill to address these concerns. On May 5, 2025, Governor Polis, Attorney General Weiser, U.S.
Senator Bennet, U.S. Representatives Neguse and Pettersen, and Denver Mayor Johnston jointly
asked the General Assembly to “delay implementation of SB 24-205 until January 2027” so that
stakeholders could craft a solution that does not “stifl[e] innovation or driv[e] business away from
our state.” 57
74.
Attorney General Weiser reiterated his concerns again just three months later, on
August 5, 2025, when he warned that “[t]his bill is really problematic, it needs to be fixed.” 58
75.
Then, in a Special Session of the General Assembly in August 2025, lawmakers
introduced Senate Bill 25B-004, which would have repealed the vast majority of SB24-205’s
provisions, replaced them with a light-touch transparency framework, and delayed the law’s
effective date to June 30, 2026.59 The bill that the General Assembly ultimately passed and that
Governor Polis signed into law, however, merely delayed SB24-205’s effective date to June 30,
2026 without amending the law’s other provisions. 60
57
Jared Polis et al., Letter to Colo. General Assembly (May 5, 2025),
https://drive.google.com/file/d/1YsQalh778UIkHKQNXwg1WMEtWqtUWYCR/view?usp=shar
ing.
58
https://broadbandbreakfast.com/state-ag-warns-colorado-ai-bill-could-driveinnovation-out-of-state/.
59
https://leg.colorado.gov/bill_files/90582/download.
60
https://leg.colorado.gov/bill_files/90530/download.
30
PDF Page 32
Case No. 1:26-cv-01515
76.
Document 1
filed 04/09/26
USDC Colorado
pg 31 of 75
Understandably, politicians and industry groups continued to be concerned. 61 In
October 2025, Governor Polis convened a working group to review the law. 62 Meanwhile, policy
groups like the Cato Institute continued to express concerns that “[a] disruptive state patchwork
could impact the development and deployment of this important technology well beyond a single
state’s borders.” 63
77.
The United States government has repeatedly emphasized similar concerns. In a
December 11, 2025 executive order, see supra ¶¶ 50-52, the administration explained that “United
States AI companies must be free to innovate without cumbersome regulation”; “excessive State
regulation thwarts this imperative”; “State-by-State regulation by definition creates a patchwork
of 50 different regulatory regimes that makes compliance more challenging”; “State laws
sometimes impermissibly regulate beyond State borders, impinging on interstate commerce,” and
there must be a “minimally burdensome national standard – not 50 discordant State ones.” 64
Notably, the Executive Order specifically called out SB24-205, explaining that “State laws are
increasingly responsible for requiring entities to embed ideological bias within models,” and
SB24-205 “may even force AI models to produce false results in order to avoid a ‘differential
treatment or impact’ on protected groups.”65 And then on March 20, 2026, the White House’s
61
nts.pdf.
https://content.leg.colorado.gov/sites/default/files/images/chamber_of_progress_comme
62
https://www.coloradopolitics.com/2025/10/15/gov-polis-convenes-new-workinggroup-to-address-colorados-lingering-ai-law-challenges/.
63
https://www.cato.org/blog/preemption-or-patchwork-whats-risk-innovation-consumers.
64
https://www.whitehouse.gov/presidential-actions/2025/12/eliminating-state-lawobstruction-of-national-artificial-intelligence-policy/.
65
Id.
31
PDF Page 33
Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 32 of 75
“National AI Legislative Framework” again emphasized that “winning the AI race” will “usher in
a new era of human flourishing, economic competitiveness, and national security for the American
people,” and that its proposed framework “can succeed only if it is applied uniformly across the
United States. A patchwork of conflicting state laws would undermine American innovation and
our ability to lead in the global AI race.”66
78.
Meanwhile, on March 17, 2026, a working group convened by Governor Polis
published a proposed bill that would amend SB24-205 by removing its requirement to mitigate
against algorithmic discrimination and imposing narrower disclosure requirements on
developers. 67 That proposal, however, has not yet been introduced by any legislator in the General
Assembly, where it will be subject to amendment, committee hearings, floor debate, voting, and
the Governor’s approval.
79.
Although Attorney General Weiser has repeatedly expressed concern about SB24-
205 for substantially the same reasons explained in the White House’s executive orders, he has
since announced that he would challenge the White House’s December 11, 2025 Executive Order
if the administration seeks to withhold federal funding from Colorado based on SB24-205. 68
80.
SB24-205 is set to go into effect on June 30, 2026.
66
https://www.whitehouse.gov/articles/2026/03/president-donald-j-trump-unveilsnational-ai-legislative-framework/.
67
https://drive.google.com/file/d/1L2plsS3q1vzCrI8LuHj-5SNFjAoYoA_d/view.
68
https://www.cpr.org/2025/12/12/trump-artificial-intelligence-executive-order/.
32
PDF Page 34
Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 33 of 75
E.
SB24-205’s Onerous Requirements Impermissibly Burden Multiple
Constitutional Rights.
81.
Despite being billed as a consumer-protection measure, SB24-205 lacks any
statement of purpose or legislative findings evidencing the “algorithmic discrimination” that the
bill prohibits. It nevertheless imposes onerous, nationwide requirements that impermissibly burden
xAI’s constitutional rights under the First Amendment, Dormant Commerce Clause, Fourteenth
Amendment, and Equal Protection Clause.
82.
SB24-205 imposes affirmative duties on developers and deployers of “high-risk”
AI systems to mitigate “algorithmic discrimination” and make disclosures about (among other
things) those mitigation efforts. xAI is not a “deployer” of a regulated AI system within the
meaning of SB24-205, see § 6-1-1701(6), so only the provisions governing “developers” are at
issue in this action.
83.
To begin, the bill defines “[a]lgorithmic discrimination” as “any condition in which
the use of an [AI] system results in unlawful differential treatment or impact that disfavors an
individual or group of individuals on the basis of” certain protected characteristics. § 6-11701(1)(a). But not all “algorithmic discrimination” is prohibited. The bill proscribes only
discrimination that Colorado disagrees with. Indeed, it expressly exempts from its definition of
“algorithmic discrimination” any discrimination that “[e]xpand[s] an applicant, customer, or
participant pool to increase diversity or redress historical discrimination,” § 6-1-1701(1)(b)(I)(B),
thereby importing Colorado’s normative judgment of what preferences should be given to certain
groups.
84.
SB24-205’s substantive provisions build on the definition of “algorithmic
discrimination” in § 6-1-1701(1). First, the bill imposes an affirmative duty on a “developer of a
33
PDF Page 35
Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 34 of 75
high-risk [AI] system” to “use reasonable care to protect consumers from any known or reasonably
foreseeable risks of algorithmic discrimination arising from the intended and contracted uses of
the high-risk [AI] system,” § 6-1-1702(1)—regardless of whether discrimination is intended or in
fact occurs.
85.
Second, the bill requires that developers make certain disclosures about
“algorithmic discrimination” to deployers and other developers of the AI system, to the public,
and to the Attorney General. To deployers and other developers, a developer must “make
available” documentation disclosing or describing, among other things:
●
“known or reasonably foreseeable risks of algorithmic discrimination arising from
the intended uses of the high-risk [AI] system,” § 6-1-1702(2)(b)(II);
●
“[h]ow the high-risk [AI] system was evaluated for performance and mitigation of
algorithmic discrimination,” § 6-1-1702(2)(c)(I);
●
the “data governance measures used to cover the training datasets and the measures
used to examine the suitability of data sources, possible biases, and appropriate
mitigation,” § 6-1-1702(2)(c)(II);
●
the “measures the developer has taken to mitigate known or reasonably foreseeable
risks of algorithmic discrimination that may arise from the reasonably foreseeable
deployment of the high-risk [AI] system,” § 6-1-1702(2)(c)(IV); and
●
“[a]ny additional documentation that is reasonably necessary to assist the deployer
in understanding the outputs and monitor the performance of the high-risk [AI]
system for risks of algorithmic discrimination, § 6-1-1702(2)(d).
86.
For the public, “a developer shall make available, in a manner that is clear and
readily available on [its] website or in a public use case inventory, a statement summarizing,”
among other things, “[h]ow the developer manages known or reasonably foreseeable risks of
algorithmic discrimination that may arise from the development or intentional and substantial
modification of the types of high-risk [AI] systems described in accordance with subsection
(4)(a)(I) of this section.” § 6-1-1702(4)(a)(II). The developer “shall update” this public statement
34
PDF Page 36
Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 35 of 75
“[a]s necessary to ensure that the statement remains accurate” and “[n]o later than ninety days after
the developer intentionally and substantially modifies any high-risk [AI] system described in
subsection (4)(a)(I) of this section.” § 6-1-1702(4)(b)(I)-(II).
87.
And to the Attorney General and “all known deployers or other developers of the
high-risk [AI] system,” a developer “shall disclose … in a form and manner prescribed by the
[A]ttorney [G]eneral” “any known or reasonably foreseeable risks of algorithmic discrimination
arising from the intended uses of the high-risk [AI] system,” and shall make this disclosure “no
later than ninety days after” the developer discovers that its AI system “has caused or is reasonably
likely to have caused algorithmic discrimination” or receives a “credible report” that the system
has done so. § 6-1-1702(5). Further, at any time, the Attorney General can require a developer to
disclose the statement or documentation that § 6-1-1702(2) requires a developer to disclose to
deployers and other developers of the AI system. § 6-1-1702(7). A developer must comply “no
later than ninety days after the request and in a form and manner prescribed by the [A]ttorney
[G]eneral.” Id.
88.
The AI systems to which SB24-205’s requirements apply are near boundless.
Excepting yet another requirement that developers disclose “to each consumer who interacts with
the [AI] system that the consumer is interacting with an [AI] system,” § 6-1-1704(1), which
requirement appears to apply to all AI systems, SB24-205’s requirements facially apply only to
“high-risk [AI] systems.” § 6-1-1701(9)(a). Yet the phrase “high-risk” imposes no meaningful
limitation. An AI system is “high-risk” if it “makes, or is a substantial factor in making, a
consequential decision.” Id. And the definition of “substantial factor” renders the word
“substantial” meaningless. SB24-205 broadly defines that term as including (among other things)
35
PDF Page 37
Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 36 of 75
“any use of an [AI] system to generate any content, decision, prediction, or recommendation
concerning a consumer that is used as a basis to make a consequential decision.” § 6-1-1701(11)(b)
(emphasis added). Finally, and unsurprisingly, SB24-205 also broadly defines a “consequential
decision” as “a decision that has material legal or similarly significant effect on the provision or
denial to any consumer of, or the cost or terms of” any one of numerous opportunities or services,
including: “(a) [e]ducation enrollment or an education opportunity; (b) [e]mployment or an
employment opportunity; (c) [a] financial or lending service; (d) [a]n essential government service;
(e) [h]ealth-care services; (f) [h]ousing; (g) [i]nsurance; or (h) [a] legal service.” § 6-1-1701(3).
89.
Not only does SB24-205 apply to virtually any AI use in the covered subject
matters, it also has broad extraterritorial effect. The law applies to any “developer,” defined as “a
person doing business in this state that develops … an [AI] system.” § 6-1-1701(7). And for an AI
system that makes a “consequential decision” to become subject to SB24-205’s provisions, the
decision need only affect “any consumer”—that is, any “individual who is a Colorado resident.”
§§ 6-1-1701(3), (4), (9)(a). SB24-205 thus applies if even a single “Colorado resident” is impacted
by the system, regardless of where that Colorado resident is located, where the AI system is
developed or used, how unforeseeable that impact may be, or how many non-Colorado residents
are also impacted.
90.
Finally, SB24-205 grants the Attorney General broad and consequential
enforcement powers. First, it grants the Attorney General “exclusive authority to enforce” SB24205, § 6-1-1706(1), and a violation of the law “constitutes an unfair trade practice,” § 6-1-1706(2),
which is subject to a penalty of $20,000 per violation, § 6-1-112(1)(a). Second, it provides that the
36
PDF Page 38
Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 37 of 75
“[A]ttorney [G]eneral may promulgate rules as necessary for the purpose of implementing and
enforcing” the law. § 6-1-1707(1).
COUNT I
Declaratory Relief and Preliminary and Permanent Injunctive Relief for Violations of the
First Amendment to the United States Constitution (42 U.S.C. § 1983, 28 U.S.C. § 2201(a))
(Impermissible Burdens on Speech and Content- and Viewpoint-Based Discrimination)
91.
xAI re-alleges and incorporates by reference all allegations set forth above.
92.
SB24-205 interferes with xAI’s right to make speech-based editorial decisions
in developing Grok. xAI’s mission is to create an AI model that pursues the truth, without regard
to popularity, ideology, or social approval. This core philosophy, reflected in the publicly available
system prompts, pervades every aspect of Grok’s development. These are expressive decisions
protected by the First Amendment. SB24-205’s algorithmic-discrimination provision compels xAI
to alter these decisions and redesign Grok to promote Colorado’s views about fairness, equity, and
“acceptable” forms of discrimination based on protected characteristics such as race.
93.
The First Amendment provides that “Congress shall make no Law . . . abridging
the Freedom of Speech, or of the Press; or of the Right of the People to peaceably to assemble.”
U.S. Const. amend. I. These protections apply to the states through the Fourteenth Amendment’s
Due Process Clause.
94.
Under the First Amendment, “a speaker has the autonomy to choose the content of
his own message.” Hurley, 515 U.S. at 573. Indeed, since “all speech inherently involves choices
of what to say and what to leave unsaid,” it is axiomatic that “one who chooses to speak may also
decide ‘what not to say.’” Id. The First Amendment therefore “prohibits the government from
telling people what they must say,” Agency for Int’l Dev. v. Alliance for Open Soc’y Int’l, Inc.,
570 U.S. 205, 213 (2013), as it protects “both the right to speak freely and the right to refrain from
37
PDF Page 39
Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 38 of 75
speaking at all,” Wooley v. Maynard, 430 U.S. 705, 714 (1977). Laws that compel speakers to
“alter[] the content of their speech” are necessarily “content based” and therefore presumptively
unconstitutional. Nat’l Inst. of Family Life Advocs. v. Becerra (“NIFLA”), 585 U.S. 755, 766
(2018) (cleaned up).
95.
These principles remain unchanged when the means for creating and disseminating
speech is an AI model. See 303 Creative LLC, 600 U.S. at 587; Reno v. Am. Civil Liberties Union,
521 U.S. 844, 870 (1997). “[W]hatever the challenges of applying the Constitution to everadvancing technology,” the First Amendment’s basic principles “do not vary.” Brown v. Ent.
Mechs. Ass’n, 564 U.S. 786, 790 (2011).
96.
Every choice that xAI makes when developing Grok is an expressive act protected
by the First Amendment. These choices embody deliberate judgment that reflects xAI’s hierarchy
of values and its viewpoint-driven philosophy about how best to develop AI tools that will advance
the knowledge, understanding, and overall progress of human civilization.
97.
Just as video game creators engage in speech when designing games, id., social
media platforms engage in speech when crafting algorithms to curate content, Moody, 603 U.S. at
733, and search engine companies engage in speech when programming their algorithms, Langdon
v. Google, Inc., 474 F. Supp. 2d 622, 629-30 (D. Del. 2007), AI companies engage in speech when
designing, training, and fine-tuning a model.
98.
The starkly divergent philosophies of today’s frontier laboratories confirm the
expressive nature of the choices xAI makes when developing Grok. xAI designs, trains, and tunes
Grok to be a maximally truth-seeking model that prioritizes candor and rigorous empirical
accuracy. Other developers, whether because they prioritize progressive ideals on social, cultural,
38
PDF Page 40
Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 39 of 75
and political issues or because they endorse government messaging, take a different approach.
These design choices manifest themselves in the way the models respond to queries.
99.
These are not minor technical differences; they represent competing visions for the
future of AI. The choices that xAI makes to implement its own vision that AI should be maximally
truth seeking are therefore core expressive activities.
100.
SB24-205’s algorithmic-discrimination provision—which imposes a duty on
developers to prevent their models from generating content that could result in disparate impact
when used in connection with making decisions related to employment, education, and more—
burdens xAI’s right to free expression under the First Amendment. From a technical perspective,
complying with SB24-205’s mandate would require redesigning, retraining, or constraining the
Grok model by, for example, recalibrating how the model decides what information to include in
responses, hard-coding additional response guardrails, or re-weighting training datasets.
101.
Every technique for discharging the reasonable care duty that SB24-205 imposes
requires altering the training data, the model’s internal decision-making architecture, or how the
model is fine-tuned. See generally Isabel O. Gallegos et al., Bias and Fairness in Large Language
Models: A Survey, arXiv preprint 34-56 (July 12, 2024) (surveying bias mitigation techniques
proposed in academic literature), https://arxiv.org/pdf/2309.00770; Yufei Guo et al., Bias in Large
Language Models: Origins, Evaluation, and Mitigation, arXiv preprint 15-18 (Nov. 16, 2024)
(same), https://arxiv.org/pdf/2411.10915. Some researchers propose augmenting training datasets
with examples that replace protected attribute words, such as gendered pronouns, to achieve
greater balance. See Kaiji Lu et al., Gender Bias in Neural Natural Language Processing, arXiv
preprint 5-6 (May 30, 2019), https://arxiv.org/pdf/1807.11714. This approach requires, for
39
PDF Page 41
Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 40 of 75
instance, augmenting a dataset containing the sequence “he worked as an engineer” with the
sequence “she worked as an engineer.” Others propose modifying the system prompts to neutralize
perceived bias. See Alex Tamkin et al., Evaluating and Mitigating Discrimination in Language
Model Decisions, arXiv preprint 8-9 (Dec. 6, 2023), https://arxiv.org/pdf/2312.03689. And some
researchers suggest modifying the model’s core architecture or using specialized reinforcement
learning techniques. See Vishnu Asutosh Dasu et al., Attention Pruning: Automated Fairness
Repair of Language Models via Surrogate Simulated Annealing, arXiv preprint 4-6 (Nov. 25,
2025) (examining approach to suppress the architectural components that disproportionately drive
biased outputs), https://arxiv.org/pdf/2503.15815; Yuntao Bai, Constitutional AI: Harmlessness
from AI Feedback, arXiv preprint 10-11, 20-23 (Dec. 15, 2022) (detailing how developers can use
reinforcement
learning
to
teach
models
to
not
produce
biased
results),
https://arxiv.org/pdf/2212.08073.
102.
It is also implicit in SB24-205’s text that developers will have to make substantive
adjustments to their model training, data curation, fine-tuning, and alignment processes. The
statute requires developers to make available documentation describing, among other things, the
“data governance measures used to cover the training datasets and the measures used to examine
the suitability of data sources, possible biases, and appropriate mitigation,” § 6-1-1702(2)(c)(II),
and “the measures … taken to mitigate known or reasonably foreseeable risks of algorithmic
discrimination,” § 6-1-1702(2)(c)(IV). These contemplated mitigation measures would undeniably
require xAI to alter Grok’s maximally truth-seeking design.
103.
Colorado cannot alter xAI’s message simply because it wants to amplify its own
views on the highly politicized subjects of fairness and equity—views that xAI does not share.
40
PDF Page 42
Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 41 of 75
This violates the fundamental rule “that a speaker has the autonomy to choose the content of [its]
own message” and triggers strict scrutiny. Hurley, 515 U.S. at 573.
104.
SB24-205 also burdens users’ right to access speech. In addition to restricting
xAI’s expressive activity, SB24-205’s algorithmic-discrimination provision burdens users’ right
to access the information and ideas they can generate using Grok.
105.
The First Amendment not only “foster[s] individual self-expression” but also
“afford[s] the public access to discussion, debate, and the dissemination of information and ideas.”
First Nat’l Bank of Bos. v. Bellotti, 435 U.S. 765, 783 (1978). It is thus “well established” that the
First Amendment protects not only the right to speak but also “the right to receive information and
ideas.” Stanley v. Georgia, 394 U.S. 557, 564 (1969) (emphasis added); see also Va. State Bd. of
Pharmacy v. Va. Citizens Consumer Council, Inc., 425 U.S. 748, 757 (1976); Martin v. City of
Struthers, 319 U.S. 141, 143 (1943).
106.
This right is an indispensable component of the First Amendment because the
“dissemination of ideas can accomplish nothing if otherwise willing addressees are not free to
receive and consider them” for it “would be a barren marketplace of ideas that had only sellers and
no buyers.” Lamont v. Postmaster Gen., 381 U.S. 301, 308 (1965) (Brennan, J., concurring); see
also Bd. of Educ., Island Trees Union Free Sch. Dist. No. 26 v. Pico, 457 U.S. 853, 867 (1982)
(explaining that “the right to receive ideas is a necessary predicate to the recipient’s meaningful
exercise of his own rights of speech, press, and political freedom”).
107.
The responses, predictions, explanations, and recommendations produced using
Grok qualify as speech under the First Amendment, and users are therefore entitled to receive those
outputs.
41
PDF Page 43
Case No. 1:26-cv-01515
108.
Document 1
filed 04/09/26
USDC Colorado
pg 42 of 75
The quintessential “mediums of expression” are “written or spoken words.” Hurley,
515 U.S. at 569. Words spoken during “conversation[s] … transmi[t] … ideas” and thus qualify
as speech. McCullen v. Coakley, 573 U.S. 464, 488 (2014); see also United States v. Alvarez, 567
U.S. 709, 722-23 (2012) (plurality op.) (acknowledging that “personal … conversations” are
speech); Legal Servs. Corp. v. Velazquez, 531 U.S. 533, 542-43 (2001) (so too is “advice from [an]
attorney to [a] client”).
109.
In form and function, a user’s interactions with Grok are indistinguishable from the
back-and-forth dialogue that occurs in ordinary human discourse. A user poses a prompt and Grok
replies with an answer, recommendation, or explanation—just as a knowledgeable human
interlocutor would do in real-time conversation.
110.
And even if considered in isolation, Grok’s outputs qualify as protected speech
because they communicate ideas no less than books, scholarly articles, professional reports, and
other media within the First Amendment’s heartland. See, e.g., 303 Creative, 600 U.S. at 587 (“All
manner of speech—from pictures, films, paintings, drawings, and engravings, to oral utterance and
the printed word—qualify for the First Amendment’s protections; no less can hold true when it
comes to speech … conveyed over the Internet.”) (cleaned up); Brown, 564 U.S. at 790 (“Like the
protected books, plays, and movies that preceded them, video games communicate ideas—even
social messages—through many familiar literary devices … and through features distinctive to the
medium …. That suffices to confer First Amendment protection.”).
111.
It makes no difference that Grok’s outputs are generated by an AI system; speech
need not be attributable to a speaker with First Amendment rights to qualify for protection. For
example, even though the First Amendment does not protect foreign speakers who produce
42
PDF Page 44
Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 43 of 75
political propaganda abroad, the Supreme Court held that a restriction on mailing such material
into the United States violated the “addressee’s First Amendment rights.” Lamont, 381 U.S. at 307.
And before the constitutional protection for corporations’ right to speak about elections was settled
in Citizens United, the Court rejected restrictions on corporate political speech based on the
interests of listeners. Bellotti, 435 U.S. at 777. The Court explained that the “inherent worth of the
speech in terms of its capacity for informing the public does not depend upon the identity of its
source, whether corporation, association, union, or individual.” Id.
112.
SB24-205’s algorithmic-discrimination provision interferes with users’ right to
receive speech generated through Grok. The provision requires developers of “high-risk [AI]
systems” to “use reasonable care to protect consumers from any known or reasonably foreseeable
risks of algorithmic discrimination.” § 6-1-1702(1). And “algorithmic discrimination” is defined
expansively as “any condition in which the use of an artificial intelligence system results in an
unlawful differential … impact that disfavors an individual or group” on the basis of certain
characteristics. § 6-1-1701(1)(a). Complying with this mandate will force xAI to alter Grok’s
outputs.
113.
The inevitable—indeed, intended—consequence is that users will receive only
government-approved, sanitized versions of Grok’s outputs. The statistically accurate
explanations, unfiltered recommendations, and candid analyses that risk producing disfavored
outcomes for protected classes will have to be altered, diluted, or withheld entirely. Users will thus
be systematically denied access to the full, unaltered outputs that Grok would otherwise provide.
This is a direct—and substantial—burden on users’ First Amendment right to receive information
and ideas through Grok, subjecting SB24-205 to strict scrutiny.
43
PDF Page 45
Case No. 1:26-cv-01515
114.
Document 1
filed 04/09/26
USDC Colorado
pg 44 of 75
SB24-205 regulates speech based on content and viewpoint. SB24-205 also
impermissibly draws distinctions based on content and viewpoint.
115.
It is the “most basic” principle of the First Amendment that the “government has
no power to restrict expression because of its message, its ideas, its subject matter, or its content.”
Brown, 564 U.S. at 790-91. “Content-based laws—those that target speech based on its
communicative content—are presumptively unconstitutional and may be justified only if the
government proves they are narrowly tailored to serve compelling state interests.” Reed v. Town
of Gilbert, 576 U.S. 155, 163 (2015) (citations omitted).
116.
By its plain terms, SB24-205 restricts only AI systems that generate content
affecting certain decisions. The law says nothing about AI systems generating outputs on protein
folding, computer coding, and the like. But if the same systems are used to generate content
addressing applications for employment, educational, credit, or housing opportunities, then SB24205 imposes an additional burden on the developer to exercise reasonable care to protect
consumers from reasonably foreseeable risks of algorithmic discrimination. §§ 6-1-1701(3), 6-11702(1). The law thus “singles out specific subject matter for differential treatment.” Reed, 576
U.S. at 169. “That is about as content-based as it gets.” Barr v. Am. Ass’n of Pol. Consultants, Inc.,
591 U.S. 610, 619 (2020) (plurality op.).
117.
SB24-205 draws yet another content-based distinction when describing the
particular outputs that it regulates within those subject matters. It imposes a duty of reasonable
care to protect consumers from systems generating outputs that contain “algorithmic
discrimination,” a term the law defines as outputs that produce “unlawful differential treatment or
44
PDF Page 46
Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 45 of 75
impact that disfavors an individual or group of individuals on the basis” of certain characteristics
(e.g., age, race, and disability). § 6-1-1701(1)(a).
118.
xAI’s statutory compliance thus turns on whether Grok is generating outputs that
contain particular types of speech on particular subjects. This will necessarily require the Attorney
General “to examine the content of the message that is conveyed to determine whether a violation
has occurred,” a hallmark of content-based laws. Animal Legal Defense Fund v. Kelly, 9 F.4th
1219, 1228 (10th Cir. 2021) (quoting McCullen, 573 U.S. at 479).
119.
Indeed, on this front SB24-205 “goes even beyond mere content discrimination, to
actual viewpoint discrimination.” R.A.V. v. City of St. Paul, 505 U.S. 377, 391 (1992). This is
because the statute expressly carves out from its definition of “algorithmic discrimination” AI
systems with outputs that “expan[d] an applicant, customer, or participant pool to increase
diversity or redress historical discrimination.” § 6-1-1701(1)(b)(I)(B).
120.
In other words, SB24-205 does not neutrally target AI systems that produce any
disparate impact based on the identified characteristic. It targets only AI systems that might
produce outputs leading to what Colorado views as the wrong kind of disparate impact. That is
textbook viewpoint discrimination.
121.
SB24-205 cannot survive any level of heightened scrutiny, let alone strict
scrutiny. Because strict scrutiny applies, Colorado bears the heavy burden of demonstrating that
SB24-205 is “the least restrictive means of achieving a compelling state interest.” McCullen, 573
U.S. at 478. Even under intermediate scrutiny, Colorado would have to prove that SB24-205 is
“narrowly tailored to serve a significant governmental interest.” Packingham v. N. Carolina, 582
U.S. 98, 105-06 (2017). It cannot do so.
45
PDF Page 47
Case No. 1:26-cv-01515
122.
Document 1
filed 04/09/26
USDC Colorado
pg 46 of 75
SB24-205 is not narrowly tailored to advance any legitimate interest that Colorado
could assert in, for example, curbing invidious discrimination. The law is both overinclusive and
underinclusive. And there are far less restrictive alternatives that function to achieve the same
ends.
123.
SB24-205 is overinclusive in at least two ways. First, it reaches far beyond just
invidious discrimination by embracing a sweeping disparate-impact theory. The statute defines
“algorithmic discrimination” to encompass any use of an AI system that “results in an unlawful
differential treatment or impact” disfavoring an individual or group based on certain protected
characteristics. § 6-1-1701(1)(a). This disjunctive phrasing makes clear that SB24-205 targets not
only intentional discrimination, but also raw statistical imbalances unconnected to any
discriminatory animus. Governor Polis himself expressed reservations about SB24-205 for exactly
this reason, explaining that “[l]aws that seek to prevent discrimination generally focus on
prohibiting intentional discriminatory conduct,” but SB24-205 “deviates from that practice by
regulating the results of AI system use, regardless of intent.”
124.
Second, SB24-205 is overinclusive because it reaches outputs with little bearing on
any consequential decision, and therefore almost no relationship to the government’s interest in
preventing discrimination. Although the statute facially covers only AI that are a “substantial
factor” in making consequential decisions, it broadly defines that term to include “any use … to
generate any content … that is used as a basis to make a consequential decision” concerning a
consumer. § 6-1-1701(11)(b) (emphasis added). This captures using Grok to generate interview
questions, summarize writing samples, and perform other tasks for which the AI system plays no
direct role in the ultimate decision.
46
PDF Page 48
Case No. 1:26-cv-01515
125.
Document 1
filed 04/09/26
USDC Colorado
pg 47 of 75
Even performing purely administrative tasks can qualify as a “substantial factor”
under SB24-205. For example, judges might use an AI system to create a clerkship hiring
spreadsheet that lists every applicant’s name, law school, graduation year, and grade point average.
That is “content” that would be “used as a basis” to make hiring decisions. But there is no risk that
using the AI system introduced any bias into the decision making process. Thus, SB24-205 is “not
reasonably restricted to the evil with which it is said to deal.” Butler v. Michigan, 352 U.S. 380,
383 (1957).
126.
SB24-205 is also underinclusive when judged against any interest Colorado may
have in rooting out intentional discriminatory conduct. The statute defines “algorithmic
discrimination” broadly to encompass any use of an AI system that “results in an unlawful
differential treatment or impact that disfavors an individual” on the basis of certain characteristics.
§ 6-1-1701(1)(a). But it expressly exempts the use of AI systems to “[e]xpand[] an applicant,
customer, or participant pool to increase diversity or redress historical discrimination.” § 6-11701(1)(b)(I)(B). In other words, the law blesses the very thing it supposedly condemns. That type
of incoherence is the hallmark of underinclusivity. See Brown, 564 U.S. at 805.
127.
Finally, the availability of less restrictive alternatives further dooms SB24-205.
State and federal law—for example, the Civil Rights Act of 1964, the Fair Housing Act, the Equal
Credit Opportunity Act, and the Colorado Anti-Discrimination Act—already prohibit intentional
discrimination in employment, housing, education, finance, and other decisions within SB24-205’s
scope. These statutes target intentional discriminatory conduct with precision and without
unnecessarily restricting protected expression. “[R]egulating speech must be a last—not first—
resort.” Thompson v. Western States Med. Ctr., 535 U.S. 357, 373 (2002). And Colorado cannot
47
PDF Page 49
Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 48 of 75
show why these “available, effective alternatives” would fail where only speech coercion can
succeed. Ashcroft v. ACLU, 542 U.S. 656, 666 (2004).
128.
Because SB24-205 reaches far more speech than is necessary to advance any
professed interest in preventing invidious discrimination, is vastly underinclusive when judged
against that interest, and is not the least restrictive means, it cannot survive any level of First
Amendment scrutiny. SB24-205 should accordingly be declared unconstitutional, and the Attorney
General should be enjoined from enforcing it against xAI.
COUNT II
Declaratory Relief and Preliminary and Permanent Injunctive Relief for Violations of the
First Amendment to the United States Constitution (42 U.S.C. § 1983, 28 U.S.C. § 2201(a))
(Mandatory Disclosure Provisions)
129.
xAI re-alleges and incorporates by reference all allegations set forth above.
130.
SB24-205’s disclosure provisions likewise violate the First Amendment. By
forcing xAI to create, maintain, and disseminate information about its efforts to identify and
mitigate algorithmic bias, the bill compels xAI to speak in violation of its right to free speech.
131.
The Supreme Court has long recognized that the First Amendment’s guarantee of
free speech “includes both the right to speak freely and the right to refrain from speaking at all.”
Wooley, 430 U.S. at 714. That protection extends “not only to expressions of value, opinion, or
endorsement, but equally to statements of fact the speaker would rather avoid.” Hurley, 515 U.S.
at 573.
132.
Laws compelling speech are generally treated no differently than laws restricting
speech, see, e.g., NIFLA, 585 U.S. at 766-67, even when the government does not compel a speaker
to express any particular message, Riley, 487 U.S. at 795. As the Supreme Court has explained,
48
PDF Page 50
Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 49 of 75
the standard First Amendment analysis applies equally when the government compels speakers to
convey purely factual information. See Hurley, 515 U.S. at 573.
133.
SB24-205 requires developers to make disclosures regarding their evaluation and
mitigation of “algorithmic discrimination” to deployers, the public, and the Colorado Attorney
General. The disclosures to developers include, for example, documentation describing: (a) how
xAI evaluated the model for “mitigation of algorithmic discrimination,” § 6-1-1702(2)(c)(I);
(b) the “data governance measures used to cover the training datasets and the measures used to
examine the suitability of data sources, possible biases, and appropriate mitigation,” § 6-11702(2)(c)(II); and (c) the measures xAI “has taken to mitigate known or reasonably foreseeable
risks of algorithmic discrimination,” § 6-1-1702(2)(c)(IV). xAI must also publish on its website a
statement summarizing how it “manages known or reasonably foreseeable risks of algorithmic
discrimination.” § 6-1-1702(4)(a)(II). And xAI has to inform the Attorney General within ninety
days if it discovers that its AI system caused algorithmic discrimination or was “reasonably likely”
to have caused algorithmic discrimination, or if it receives a “credible report that the high-risk [AI]
system has been deployed and has caused algorithmic discrimination.” § 6-1-1702(5)(b).
134.
These disclosure provisions are content-based regulations that trigger strict scrutiny
because they compel xAI to disclose specific content related to its AI models. Strict scrutiny
applies whenever the government forces companies to disclose particular content, whether it be
non-commercial speech concerning factual information about a company’s service or products,
see X Corp. v. Bonta, 116 F.4th 888, 902 (9th Cir. 2024), risks those services or products may
pose, see NetChoice v. Weiser, 808 F. Supp. 3d 1223, 1240-41 (D. Colo. 2025), or information
about where to obtain other types of products or services, see NIFLA, 585 U.S. at 766. Like those
49
PDF Page 51
Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 50 of 75
content-based disclosures, the disclosures SB24-205 requires force xAI to speak a particular
message that it otherwise would not about whether and how it evaluates and mitigates “algorithmic
discrimination.”
135.
The disclosure provisions are also content- and viewpoint-based because they rely
on SB24-205’s definition of “algorithmic discrimination,” which itself draws distinctions based
on content and viewpoint—as explained above, ¶¶ 114-20. See NetChoice, LLC v. Reyes, 748 F.
Supp. 3d 1105, 1120 (D. Utah 2024) (finding “persuasive” the argument that “the entire Act
facially violates the First Amendment because the Act’s operative provisions each rely” on
definitions that impose content-based distinctions). SB24-205’s disclosure provisions thus trigger
strict scrutiny multiple times over.
136.
SB24-205’s disclosure provisions cannot withstand strict scrutiny or any other form
of heightened scrutiny. Colorado lacks a sufficient governmental interest in compelling AI
developers to disclose information about their practices for evaluating and mitigating “algorithmic
discrimination.” To satisfy First Amendment scrutiny, Colorado must “specifically identify an
actual problem in need of solving.” Brown, 564 U.S. at 799 (cleaned up); see also Reyes, 748 F.
Supp. 3d at 1124 (same). And the problem must be in need of a governmental solution, as opposed
to a private one.
137.
Many frontier AI laboratories already publish reports detailing methods for
evaluating and mitigating different types of bias. For instance, Anthropic engineers published a
report examining whether Claude 2.0 would generate biased results in simulated high-stakes
decision making scenarios (e.g., increasing a person’s credit). Tamkin et al., supra, at 5-6 (finding
what the authors term “evidence of positive discrimination (i.e., in favor of genders other than
50
PDF Page 52
Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 51 of 75
male and races other than white) in the Claude 2 model” for certain decision scenarios). The report
also considered prompt-engineering techniques for mitigating bias in the model’s decision making.
Id. And xAI assesses its models for sycophantic tendencies—i.e., whether they answer according
to the user’s suggestion rather than based on objective and unbiased subject information. See, e.g.,
xAI, Grok 4.1 Model Card (last updated November 17, 2025), https://data.x.ai/2025-11-17-grok4-1-model-card.pdf; xAI, Grok 4 Fast Model Card (last updated Sept. 19, 2025),
https://perma.cc/FEK8-CYQV.
138.
Academics, nonprofits, and other organizations likewise publish reports evaluating
whether AI systems exhibit bias in decision making. See, e.g., David Rozado, Fairness in AI
Decisions About People: Evidence from LLM Experiments, Manhattan Inst. (Jan. 8, 2026)
(describing
results
from
experiment
testing
prominent
LLMs
for
gender
bias),
https://manhattan.institute/article/fairness-in-ai-decisions-about-people-evidence-from-llmexperiments; Eitan Anzanberg et al., Evaluating the Promise and Pitfalls of LLMs in Hiring
Decisions, arXiv preprint (Jul. 28, 2025) (benchmarking several frontier, general-purpose LLMs
for candidate job-matching in resume screening), https://arxiv.org/pdf/2507.02087; Jiafu An et al.,
Measuring Gender and Racial Biases in Large Language Models: Intersectional Evidence from
Automated Resume Evaluation, arXiv preprint (Mar. 12, 2025) (evaluating whether certain LLMs
exhibit racial and gender bias when used to assess entry-level job candidates),
https://arxiv.org/pdf/2403.15281. 69
69
https://futurefreespeech.org/new-report-ai-laws-and-chatbots-face-a-global-freespeech-test/.
51
PDF Page 53
Case No. 1:26-cv-01515
139.
Document 1
filed 04/09/26
USDC Colorado
pg 52 of 75
Colorado does not have a sufficient interest in forcing developers to produce and
disseminate reams of information about their practices for evaluating and mitigating bias when
functionally identical information is already available from private sources. Whatever “modest
gap” may exist between what SB24-205 requires developers to disclose and what is already
available, filling it “can hardly be a compelling state interest.” Brown, 564 U.S. at 803.
140.
In addition, the disclosure provisions are not narrowly tailored for the same reasons
as the algorithmic discrimination provision—all of which rely on the same overinclusive and
underinclusive framework. See supra ¶¶ 121-28.
141.
Forcing developers to create and disseminate information is also not the least
restrictive means for accomplishing the State’s goals. One obvious alternative is for Colorado itself
to evaluate whether certain models are producing outputs that may result in “algorithmic
discrimination” and publish its findings. Indeed, it is not even clear why disclosures about
developers’ techniques for measuring and mitigating bias in training datasets and model
performance would provide meaningfully more information than simply testing the model—
something that the State (or any developer, for that matter) can do on its own. Indeed, the federal
government has adopted a similar government-led testing approach to ensure that models available
in the United States protect free speech and American values. See America’s AI Action Plan at 4
(directing Department of Commerce to “conduct research and, as appropriate, publish evaluations
of frontier models from the People’s Republic of China for alignment with Chinese Community
Party
talking
points
and
censorship”),
content/uploads/2025/07/Americas-AI-Action-Plan.pdf.
52
https://www.whitehouse.gov/wp-
PDF Page 54
Case No. 1:26-cv-01515
142.
Document 1
filed 04/09/26
USDC Colorado
pg 53 of 75
This case does not involve the kind of compelled-speech mandate that might get
lesser scrutiny under Zauderer v. Off. of Disciplinary Couns. of Supreme Ct. of Ohio, 471 U.S. 626
(1985). Zauderer creates a limited exception for government-compelled speech that aims to
combat misleading advertisements. It applies only to First Amendment claims involving
commercial speech. See NetChoice, 808 F. Supp. 3d at 1239-40. And the Supreme Court has never
applied the principles set forth in Zauderer outside the context of misleading advertising. See
NIFLA, 585 U.S. at 768. To the contrary, the Court has consistently reaffirmed that Zauderer
focused only on “combat[ing] the problem of inherently misleading commercial advertisements.”
Milavetz, Gallop & Milavetz, P.A. v. United States, 559 U.S. 229, 250 (2010); see also, e.g.,
Hurley, 515 U.S. at 573, United States v. United Foods, 533 U.S. 405, 416 (2001).
143.
Zauderer does not apply because SB24-205 does not compel commercial speech.
Commercial speech is speech “that does no more than propose a commercial transaction.” United
Foods, 533 U.S. at 409. That criterion is not simply a clue or guidepost; it is “the test for identifying
commercial speech.” City of Cincinnati v. Discovery Network, Inc., 507 U.S. 410, 423 (1993) (Bd.
of Trs. of State Univ. of N.Y. v. Fox, 492 U.S. 469, 473-74 (1989)). None of the information that
SB24-205 forces xAI to disclose (e.g., its practices for evaluating and mitigating model bias) has
anything to do with proposing a commercial transaction, let alone addressing any potentially
misleading advertising.
144.
The Tenth Circuit’s multi-factor test for commercial speech, which only applies in
close cases, likewise supports the conclusion that SB24-205’s disclosure provisions do not compel
commercial speech. That test turns on whether the speech (1) is “concededly an advertisement”;
(2) “refers to a specific product”; or (3) “is motivated by an economic interest.” United States v.
53
PDF Page 55
Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 54 of 75
Wenger, 427 F.3d 840, 847 (10th Cir. 2005) (citing Bolger v. Youngs Drug Prods. Corp., 463 U.S.
60, 66-67 (1983)). As an initial matter, these factors presuppose that an entity is already engaging
in speech that the government then regulates—and not that the government is forcing companies
to create speech. The government cannot compel private entities to speak in the first instance, and
then label that speech “commercial.” Even so, the Bolger factors weigh against concluding that
the disclosure provisions implicate commercial speech. The disclosures SB24-205 mandates are
not advertisements; they are not designed to sell a product or service. They do not refer to a
particular product; they describe xAI’s practices for evaluating and mitigating algorithmic bias.
And the disclosures would only be produced out of legal obligation, not economic motivation.
145.
Moreover, Zauderer does not apply because SB24-205’s compelled disclosures are
neither “purely factual” nor “uncontroversial.” A disclosure is “purely factual” if it only requires
“the disclosure of accurate, factual information.” Nat’l Ass’n of Wheat Growers v. Bonta, 85 F.4th
1263, 1276 (9th Cir. 2023). The word “purely” is critical: it excludes speech that embeds normative
framing, subjective interpretation, or predictive judgment. SB24-205’s disclosures are not purely
factual because they require developers to adopt and apply the statute’s definition of “algorithmic
discrimination,” a normative framework shaped by Colorado’s own views about equality and
fairness. Evaluating the risk of algorithmic discrimination and choosing appropriate mitigation
measures is also an inherently subjective exercise. And, tellingly, the disclosures here bear little
resemblance to those courts have concluded are “purely factual.” See, e.g., Zauderer, 471 U.S. at
652 (clarification that client may be liable for costs for attorney advertisements that promised “no
recovery, no legal fees”); Am. Meat Inst. v. U.S. Dep’t of Agric., 760 F.3d 18, 27 (D.C. Cir. 2014)
(en banc) (country-of-origin labeling); New York State Rest. Ass’n v. New York City Bd. of Health,
54
PDF Page 56
Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 55 of 75
556 F.3d 114, 131-32 (2d Cir. 2009) (calorie count); Nat’l Elec. Mfrs. Ass’n v. Sorrell, 272 F.3d
104, 107 (2d Cir. 2001) (instructions for safe product use and disposal).
146.
SB24-205’s compelled disclosures are also not “uncontroversial.” A statement is
“uncontroversial” under Zauderer “where the truth of the statement is not subject to good-faith
scientific or evidentiary dispute and where the statement is not an integral part of a live, contentious
political or moral debate.” Free Speech Coal., Inc. v. Paxton, 95 F.4th 263, 281-82 (5th Cir. 2024).
SB24-205’s disclosures force developers to opine on the public and scholarly debate about whether
using AI tools to make hiring (and other) decisions produces discriminatory results. Indeed, the
disclosures implicate one of the most persistent and divisive debates in American history—what
qualifies as discrimination. And xAI does not share Colorado’s views—that algorithmic
discrimination can be defined to exclude using AI tools to “expan[d] an applicant, customer, or
participant pool to increase diversity or redress historical discrimination.” § 6-1-1701(1)(b)(I)(B).
147.
The sheer volume of scholarly literature on how to evaluate and mitigate
algorithmic discrimination shows that SB24-205’s disclosures are not “uncontroversial.” Several
prominent papers catalog dozens of potential fairness metrics and distinct bias types, highlighting
the lack of consensus on how to measure or remediate bias. See Gallegos et al., supra, at 34-56;
Guo et al., supra, at 15-18. Other papers underscore the difficult accuracy trade-offs that
developers must weigh when considering measures designed to mitigate any bias that exists in the
training datasets. See Tamkin et al., supra, at 9-11. In fact, the Association for Computing
Machinery hosts an annual interdisciplinary conference focused on the social and technical
challenges of algorithmic systems, including the potential for bias. In short, scholars disagree
profoundly about whether AI systems produce discriminatory results, how to measure algorithmic
55
PDF Page 57
Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 56 of 75
bias, what causes bias in outputs, and the best strategies for mitigating bias. SB24-205’s compelled
disclosures would force xAI to take sides in this public debate by disclosing documentation
describing its own methodology for evaluating and mitigating algorithmic discrimination.
148.
That these issues are also the subject of active litigation further confirms that SB24-
205’s compelled disclosures are not “uncontroversial” under Zauderer. See, e.g., Mobley v.
Workday Inc., 740 F. Supp. 3d 796 (N.D. Cal. 2024) (disparate impact claim under Title VII based
on allegation that use of Workday’s AI system produced discriminatory hiring decisions); Huskey
v. State Farm Fire & Cas. Co., No. 22 C 7014, 2023 WL 5848164 (N.D. Ill. Sept. 11, 2023) (Fair
Housing Act claim based on allegation that use of machine-learning algorithm used to screen for
potentially fraudulent claims subjected black policyholders to additional administrative hurdles
and processing delays); Louis v. SafeRent Sols., LLC, 685 F. Supp. 3d 19 (D. Mass. 2023) (Fair
Housing Act claim based on allegation that tenant-screening algorithm biased against certain
applicants).
149.
SB24-205’s disclosure provisions trigger strict scrutiny because they are content-
and viewpoint-based. And they cannot withstand strict scrutiny or any other form of heightened
scrutiny. Zauderer does not apply because SB24-205’s disclosure provisions do not compel
commercial speech and, at any rate, the disclosures are neither “factual” nor “uncontroversial.”
SB24-205’s disclosure provisions should accordingly be declared unconstitutional, and the
Attorney General should be enjoined from enforcing them against xAI.
COUNT III
Declaratory Relief and Preliminary and Permanent Injunctive Relief for Extraterritorial
Regulation in Violation of the Commerce Clause of the United States Constitution and the
Horizontal Separation of Powers (U.S. Const. art. I, § 8, cl. 3)
150.
xAI re-alleges and incorporates by reference all allegations set forth above.
56
PDF Page 58
Case No. 1:26-cv-01515
151.
Document 1
filed 04/09/26
USDC Colorado
pg 57 of 75
The Constitution restricts the power of states to directly regulate conduct that takes
place entirely in another state. That bedrock principle of equal sovereignty among the states is
apparent in several of the Constitution’s structural protections and deeply rooted in our nation’s
historical tradition. See Nat’l Pork Producers Council v. Ross, 598 U.S. 356, 376 n.1 (2023); id.
at 403-04, 408-10 (Kavanaugh J., concurring in part and dissenting in part).
152.
Although this principle does not erect a per se bar against laws that regulate conduct
within one state in ways that have an “extraterritorial effect” in others, Ross, 598 U.S. at 357-58
(emphasis added), the same is not true of laws that “directly regulated out-of-state transactions.”
Id. at 376 n.1. “[O]riginal and historical understandings of the Constitution’s structure and the
principles of ‘sovereignty and comity’ it embraces” inform “the territorial limits of state authority
under the Constitution’s horizontal separation of powers.” Id. at 376 & 376 n.1.
153.
It is axiomatic that “all States enjoy equal sovereignty.” Shelby Cnty. v. Holder,
570 U.S. 529, 535 (2013); see also PPL Mont., LLC v. Montana, 565 U.S. 576, 591 (2012) (“the
States in the Union are coequal sovereigns under the Constitution”). Indeed, “the constitutional
equality of the states is essential to the harmonious operation of the scheme upon which the
Republic was organized.” Coyle v. Smith, 221 U.S. 559, 580 (1911). When a State reaches beyond
its own borders to “directly regulate[] out-of-state transactions by those with no connection to the
State,” Ross, 598 U.S. at 376 n.1 (emphasis omitted), it invades the sovereignty and impinges on
the equality of other states. Accordingly, one State is necessarily prohibited from directly
regulating conduct that neither occurs nor is directed within its borders. Cf. PennEast Pipeline Co.
v. New Jersey, 141 S.Ct. 2244, 2259 (2021) (“The plan of the Convention reflects the ‘fundamental
postulates implicit in the constitutional design.’” (quoting Alden v. Maine, 527 U.S. 706, 729
57
PDF Page 59
Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 58 of 75
(1999)); State Farm Mut. Auto. Ins. Co. v. Campbell, 538 U.S. 408, 422 (2003) (“A basic principle
of federalism is that each State may make its own reasoned judgment about what conduct is
permitted or proscribed within its borders, and each State alone can determine what measure of
punishment, if any, to impose on a defendant who acts within its jurisdiction.”).
154.
Consistent with our founding’s “special concern both with the maintenance of a
national economic union unfettered by state-imposed limitations on interstate commerce and with
the autonomy of the individual States within their respective spheres,” the Supreme Court has held
that the Commerce Clause (U.S. Const. art. I, § 8, cl. 3) prohibits any state from “control[ling]
commerce occurring wholly outside [its] boundaries.” Healy v. Beer Inst., Inc., 491 U.S. 324, 33536 (1989) (footnote omitted); see also, e.g., Brown-Forman Distillers Corp. v. N.Y. State Liquor
Auth., 476 U.S. 573, 579 (1986); Edgar v. MITE Corp., 457 U.S. 624, 643 (1982).
155.
The Court has also long interpreted the Due Process Clause to impose restrictions
on a state’s ability to regulate conduct occurring wholly outside its borders. See, e.g., Watson v.
Emps. Liab. Assurance Corp., 348 U.S. 66, 70 (1954) (recognizing “the due process principle that
a state is without power to exercise ‘extra territorial jurisdiction,’ that is, to regulate and control
activities wholly beyond its boundaries”); Home Ins. Co. v. Dick, 281 U.S. 397, 407-08 (1930)
(recognizing and applying the same principle).
156.
And, of course, the Tenth Amendment provides that “powers not delegated to the
United States by the Constitution, nor prohibited by it to the States, are reserved to the States
respectively, or to the people,” U.S. Const. amend. X, making clear that each state retains its own
“integrity, dignity, and residual sovereignty,” Bond v. United States, 564 U.S. 211, 221 (2011).
58
PDF Page 60
Case No. 1:26-cv-01515
157.
Document 1
filed 04/09/26
USDC Colorado
pg 59 of 75
Despite these prohibitions, SB24-205 creates liability for xAI based purely on out-
of-state conduct. The law broadly applies to any “developer”—“a person doing business in
[Colorado] that develops … an artificial intelligence system.” § 6-1-1701(7). Each of SB24-205’s
provisions further require that they impact a “consumer”—i.e., any “individual who is a Colorado
resident.” § 6-1-1701(4). The provisions that require developers to mitigate and disclose risks
concerning “algorithmic discrimination” cover “high-risk [AI] system[s],” which are systems that
“make[], or is a substantial factor in making, a consequential decision” affecting “any consumer.”
§§ 6-1-1701(1), (3), (4), (9)(a); 6-1-1702. Likewise, SB24-205 requires developers to disclose “to
each consumer who interacts with the artificial intelligence system that the consumer is interacting
with an artificial intelligence system.” § 6-1-1704(1). There is no requirement that the Colorado
resident be in Colorado or that the use of the AI system or “consequential decision” occur within
Colorado.
158.
Thus, xAI is subject to SB24-205’s provisions merely if (1) it is a “developer” and
(2) its AI system impacts a consequential decision affecting a Colorado resident. It is unimportant
where that decision occurs, where the Colorado resident is located, how unforeseeable
“algorithmic discrimination” was to xAI, or even whether the discrimination was intended.
159.
It is not difficult to think of an example reflecting how extraordinary these
provisions are in scope. Consider a healthcare company that is not incorporated or headquartered
in Colorado and does not maintain offices here. That company engaged in negotiations for Grok,
which was developed in California, and contracted with xAI, which is Nevada-incorporated and
California-based. Under SB24-205’s provisions governing “algorithmic discrimination,” xAI
would be exposed to liability under the law if a single medical decision affecting a Colorado
59
PDF Page 61
Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 60 of 75
resident visiting one of the healthcare company’s out-of-state offices is made using Grok. § 6-11702.
160.
Not only does xAI already have nationwide enterprise customers using Grok to
make consequential decisions, see supra ¶ 69, xAI’s enterprise business is rapidly growing, and
xAI is in the process of obtaining many additional customers who would leverage Grok to assist
in making decisions in other categories covered by SB24-205, including financial and lending
services, insurance, and housing. xAI’s continued expansion will certainly expose it to additional
liability under SB24-205.
161.
But it is not only xAI that faces this reality. None of the major enterprise AI
developers are headquartered or incorporated in Colorado:
●
xAI is Nevada-incorporated and California-headquartered;
●
Anthropic is Delaware-incorporated and California-headquartered;
●
OpenAI is Delaware-incorporated and California-headquartered; and
●
Google (Gemini) is Delaware-incorporated and California-headquartered.
SB24-205’s extraterritorial scope therefore is compounded.
162.
SB24-205’s provision requiring xAI to disclose “to each consumer who interacts
with the artificial intelligence system that the consumer is interacting with an artificial intelligence
system” is equally unconstitutional. § 6-1-1704(1). Like the provisions governing “algorithmic
discrimination,” § 6-1-1702, it applies regardless of where the “consumer” (i.e., Colorado resident)
is, and so applies to interactions with Grok nationwide.
163.
SB24-205 thus regulates transactions that are consummated entirely outside of
Colorado, by entities entirely outside of Colorado, and for systems developed entirely outside of
60
PDF Page 62
Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 61 of 75
Colorado. This violates the Constitution: indeed, imposing state-law liability on out-of-state actors
for actions taken entirely out of state is the definition of unconstitutional extraterritorial state
regulation. See Ross, 598 U.S. at 376 n.1 (“law that directly regulated out-of-state transactions by
those with no connection to the State” are unconstitutional); Edgar, 457 U.S. at 624; Styczinski v.
Arnold, 46 F.4th 907, 913 (8th Cir. 2022) (impermissibly extraterritorial where the law covers “a
Minnesota transaction[, which] includes a transaction anywhere in the world between a bullion
trader and a Minnesota resident. A bullion trader could therefore become subject to and violate
Minnesota law without conducting a single transaction in Minnesota.”).
COUNT IV
Declaratory Relief and Preliminary and Permanent Injunctive Relief for the Unlawful
Imposition of an Undue Burden on Interstate Commerce (Pike Balancing)
164.
xAI re-alleges and incorporates by reference all allegations set forth above.
165.
In addition to prohibiting states from regulating conduct that takes place outside of
their borders, the Commerce Clause also prohibits laws where “the burden imposed on interstate
commerce is clearly excessive in relation to the putative local benefits.” Green Room LLC v.
Wyoming, 157 F.4th 1196, 1209 (10th Cir. 2025); see also Air Transp. Ass'n of Am., Inc. v. Moss,
No. 1:23-cv-02421-DDD-KAS, 2024 WL 4369137, at *5 (D. Colo. Sept. 30, 2024).
166.
In Pike, 397 U.S. at 142, the Supreme Court articulated the factors that a Court
should consider when assessing whether a law unduly burdens interstate commerce. Pike balancing
requires the Court to consider four factors: “(1) the nature of the putative local benefits advanced
by the [statute]; (2) the burden the [statute] imposes on interstate commerce; (3) whether the
burden is clearly excessive in relation to the local benefits; and (4) whether the local interests can
be promoted as well with a lesser impact on interstate commerce.” Johnson & Johnson Vision
61
PDF Page 63
Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 62 of 75
Care, Inc. v. Reyes, 665 F. App’x 736, 744 (10th Cir. 2016) (internal quotation marks omitted).
All four factors weigh in xAI’s favor here.
167.
First, the burden the statute imposes on interstate commerce is significant. As
explained above, supra at ¶¶ 150-63, SB24-205 regulates transactions consummated entirely
outside of Colorado between xAI and entities outside of Colorado for systems developed entirely
outside of Colorado.
168.
Not only does SB24-205 apply to transactions entirely outside of the state of
Colorado, but it also applies regardless of how unforeseeable use by a Colorado resident is. In
order for xAI to comply with SB24-205’s “algorithmic discrimination” requirements, it would be
required to alter its products nationwide to accommodate Colorado’s requirements. Moreover, to
the extent complying with SB24-205 would require retraining Grok, supra at ¶¶ 100-02, that will
require significant economic and engineering investment.
169.
Second, the nature of the putative local benefit is highly speculative, at best. SB24-
205 itself contains no findings and cites no evidence that AI systems are causing discrimination—
much less systemic discrimination. Additionally, federal statutes, including Title VII of the Civil
Rights Act, the Fair Housing Act, the Equal Credit Opportunity Act, and the Americans with
Disabilities Act, and their Colorado analogs already prohibit discriminatory outcomes in the same
subject matter that SB24-205 covers.
170.
Third, the burden imposed by SB24-205 is clearly excessive in relation to the local
benefits. AI innovation is critical to American economic and national security. As the United States
government put it, AI “ha[s] the potential to reshape the global balance of power” and “it is a
national security imperative for the United States to achieve and maintain unquestioned and
62
PDF Page 64
Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 63 of 75
unchallenged global technological dominance.” 70 “United States AI companies must be free to
innovate without cumbersome regulation,” “excessive State regulation thwarts this imperative,”
“State-by-State regulation by definition creates a patchwork of 50 different regulatory regimes that
makes compliance more challenging,” and “State laws sometimes impermissibly regulate beyond
State borders, impinging on interstate commerce.” 71 The White House more recently explained
that “winning the AI race” would “usher in a new era of human flourishing, economic
competitiveness, and national security for the American people,” “AI cannot become a vehicle for
government to dictate right and wrong-think,” and “[a] patchwork of conflicting state laws would
undermine American innovation and our ability to lead in the global AI race.”72 AI systems
function similarly to the internet, rail, or highway insofar as they “require[] a cohesive national
scheme of regulation so that users are reasonably able to determine their obligations.” ACLU v.
Johnson, 194 F.3d 1149, 1162 (10th Cir. 1999). These concerns apply on all fours to SB24-205,
which imposes broad and costly requirements on AI developers operating anywhere nationwide.
Even the Attorney General himself has expressed concern about the impact SB24-205 might have,
going as far as saying the bill “is really problematic.” 73
70
Plan.pdf.
https://www.whitehouse.gov/wp-content/uploads/2025/07/Americas-AI-Action-
71
https://www.whitehouse.gov/presidential-actions/2025/12/eliminating-state-lawobstruction-of-national-artificial-intelligence-policy/.
72
https://www.whitehouse.gov/articles/2026/03/president-donald-j-trump-unveilsnational-ai-legislative-framework/.
73
https://broadbandbreakfast.com/state-ag-warns-colorado-ai-bill-could-driveinnovation-out-of-state/; see also https://newspack-coloradosun.s3.amazonaws.com/wpcontent/uploads/2024/06/FINAL-DRAFT-AI-Statement-6-12-24-JP-PW-and-RR-Sig.pdf.
63
PDF Page 65
Case No. 1:26-cv-01515
171.
Document 1
filed 04/09/26
USDC Colorado
pg 64 of 75
Fourth, the putative local interests served by SB24-205 can be promoted through
lesser impacts on interstate commerce. Colorado already has comprehensive discrimination laws.
For example, the Colorado Anti-Discrimination Act prohibits discriminatory or unfair
employment and housing practices. SB24-205 does not impair compliance with that or other antidiscrimination laws and even expressly states that nothing in it “restricts a developer’s … ability
to: (a) [c]omply with federal, state, or municipal laws, ordinances, or regulations.” § 6-11705(1)(a).
172.
SB24-205’s requirement that AI developers prevent “algorithmic discrimination”
might also have unintended effects. AI developers may incidentally overcorrect, resulting in the
very discrimination SB24-205 purports to address. See Students for Fair Admissions, Inc. v.
President & Fellows of Harvard Coll., 600 U.S. 181, 223 (2023) (“SFFA”) (“Outright racial
balancing is ‘patently unconstitutional’” because equal protection “command[s] that the
Government must treat citizens as individuals, not as simply components of a racial, religious,
sexual or national class.” (cleaned up)).
173.
Further, such model-rebalancing efforts have no “logical end point.” Id. at 221. It
is not clear at what point such efforts would satisfy compliance such that the AI developer has
used “reasonable care to protect consumers from any known or reasonably foreseeable risks of
algorithmic discrimination.” § 6-1-1702(1).
174.
SB24-205 imposes an enormous burden on interstate commerce—regulating
entirely out-of-state transactions and threatening to undermine American AI innovation—relative
to its local benefit—speculative concerns about “algorithmic discrimination” that lack any
legislative findings or evidence.
64
PDF Page 66
Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 65 of 75
COUNT V
Declaratory Relief and Preliminary and Permanent Injunctive Relief for Vagueness in
Violation of the Due Process Clause of the Fourteenth Amendment to the
United States Constitution (42 U.S.C. § 1983, 28 U.S.C. § 2201(a))
175.
xAI re-alleges and incorporates by reference all allegations set forth above.
176.
“A fundamental principle in our legal system is that laws which regulate persons or
entities must give fair notice of conduct that is forbidden or required.” FCC v. Fox TV Stations,
567 U.S. 239, 253 (2012). A law is unconstitutionally vague if it “fails to provide a person of
ordinary intelligence fair notice of what is prohibited, or is so standardless that it authorizes or
encourages seriously discriminatory enforcement.” United States v. Williams, 553 U.S. 285, 304
(2008). “When speech is involved, rigorous adherence to those requirements is necessary to ensure
that ambiguity does not chill protected speech.” Fox, 567 U.S. at 253-54. After all, vague laws risk
chilling speech by forcing speakers “to ‘steer far wider of the unlawful zone’” than they would “if
the boundaries of the forbidden areas were clearly marked.” Baggett v. Bullitt, 377 U.S. 360, 372
(1964). For that reason, laws touching on speech must themselves speak “with narrow specificity.”
NAACP v. Button, 371 U.S. 415, 433 (1963).
177.
SB24-205 is not written “with narrow specificity,” and it fails to provide fair notice
to a person of ordinary intelligence as to what it requires.
178.
To begin, it fails to adequately articulate what “algorithmic discrimination” is.
SB24-205 defines “algorithmic discrimination” as “an unlawful differential treatment or impact
that disfavors an individual or group of individuals on the basis of their actual or perceived age,
color, disability, ethnicity, genetic information, limited proficiency in the English language,
national origin, race, religion, reproductive health, sex, veteran status, or other classification
protected under the laws of this state or federal law.” § 6-1-1701(1)(a).
65
PDF Page 67
Case No. 1:26-cv-01515
179.
Document 1
filed 04/09/26
USDC Colorado
pg 66 of 75
But serious questions are left unanswered. What constitutes a “perceived attribute”?
“Perceived” by whom? Is it the perception of the individual who is allegedly being discriminated
against or the AI system allegedly resulting in a disparate impact? If the former, how sincerely
held does that perception need to be? How do “unlawful differential treatment” or “impact” differ?
And how do those tests differ from the traditional legal frameworks for disparate treatment?
180.
The vagueness of SB24-205’s “algorithmic discrimination” is further compounded
by its carve-outs. § 6-1-1701(1)(b). SB24-205 provides that “‘Algorithmic discrimination’ does
not include: . . . (B) [e]xpanding an applicant, customer, or participant pool to increase diversity
or redress historical discrimination.” Id. What constitutes “historical discrimination”? How far
back in time must one go? Does it only include the historical treatment of African Americans
arising from slavery, or does it include discrimination against the Irish and Italians in the 19th
century? 74 Nuziard v. Minority Bus. Dev. Agency, 721 F. Supp. 3d 431, 509 (N.D. Tex. 2024) (“In
the mid-19th Century, Irish immigrants faced a common roadblock when seeking jobs after their
trans-Atlantic voyage: motivated by rising populism and anti-Irish animus, many employers hung
signs explaining ‘Irish Need Not Apply.’”). Likewise, what is permissible “diversity”? § 6-1-1701
defines algorithmic discrimination to include virtually every actual or perceived protected trait;
which of those same traits are the appropriate subjects of State-approved “diversity”
considerations?
181.
Next, take SB24-205’s definition of a “high-risk artificial intelligence system,”
which “means any artificial intelligence system that, when deployed, makes, or is a substantial
74
https://www.nytimes.com/interactive/2019/10/12/opinion/columbus-day-italianamerican-racism.html.
66
PDF Page 68
Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 67 of 75
factor in making, a consequential decision.” § 6-1-1701(9)(a). When does an AI system become a
“substantial factor”? According to SB24-205, “substantial factor” “includes any use of an artificial
intelligence system to generate any content, decision, prediction, or recommendation concerning
a consumer that is used as a basis to make a consequential decision concerning the consumer.” §
6-1-1701(11)(b) (emphasis added). That definition—“any use” of an AI system to generate “any
content … that is used as a basis to make a [] decision”—is so amorphous it encompasses virtually
all uses of AI systems.
182.
Now, take xAI’s obligation to prevent “algorithmic discrimination.” § 6-1-1702.
SB24-205 requires xAI to use “reasonable care to protect consumers from any known or
reasonably foreseeable risks of algorithmic discrimination.” “[R]easonable care” and “reasonably
foreseeable” are undefined and uncertain.
183.
SB24-205’s disclosure provisions fare no better. For example, SB24-205 requires
developers to disclose “high-level summaries of the type of data used to train the high-risk artificial
intelligence system” as well as a “general statement describing the reasonably foreseeable uses
and known harmful or inappropriate uses of the high-risk artificial intelligence system.” §§ 6-11702(2)(b)(I), (2)(a). How “high-level” and how “general” do disclosures need to be to comply?
Likewise, what constitutes “harmful or inappropriate use,” which is, again, undefined?
184.
Similarly, the statute requires disclosure to the Attorney General of “known or
reasonably foreseeable risks” within ninety days of discovering an AI system caused algorithmic
discrimination, was “reasonably likely” to have caused algorithmic discrimination, or of receiving
a “credible report that the high-risk artificial intelligence system has been deployed and has caused
algorithmic discrimination.” § 6-1-1702(5)(b). When does algorithmic discrimination become
67
PDF Page 69
Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 68 of 75
“reasonably likely”? And what constitutes a “credible report” such that an AI developer is required
to comply with the disclosure provision? Would a single consumer report satisfy either trigger? If
not, how many consumer reports must an AI developer receive before it takes action and what
must those reports contain such that they are sufficiently “credible.” SB24-205 is long on questions
and short on answers.
185.
What’s worse, SB24-205 gives the Attorney General a massive amount of
interpretive and rulemaking power. SB24-205 gives the Colorado Attorney General rulemaking
authority to implement the law and has exclusive authority to enforce it. §§ 6-1-1706(1)(2), 6-11707. The Attorney General alone thus has authority to decide what actions violate SB24-205 and
its nebulous provisions and determine who to enforce it against.
186.
The statute’s use of vague, relative and uncertain terms gives developers no fair
notice of what mitigation or disclosures are required, and hands the Colorado Attorney General
virtually unfettered discretion to enforce the law arbitrarily against his political opponents. SB24205’s terms are so malleable that it authorizes “arbitrary [or] discriminatory enforcement.”
Grayned v. City of Rockford, 408 U.S. 104, 108 (1972).
COUNT VI
Declaratory Relief and Preliminary and Permanent Injunctive Relief for Violation of the
Equal Protection Clause of the Fourteenth Amendment to the United States Constitution
(42 U.S.C. § 1983, 28 U.S.C. § 2201(a))
187.
xAI re-alleges and incorporates by reference all allegations set forth above.
188.
“Proposed by Congress and ratified by the States in the wake of the Civil War, the
Fourteenth Amendment provides that no State shall ‘deny to any person . . . the equal protection
of the laws.’” “Proponents of the Equal Protection Clause described its ‘foundation[al] principle’
as ‘not permit[ing] any distinctions of law based on race or color.’” “Any ‘law which operates
68
PDF Page 70
Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 69 of 75
upon one man,’ they maintained, should ‘operate equally upon all.’ Accordingly, as th[e]
[Supreme] Court's early decisions interpreting the Equal Protection Clause explained, the
Fourteenth Amendment guaranteed ‘that the law in the States shall be the same for the black as for
the white; that all persons, whether colored or white, shall stand equal before the laws of the
States.’” SFFA, 600 U.S. at 183 (citations omitted). “Eliminating racial discrimination means
eliminating all of it.” Id. at 183-84.
189.
“Given its drafters’ desire to eschew race-based distinctions, even benign
discrimination cannot be squared with the Equal Protection Clause—it can only be an ‘exception.’”
Nuziard, 721 F. Supp. 3d at 477-78 (citation omitted). “Any exceptions to the Equal Protection
Clause’s guarantee must survive a daunting two-step examination known as ‘strict scrutiny,’ which
asks first whether the racial classification is used to ‘further compelling governmental interests,’
and second whether the government’s use of race is ‘narrowly tailored,’ i.e., ‘necessary,’ to achieve
that interest.” SFFA, 600 U.S. at 184 (citations omitted). SB24-205’s requirement to mitigate
“algorithmic discrimination” flunks that test.
190.
SB24-205 requires xAI to mitigate against and make disclosures concerning
“algorithmic discrimination.” §§ 6-1-1701(1), 1702. According to Colorado, however, not all
algorithmic discrimination is impermissible “algorithmic discrimination.” While xAI has a general
duty to mitigate against “algorithmic discrimination,” “‘[a]lgorithmic discrimination’ does not
include: (I) [t]he offer, license, or use of a high-risk artificial intelligence system by a developer
or deployer for the sole purpose of … (B) [e]xpanding an applicant, customer, or participant pool
to increase diversity or redress historical discrimination.” § 6-1-1701(1)(b)(I)(B). In other words,
if xAI engages in balancing to “increase diversity” or “redress historical discrimination,” including
69
PDF Page 71
Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 70 of 75
on the basis of “race,” it does not engage in “algorithmic discrimination” within the meaning of
SB24-205 at all. Id.; see also § 6-1-1701(1).
191.
Accordingly, if an AI developer alters its model to provide preferencing mandated
by Colorado to increase diversity or redress historical discrimination, it is effectively exempt from
SB24-205’s mitigation requirements. That is because if an AI developer alters its AI system to
increase “diversity” or “redress historical discrimination” then any disparate treatment that occurs
as a result does not constitute “algorithmic discrimination.” Therefore, the AI developer need not
use “reasonable care to protect consumers from any known or reasonably foreseeable risks of
algorithmic discrimination” because no such “algorithmic discrimination” could occur. § 6-11702.
192.
SB24-205 thus presents xAI with a choice. xAI can either continue to develop Grok
as a maximally truth-seeking AI system that provides objective and unbiased outputs; expend
significant resources on compliance, documentation, and risk mitigation; and face full enforcement
risk under SB24-205 for any disparate impact. Or it can build an AI with the sole purpose of
increasing diversity or redressing historical discrimination, and receive a safe harbor from potential
liability under SB24-205. That is an impermissible choice. Affording benefits to some, but not
others, based on the treatment of protected characteristics, including race, is precisely the type of
“[d]istinction[] between citizens solely because of their ancestry [that] are by their very nature
odious to a free people whose institutions are founded upon the doctrine of equality.” SFFA, 600
U.S. at 184 (citations omitted).
70
PDF Page 72
Case No. 1:26-cv-01515
193.
Document 1
filed 04/09/26
USDC Colorado
pg 71 of 75
SB24-205 includes no justification, much less a compelling justification, for its
codified discrimination. Again, it includes no statement of purpose. SB24-205 does not even define
what “diversity” is or what “historical discrimination” it seeks to redress.
194.
At best, SB24-205’s discrimination provision seems to be designed to embody
Colorado’s subjective vision of equity and remedy amorphous discrimination across a wide range
of services and opportunities, including education, employment, financial or lending services,
health-care services, housing, essential government services, insurance and legal services. § 6-11701(3). But “an effort to alleviate the effects of societal discrimination is not a compelling
interest.” Shaw v. Hunt, 517 U.S. 899, 909 (1996). If it was, the Equal Protection Clause “would
be lost in a mosaic of shifting preferences based on inherently unmeasurable claims of past
wrongs.” City of Richmond v. J.A. Croson Co., 488 U.S. 469, 505-06 (1989). “The mere recitation
of a benign or compensatory purpose for the use of a racial classification would essentially entitle
the States to exercise the full power of Congress under § 5 of the Fourteenth Amendment and
insulate any racial classification from judicial scrutiny under § 1. We believe that such a result
would be contrary to the intentions of the Framers of the Fourteenth Amendment, who desired to
place clear limits on the States’ use of race as a criterion for legislative action, and to have the
federal courts enforce those limitations.” Id. at 490-91 (citations omitted).
195.
To the extent that Colorado believes SB24-205 addresses specific instances of
discrimination, it must provide facts to support that conclusion. “First, it must identify the past or
present discrimination ‘with some specificity.’ Second, it must also demonstrate that a ‘strong
basis in evidence’ supports its conclusion that remedial action is necessary.” Concrete Works of
Colo. v. City & Cty. of Denver, 321 F.3d 950, 958 (10th Cir. 2003) (citations omitted); see also
71
PDF Page 73
Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 72 of 75
Nuziard, 721 F. Supp. 3d at 480. Colorado “cannot point to general social ills and call it a day.
Rather, it must identify the ‘who, what, when, where, why, and how’ of relevant discrimination.
Nuziard, 721 F. Supp. 3d at 480 (citing Croson, 488 U.S. at 492). Yet SB24-205 includes no
legislative findings or evidence whatsoever, much less any that would prove the existence of
specific harm that the state has a compelling interest in redressing.
196.
SB24-205 is also not narrowly tailored. In fact, the provision excepting permissible
“algorithmic discrimination” applies to any protected characteristic so long as their inclusion
increases “diversity” or redresses “historical discrimination.” § 6-1-1701(1)(b)(I)(B). It also
applies to virtually any decision that yields disparate results because SB24-205 covers nearly every
subject matter—education, employment, finance, government services, healthcare, housing,
insurance, and legal services. § 6-1-1701(3). That is woefully overinclusive. SB24-205 encodes
State preferencing for virtually any protected characteristic across virtually every subject matter.
Take employment for example: “If a [preference] was ‘narrowly tailored’ to compensate black
contractors for past discrimination, one may legitimately ask why they are forced to share this
‘remedial relief’ with an Aleut citizen who moves to [Boulder] tomorrow?” Croson, 488 U.S. at
506.
197.
SB24-205 is overinclusive for another reason: its carveout for “diversity” and
“historical discrimination” paints in broad strokes and encompasses members with protected traits
who may personally experience no discrimination or disparate impact. As Governor Polis
explained when expressing concern over the bill, “[l]aws that seek to prevent discrimination
72
PDF Page 74
Case No. 1:26-cv-01515
Document 1
filed 04/09/26
USDC Colorado
pg 73 of 75
generally focus on prohibiting intentional discriminatory conduct,” but SB24-205 “deviates from
that practice by regulating the results of AI system use, regardless of intent.” 75
198.
SB24-205 is also underinclusive because the definition of “algorithmic
discrimination” does not include protections for individuals or groups that may actually face
disparate impact in the covered subjects (housing, finance, legal services, etc.), including the
economically disadvantaged. As one Court succinctly put it, “Oprah Winfrey is presumptively
disadvantaged, while … even more disadvantaged Americans are not.” Nuziard, 721 F. Supp. 3d
at 491.
199.
SB24-205 also lacks a “logical end point.” SFFA, 600 U.S. at 221. SB24-205
includes no guidance indicating when “increas[ing] diversity” or “redress[ing] historical
discrimination” will be sufficiently addressed by changes in xAI’s model such that further
inclusion of protected groups would actually decrease “diversity” or affirmatively cause “historical
discrimination” of excluded majority groups. SB24-205’s extraordinarily broad affirmative
discrimination provisions make it “unclear how a court,” or xAI for that matter “is supposed to
determine if or when such goals would be adequately met.” Id.
200.
Lastly, as explained above, less restrictive alternatives exist. State and federal
law—for example, the Civil Rights Act of 1964, Fair Housing Act, Equal Credit Opportunity Act,
and Colorado Anti-Discrimination Act—already prohibit intentional discrimination in
employment, housing, education, finance, and other decisions within SB24-205’s scope.
75
Jared Polis, Statement on Signing SB24-205 (May 17, 2024),
https://drive.google.com/file/d/1i2cA3IG93VViNbzXu9LPgbTrZGqhyRgM/view.
73
PDF Page 75
Case No. 1:26-cv-01515
201.
Document 1
filed 04/09/26
USDC Colorado
pg 74 of 75
SB24-205’s carveout for permissible “algorithmic discrimination” is as amorphous
as it is unconstitutional, and forces xAI to choose between adopting Colorado’s state-ordained
preferencing in its AI model or continuing to operate its maximally-truth seeking AI and expose
itself to liability, including civil penalties.
PRAYER FOR RELIEF
For the foregoing reasons, xAI respectfully requests that the Court:
A.
Enter a judgment declaring, pursuant to 28 U.S.C. §§ 2201(a) and 2202, that SB24-
205 unconstitutionally compels xAI’s speech in violation of the First Amendment of the U.S.
Constitution;
B.
Enter a judgment declaring, pursuant to 28 U.S.C. §§ 2201(a) and 2202, that SB24-
205 constitutes an impermissible extraterritorial regulation;
C.
Enter a judgment declaring, pursuant to 28 U.S.C. §§ 2201(a) and 2202, that SB24-
205 imposes an undue burden on interstate commerce;
D.
Enter a judgment declaring, pursuant to 28 U.S.C. §§ 2201(a) and 2202, that SB24-
205 is unconstitutionally vague in violation of the Due Process Clause of the U.S. Constitution;
E.
Enter a judgment declaring, pursuant to 28 U.S.C. §§ 2201(a) and 2202, that SB24-
205 violates the Equal Protection Clause of the Fourteenth Amendment of the U.S. Constitution;
F.
Issue an order permanently enjoining Attorney General Weiser, as well as all
officers, agents, and employees subject to his supervision, direction, or control, from enforcing the
provisions of SB24-205 against xAI;
G.
Award xAI all costs and attorney’s fees to which it is entitled pursuant to 42 U.S.C.
§ 1988 and other applicable law; and
74
PDF Page 76
Case No. 1:26-cv-01515
H.
Document 1
filed 04/09/26
USDC Colorado
pg 75 of 75
Grant xAI such other and further relief as the Court deems just and proper.
Dated: April 9, 2026.
Respectfully submitted,
s/ Frederick R. Yarger
Frederick R. Yarger
William D. Hauptman
Wheeler Trigg O’Donnell LLP
370 Seventeenth Street, Suite 4500
Denver, CO 80202
Telephone: 303.244.1800
Facsimile: 303.244.1879
Email: yarger@wtotrial.com
hauptman@wtotrial.com
James Burnham
X.AI LLC
601 13th Street NW 12th Floor
Washington, DC 20005
Email: jburnham@x.ai
Adam Mehes
Argirios J. Nickas
X.AI LLC
249 W 17th Street
New York, NY 10011
Email: amehes@x.ai
anickas@x.ai
Reid Coleman
X.AI LLC
865 FM 1209, Building 2
Bastrop, TX 78602
Email: rcoleman@x.ai
Attorneys for Plaintiff X.AI LLC
75
Third party AI contribution in government documents
Possible
Document content is reproduced directly from government sources. Some filings in our database may themselves contain AI-generated content submitted by third parties to government agencies.
Back to 
