ICLE Comments to the FTC on AI Suppression
Introduction
We thank the Federal Trade Commission (“FTC” or “Commission”) for the opportunity to comment on its proposed Policy Statement Concerning the Suppression of Accuracy in Artificial Intelligence Systems (“AI Policy Statement” or “Statement”). The comment period gives the public an opportunity to inform the Commission’s consideration of how the First Amendment constrains government regulation of artificial intelligence (“AI”).[1]
The International Center for Law & Economics (“ICLE”) is a nonprofit, nonpartisan research and policy center dedicated to developing economically grounded approaches to law and public policy. ICLE applies law & economics methods and economic research to policy debates and has longstanding expertise in competition, consumer protection, innovation, and technology policy.
ICLE has a strong interest in ensuring that First Amendment and consumer protection law serve the public interest through clear legal rules and sound economic analysis. ICLE scholars have written extensively about the regulation of online speech platforms, free expression, competition, and consumer protection. They have also examined how the FTC’s competition and consumer protection missions complement one another and how excessive consumer protection regulation—even when well intentioned—can impede innovation and competition.[2] The AI Policy Statement recognizes several of these concerns. ICLE is therefore well positioned to address the legal and economic issues raised by the Statement.
The FTC gets several important points right.
First, uniform federal AI policies may protect consumers and competition more effectively than a patchwork of state regulations. The Statement correctly calls for a “national AI framework” that would “protect innovation and competition by providing national regulatory clarity and certainty and avoiding a balkanized or patchwork regulatory approach driven by States—or, most dangerously, imposed by certain anti-innovation State governments on the rest of the country.”[3]
Second, excessive AI regulation at either the federal or state level can impede innovation, weaken competition, and harm consumers across many markets.[4] Such regulation may also raise concerns beyond antitrust and the Commission’s competition and consumer protection authority.[5] Poorly designed rules can increase consumers’ information costs by restricting access to truthful, non-misleading speech.
Third, the Commission correctly recognizes that AI providers generally remain subject to the Federal Trade Commission Act. As the Statement explains, AI is “an umbrella term covering a universe of different tools and systems.”[6] It includes a broad and evolving range of products and services. AI providers generally do not fall within the categories of persons, partnerships, or corporations excluded from the Commission’s authority under Section 5(a)(2). Nor do they typically qualify for the bona fide nonprofit exemption under Section 4.[7]
The proposed Statement nevertheless requires substantial revision before the Commission finalizes it.
First, the Statement offers little practical guidance about how the Commission will apply its deception authority to AI. Except for a few straightforward examples, it does not identify which acts or practices the Commission would consider violations of Section 5.
Second, the Statement largely avoids the difficult question of when the conduct of AI providers, or their interactions with users, constitutes speech protected by the First Amendment. It also does not explain how those protections would constrain FTC enforcement. The Statement’s focus on “ideologically motivated distortions” suggests that the Commission’s concerns extend beyond factual misrepresentations in marketing to speech that may receive the highest degree of First Amendment protection.[8]
Third, the Statement’s discussion of federal preemption requires substantial clarification. We understand that the Commission addresses preemption at least partly in response to Executive Order 14365, Ensuring a National Policy Framework for Artificial Intelligence.[9] Even accounting for the executive order’s requirements, the Statement’s discussion is so brief that it may mislead regulated parties and courts.
The Commission may intend to say that compliance with state consumer protection laws governing AI products and services would not provide an affirmative defense to liability under Section 5. If so, a more clearly explained and illustrated version of that point could provide useful guidance. Established forms of implied preemption, though, fit poorly with the FTC Act and settled principles of federalism.
Courts have not held that the FTC Act occupies the field of state advertising, marketing, or related consumer protection law so completely that it leaves no room for state regulation.[10] Conflict preemption also appears inapposite. We are unaware of any relevant state AI law or regulation—however sound or unsound as policy—that makes simultaneous compliance with state law and the FTC Act impossible.[11] As Part III explains, obstacle preemption appears even less applicable because state AI laws do not necessarily frustrate the purposes of the FTC Act.[12] In the end, the Statement’s current discussion is too vague to provide the clear and useful explanation contemplated by the executive order.
These comments explain why constitutional law and economic analysis impose meaningful limits on FTC enforcement under the proposed Statement. The Commission should provide clearer guidance about how it will apply its deception authority within those limits.
Part I explains the First Amendment principles that protect the “marketplace of ideas.” Those principles protect, to varying degrees, both the editorial discretion of AI companies that train and develop large language models (“LLMs”) and the rights of users who interact with them. Market forces also discipline AI companies whose systems repeatedly provide false or unreliable responses.
Part II examines the difficulty of applying the Commission’s deception authority to generative AI providers based on disfavored speech. Section 5 permits the Commission to police deceptive acts or practices, but liability ordinarily requires a material representation, omission, or promise likely to mislead reasonable consumers. We examine the representations made by major AI providers and conclude that the Commission may lack a viable basis for enforcement under the proposed Statement.
Part III analyzes the Statement’s treatment of preemption. State AI laws generally do not make compliance with the FTC Act impossible, as conflict preemption requires. The Statement is also agency guidance and therefore is unlikely to receive substantial judicial deference.
I. The First Amendment Protects AI Editorial Judgment and User Access[13]
The defense of free speech against government censorship has a long pedigree. Its advocates include John Stuart Mill and, long before the Bill of Rights, John Milton.[14] Although neither used the precise phrase, their defenses of free expression helped establish the theory later described as the “marketplace of ideas.” That theory bears directly on the AI Policy Statement because AI chatbots have become a new forum for producing, receiving, and testing ideas. Complaints about whether chatbots are ideologically neutral[15] must therefore account for the First Amendment.
Mill offered four grounds for protecting free speech. A censored opinion may be true. An erroneous opinion may contain part of the truth, which can emerge only through confrontation with opposing views. A true opinion must still face challenge so that those who hold it understand its basis. And even a true opinion can become an inherited dogma rather than a genuine conviction if no one contests it.[16] The exchange of competing ideas allows truth to emerge and remain meaningful.
Thomas Jefferson advanced a similar principle in his First Inaugural Address after a divisive election. He urged tolerance for divergent political views and argued that even those who favored dissolving the Union or changing its republican form should remain “undisturbed as monuments of the safety with which error of opinion may be tolerated where reason is left free to combat it.”[17]
The Supreme Court later adopted the marketplace-of-ideas metaphor, initially through Justice Oliver Wendell Holmes’ dissenting opinions.[18] Variations of the phrase have since appeared in thousands of federal First Amendment decisions.[19]
The Constitution protects private parties’ authority to decide which speech they will disseminate, largely free from government control. As the Court explained in Manhattan Community Access Corp. v. Halleck, “[t]he Free Speech Clause of the First Amendment constrains governmental actors and protects private actors . . . .”[20]
Private parties often participate in the marketplace of ideas by creating expressive products. In Moody v. NetChoice,[21] the Supreme Court considered whether state laws regulating social-media platforms and other websites facially violated the First Amendment.[22] The Court explained that the First Amendment protects an entity engaged in expressive activity, including one that compiles and curates others’ speech, when the government directs it to carry messages it would prefer to exclude.[23]
After reviewing its compelled-speech precedents, the Court concluded that government efforts to alter the views social-media platforms include or exclude from their principal feeds interfere with protected speech.[24] It also warned that, “[h]owever imperfect the private marketplace of ideas,” government decisions about whether speech is imbalanced, followed by coercion requiring more or less of particular views, offer “a worse proposal.”[25]
Generative AI companies likewise exercise protected editorial judgment. They select training materials, determine how models process those materials, and establish policies governing the outputs their systems generate. Government efforts to rebalance those choices would interfere with that editorial discretion. The Commission should therefore exercise caution before bringing an action against an AI provider for offering “ideologically motivated distortions in a response to a factual question.”[26]
Justice Amy Coney Barrett observed in her Moody concurrence that a “function qualifies for First Amendment protection only if it is inherently expressive.”[27] AI providers’ activities readily meet that standard. Selecting inputs, training models, and reviewing outputs for compliance with company policies all require editorial judgment. An AI chatbot’s answers are themselves an expressive product.[28]
AI providers resemble search engines, which also produce expressive responses to user queries. Several federal district courts have recognized that search engines possess a First Amendment interest in their results. Some courts have treated the results themselves as protected speech. Others have focused on the editorial judgment used to generate and arrange them. Under either approach, generative AI responses warrant similar protection.
In Jian Zhang v. Baidu.com,[29] the U.S. District Court for the Southern District of New York held that applying a New York public-accommodations law to a Chinese search engine accused of censoring pro-democracy speech would conflict with the search engine’s editorial discretion. The court explained that “there is a strong argument to be made that the First Amendment fully immunizes search-engine results from most, if not all, kinds of civil liability and government regulation.”[30]
The court also described the editorial judgments inherent in search. A search engine retrieves information from a vast body of online material, determines which information is relevant, organizes it for users, and decides where each result will appear.[31] Other courts have likewise recognized search engines’ right to exercise editorial discretion over their results.[32]
At least one court has treated search results themselves as protected opinions. In Search King Inc. v. Google Technology, Inc.,[33] the court explained that search results “are opinions—opinions of the significance of particular web sites as they correspond to a search query.” Different search engines may reach different conclusions because each uses its own method to determine relevance and significance.[34]
AI chatbot responses should receive no less protection than social-media feeds or search-engine results. AI providers exercise editorial judgment when they create, train, and maintain their systems, and the systems’ outputs are expressive. An allegation of ideological bias in an expressive product, standing alone, cannot supply a permissible basis for government enforcement.[35]
Users also have a First Amendment interest in receiving chatbot responses. The Supreme Court has held that the First Amendment protects the right to receive speech.[36] That protection covers listening to speakers,[37] reading pamphlets[38] and books,[39] receiving advertisements,[40] playing video games,[41] and using social media.[42] Interacting with ideas generated by an AI chatbot warrants the same protection.
The Commission is therefore correct to focus on its deception authority under Section 5 rather than its unfairness authority.[43] An unfairness claim would require the Commission to establish three elements. The practice must cause or be likely to cause substantial consumer injury. That injury must not be outweighed by countervailing benefits to consumers or competition. Consumers also must be unable reasonably to avoid the injury.[44]
A claim based on inaccurate or misleading chatbot responses would treat speech itself as the substantial injury. That theory conflicts with First Amendment precedent.
Even false noncommercial speech presumptively receives full First Amendment protection.[45] In United States v. Alvarez,[46] a Supreme Court plurality explained that prior statements denying protection to false speech concerned only the limited “historic and traditional categories [of expression] long familiar to the bar.”[47] The Court rejected a general First Amendment exception for false statements.[48] A rule directed at false speech as such would therefore face strict scrutiny.
To avoid strict scrutiny, an unfairness claim would need an independent justification tied to a category of low-value speech that the Supreme Court has recognized as unprotected.[49] Those categories include defamation, fraud, false light, false statements to government officials, perjury, impersonating government officials, and speech integral to criminal conduct.[50] A claim that inaccurate information alone constitutes substantial consumer injury would likely fail.
Alvarez illustrates the problem. The defendant falsely claimed at a public meeting that he had received the Congressional Medal of Honor.[51] The statement left “no room to argue about interpretation or shades of meaning.” Even so, the Court held that the restriction was content based and subject to strict scrutiny.[52] An FTC claim treating false or misleading chatbot output as a substantial injury would rest on similarly uncertain constitutional ground.
Even assuming a legally cognizable injury, the Commission would likely struggle to show that the injury outweighs consumer and competitive benefits. The First Amendment’s preference for less speech-restrictive alternatives would compound that difficulty.
AI companies participate in the marketplace of ideas by offering users what they regard as the best answers to their questions. Providers that repeatedly fail to satisfy users risk losing them to competing chatbots. Market discipline has already operated in prominent cases. When Google’s chatbot generated ideologically charged false outputs, substantial public criticism prompted the company to revise the product quickly.[53]
False information may also contribute to public debate by prompting correction and exposing weaknesses in competing claims. Suppression is rarely the least speech-restrictive response. As the Alvarez Court explained:
The remedy for speech that is false is speech that is true… Freedom of speech and thought flows not from the beneficence of the state but from the inalienable rights of the person. And suppression of speech by the government can make exposure of falsity more difficult, not less so. Society has the right and civic duty to engage in open, dynamic, rational discourse. These ends are not well served when the government seeks to orchestrate public discussion through content-based mandates.[54]
Competition among AI providers also gives consumers meaningful ways to avoid services they regard as inaccurate or ideologically slanted. The Statement recognizes that AI companies increasingly market their products as more accurate or ideologically neutral than competing systems.[55] Consumers can compare those claims, test competing products, and switch providers when a chatbot fails to meet their expectations.
II. Section 5 Deception Claims Require Concrete, Material Misrepresentations[56]
Quoting the “AI National Policy Fact Sheet,”[57] the AI Policy Statement states that “States such as California and Colorado are considering requiring AI companies to censor outputs and insert left-wing ideology in their programming.”[58] Such requirements may raise legitimate concerns because the First Amendment protects speakers against censorship and compelled speech.
The Statement, though, does not analyze the relevant state laws or explain how they might require AI providers to engage in conduct that qualifies as deception under Section 5. Its vague reference to “left-wing ideology” instead suggests concern about political viewpoints expressed in chatbot outputs. The First Amendment would generally protect such speech against FTC enforcement, regardless of the merits of the ideology expressed.
AI providers remain subject to the FTC Act, and the First Amendment does not protect every use of speech. The Supreme Court has long distinguished commercial speech, most commonly advertising, from other protected expression.[59] Commercial speech receives First Amendment protection, but courts generally afford it less protection than noncommercial speech.
The Commission may therefore bring careful, fact-specific deception cases involving false or misleading claims in advertising or marketing materials, consistent with its Policy Statement on Deception (“Deception Statement”).[60] Such enforcement also accords with decisions permitting the government to regulate fraud and other unlawful conduct carried out through speech.
Advertising and marketing materials may violate Section 5 when they expressly or implicitly make false or misleading factual claims about material attributes of AI products or services. The government has a legitimate interest in regulating such false commercial speech. Proper application of Section 5 under the Deception Statement can thereby protect consumers and competition against commercial fraud without intruding on protected expression.
Even where false speech falls outside First Amendment protection,[61] a fraud claim[62] requires more than proof of falsity.[63] As the Supreme Court explained in Illinois ex rel. Madigan v. Telemarketing Assocs., “[s]imply labeling an action one for ‘fraud’ . . . will not carry the day.”[64] The Court has, for example, repeatedly invalidated prophylactic restrictions on charitable solicitation because they imposed prior restraints without requiring proof of fraud.[65]
A properly tailored fraud action places the full burden of proof on the government. A false statement alone does not establish liability. Exacting proof requirements are necessary to preserve sufficient breathing room for protected speech.[66]
The Commission must therefore distinguish commercial representations from fully protected expression before applying its deception authority to an AI provider. It should avoid guidance, and especially enforcement actions, that overlooks the First Amendment’s restrictions on government efforts to regulate or chill expressive speech.
The Deception Statement establishes demanding proof requirements even when the challenged speech is commercial.[67] It identifies three elements of deception.
First, “there must be a representation, omission or practice that is likely to mislead the consumer.”[68] Second, the Commission evaluates the conduct “from the perspective of a consumer [or specific group] acting reasonably in the circumstances.”[69] A claim is material when it is “likely to affect the consumer’s conduct or decision with regard to a product or service,” such that “consumer injury is likely, because consumers are likely to have chosen differently [in the marketplace] but for the deception.”[70]
The Commission must begin by identifying a representation, omission, or practice. Express claims about an AI product’s policies or capabilities may satisfy that requirement, particularly when they appear in consumer advertising. Implied claims and omissions may also be actionable, but the Commission should exercise caution before inferring what a reasonable consumer understands or expects in a new and rapidly changing market.
Some objective product claims may permit straightforward interpretation under established enforcement principles. Others may require substantial investigation, economic analysis, and consumer testing. Even in advertising cases, the Commission should account for its longstanding recognition that advertising can improve consumer welfare and competition by providing information and facilitating comparison.[71]
Section 5 enforcement can produce costs through both false positives and false negatives. An overbroad approach to deception may suppress useful speech and encourage AI providers to offer consumers less information. Commission guidance should therefore acknowledge the variety of AI products and uses that the Statement itself recognizes.
General assertions of liability should be carefully defined and supported. The Commission should examine what AI providers actually say in their policies, advertisements, and other marketing materials, along with the likely and demonstrable effects of those representations.
Published AI policies and marketing materials could, in principle, mislead reasonable consumers. The representations cited in the Statement, though, appear limited. Major providers do not promise certainty or perfect accuracy.[72] Instead, they expressly warn users that outputs may contain errors.
For example:
- OpenAI states that “[a]ccuracy will never reach 100%.”[73] Its terms of use warn that outputs “may not always be accurate” and instruct users to evaluate them for “accuracy and appropriateness.”[74]
- Anthropic’s consumer terms state that “[o]utputs may not always be accurate and may contain material inaccuracies even if they appear accurate.” They also instruct users not to rely on outputs or actions without independently confirming their accuracy.[75]
- DeepSeek’s terms state that outputs “may contain errors or omissions and are for your reference only.” They warn users not to treat outputs as professional advice and explain that outputs may contain “incorrect, incomplete or inaccurate content.”[76]
- Gab AI’s terms state that outputs “may not always be accurate” and place responsibility on users to evaluate outputs for accuracy and appropriateness. They also warn that the service may provide incomplete, incorrect, or offensive outputs.[77]
- Grok’s terms state that outputs “may not always be accurate,” instruct users to conduct their own research, and warn them not to rely on outputs as truth. The terms also state that the service may provide incomplete, incorrect, or offensive outputs.[78]
- Google warns that Gemini “may sometimes provide inaccurate or offensive content.” It also instructs users not to rely on the service for medical, legal, financial, or other professional advice.[79]
A provider’s policies may contain isolated statements that appear to promise accuracy. A court would still need to read those statements in context and determine whether they are material. The FTC often presumes that express claims are material, but a court may not treat a general assertion such as Grok’s description of itself as a “truth-seeking AI companion for unfiltered answers”[80] as material. Without a more specific assurance, the statement resembles nonactionable marketing puffery.[81]
Terms of service cannot always cure a deceptive representation or omission. Still, the broad claims about accuracy cited in the Statement appear unlikely to overcome providers’ much more specific and detailed disclaimers.
The Commission should explain when an AI provider’s representations could become deceptive despite express warnings that its products may produce inaccurate outputs. None of the policies claims perfect accuracy or complete ideological neutrality. The record therefore offers little basis to conclude that a reasonable consumer would expect either.
AI providers compete to offer chatbots that users regard as accurate, reliable, and useful. Advertising those products while clearly disclosing that their outputs may contain errors does not, without more, constitute deception. The Commission may investigate whether providers honor their concrete representations. Mere disagreement with the accuracy of a particular chatbot response would not ordinarily establish a Section 5 violation.
An enforcement theory based on the Commission’s own assessment of whether individual chatbot outputs are sufficiently accurate would require the agency to determine the truth of potentially vast numbers of expressive responses. That approach would recreate the First Amendment problems discussed in Part I.
III. The Statement’s Preemption Analysis Offers Little Practical Guidance
The AI Policy Statement’s treatment of federal preemption does not provide useful guidance on a difficult question of federalism and constitutional law. With little explanation and no directly applicable precedent, the Statement declares that “[a] state law that requires an AI firm to deceive its customers obviously conflicts with Section 5’s express purpose of protecting consumers from such conduct.”[82] That broad assertion may mislead regulated parties about the circumstances in which the FTC Act preempts state law.
Federal law, including the FTC Act, may preempt conflicting state laws and regulations.[83] Courts generally disfavor implied preemption, however, and usually analyze it under one of two doctrines. Field preemption applies when federal law occupies an entire area of regulation. Conflict preemption applies when compliance with both federal and state law is impossible or when state law obstructs Congress’ purposes. Neither doctrine fits the Statement’s analysis. The Statement also does not explain which state laws might be preempted, which would remain valid, or why.
We recognize the difficulty of the Commission’s assignment. The Statement responds to Executive Order 14365, Ensuring a National Policy Framework for Artificial Intelligence, which directs the FTC chair, in consultation with the special adviser for AI and crypto, to explain “how state laws requiring alterations to the accurate outputs of AI models can conflict with the requirements of the FTC Act.”[84]
Section 7 of the executive order specifically requires a policy statement explaining “the circumstances under which State laws that require alterations to the truthful outputs of AI models are preempted by the Federal Trade Commission Act’s prohibition on engaging in deceptive acts or practices affecting commerce.” Meeting that directive within the prescribed 90-day period presented a substantial challenge.
Part of the difficulty comes from the breadth of the subject. As the Commission recognizes, AI is “an umbrella term covering a universe of different tools and systems.”[85] AI products, services, models, and outputs vary widely and continue to develop. Section 5 enforcement also depends heavily on the facts of each case. Providing useful guidance in an unsettled area therefore requires careful attention to particular laws, representations, products, and applications.
The First Amendment presents another difficulty. As Part I explains, many AI models, applications, and outputs may constitute expressive speech entitled to the highest degree of constitutional protection. Both federal and state regulation must comply with those protections. Even the federal legislation contemplated by Section 8 of the executive order would face constitutional limits if it restricted protected expression.
The principal difficulty arises from preemption doctrine and the structure of the FTC Act. The Statement correctly acknowledges that “the FTC Act does not expressly preempt state law.”[86] Although implied preemption remains possible, courts generally apply a presumption against preemption in fields traditionally regulated by the states. Consumer protection is one such field.[87]
A. The FTC Act Does Not Occupy the Field of Consumer Protection
The Statement relies on Schneidewind v. ANR Pipeline Co.,[88] an implied-preemption decision that did not involve the FTC Act. In Schneidewind, the Supreme Court held that the federal Natural Gas Act of 1938 (“NGA”) preempted a Michigan law because the NGA established “a comprehensive scheme of federal regulation” and occupied the relevant field to the exclusion of state law.[89]
No comparable authority holds that Section 5 occupies the fields of advertising, marketing, or consumer protection so completely that it leaves no room for state regulation. The history and structure of consumer protection law point in the opposite direction.[90]
Courts generally recognize that state consumer protection statutes, including laws modeled on the FTC Act, operate alongside federal law when their substantive and jurisdictional requirements are satisfied. Congress has allowed the FTC Act and state consumer protection laws to coexist, subject to limited exceptions.[91]
The Supreme Court reached an analogous conclusion in California v. ARC America Corp.[92] There, the Court held that federal antitrust law, including the Sherman Act, did not preempt distinct state antitrust remedies.[93] The FTC Act likewise supplies no basis for treating federal consumer protection law as exclusive.
B. The Statement Identifies No Actual Conflict with State Law
The Statement also cites Schneidewind for the rule that state law is preempted when it actually conflicts with federal law. Such a conflict may arise when a private party cannot comply with both federal and state requirements or when state law obstructs Congress’ purposes and objectives.[94]
PLIVA, Inc. v. Mensing illustrates impossibility preemption.[95] The Supreme Court concluded that federal drug-labeling requirements prevented generic-drug manufacturers from independently making the label changes required by state tort law. Because the manufacturers could not comply with both sets of requirements, federal law preempted the state-law claims.[96]
The Statement identifies no comparable conflict between the FTC Act and any state AI law. We are unaware of a state law or regulation—regardless of its policy merits—that makes compliance with both state law and Section 5 impossible.
The Statement predicts that an AI provider might “suppress accuracy and interpose other objectives, such as so-called ‘equity,’ to avoid liability under this law, but fail to disclose these ulterior objectives in order to hide the loss of accuracy they necessitate.”[97] An AI provider could conceivably respond to a state law in that manner. But the Statement does not identify an existing law that requires such conduct or a concrete representation that would deceive consumers.
Any Section 5 analysis would depend on the provider’s actual statements, omissions, practices, and disclosures. It would also require an assessment of how reasonable consumers understood those representations and whether they were material. Speculation about how a provider might respond to an unidentified state law does not establish that simultaneous compliance is impossible.
The Statement also asserts that “[a] state law that requires an AI firm to deceive its consumers obviously conflicts with Section 5’s express purpose of protecting consumers from such conduct.”[98] An appeal to Section 5’s general purpose does not resolve the preemption question.
Section 5 broadly prohibits unfair methods of competition and unfair or deceptive acts or practices. It also defines the Commission’s enforcement authority and limits that authority. Its broad terms, case-specific application, and longstanding coexistence with state competition and consumer protection laws make obstacle preemption especially difficult to establish.
Those features also help explain courts’ growing skepticism toward preemption theories based on broad assertions about statutory purposes. As Justice Neil Gorsuch cautioned in Virginia Uranium, Inc. v. Warren, “‘pre-emptive purpose,’ whether express or implied, must therefore be ‘sought in the text and structure of the statute at issue.’”[99] The Statement does not identify language or structure in the FTC Act that supports its expansive suggestion of preemption.
AI providers can violate the FTC Act, and a particular state law could conceivably conflict with a federal law or regulation that the Commission enforces. Some existing or proposed state laws may also violate the First Amendment without regard to statutory preemption.
The Statement does not identify such a law, describe the conflict, or explain how established preemption doctrine would apply. Its general assurances therefore offer little guidance to AI providers, consumers, or federal and state lawmakers.
Conclusion
The Supreme Court has cautioned that “[o]ur constitutional tradition stands against the idea that we need Oceania’s Ministry of Truth.”[100] The Commission should heed that warning. It should not assume responsibility for deciding whether individual chatbot responses are true, false, balanced, or ideologically acceptable.
AI providers exercise editorial judgment when they develop models and generate responses, and users have a First Amendment interest in receiving those responses. Competition among providers also allows consumers to compare systems, test claims of accuracy and neutrality, and switch services when a product fails to meet their expectations.
The FTC retains authority to challenge concrete, material misrepresentations in advertising and marketing. Any enforcement action should follow the Deception Statement, distinguish commercial claims from protected expression, and account for providers’ express warnings that chatbot outputs may contain errors. Disagreement with an output, standing alone, does not establish deception.
The Commission should also revise the Statement’s preemption analysis. It should identify the state laws and factual circumstances that could create an actual conflict with Section 5 rather than rely on broad appeals to statutory purpose.
The FTC should revise the AI Policy Statement to provide clear, fact-specific guidance and to confine enforcement to conduct that falls within Section 5 and complies with the First Amendment. That approach would protect consumers without making the Commission the arbiter of truth in the marketplace of ideas.
[1] Policy Statement Concerning the Suppression of Accuracy in Artificial Intelligence Systems, 91 Fed. Reg. 41,638 (July 7, 2026) [hereinafter AI Policy Statement].
[2] See, e.g., American Medical Ass’n v. FTC, 638 F.2d 443 (2d Cir. 1980); Iowa Chapter of Am. Physical Therapy Ass’n, 111 F.T.C. 199 (1988); Fed. Trade Comm’n Bureau of Econ., Staff Report: Effects of Restrictions on Advertising and Commercial Practice in the Professions: The Case of Optometry (1980), https://www.ftc.gov/sites/default/files/documents/reports/effects-restrictions-advertising-and-commercial-practice-professions-case-optometry/198009optometry.pdf; P. Ippolito & A. Mathios, Health Claims in Advertising and Labeling: A Study of the Cereal Market (FTC Bureau of Econ. Study, 1989); FTC Staff Comment Before the Massachusetts Department of Public Health Concerning Proposed Regulation of Limited Service Clinics (Oct. 1, 2007), https://www.ftc.gov/sites/default/files/documents/advocacy_documents/ftc-staff-comment-massachusetts-department-public-health-concerning-proposed-regulation-limited/v070015massclinic.pdf; FTC Staff Comments to the Sunset Advisory Commission of the State of Texas (Aug. 14, 1992) (addressing licensing restrictions on business practices imposed by the state boards of optometry, dentistry, and medicine).
[3] AI Policy Statement, supra note 1, at 41,639.
[4] See id. at 41,638 (“Excessive AI regulation would undermine American AI supremacy by deterring and suppressing the same ingenuity responsible for making American AI great.”).
[5] Id. at 41,641 n.44 (noting that “a law restricting truthful speech because it might lead another person to commit disparate impact discrimination would not survive First Amendment scrutiny,” but explaining that issue “is orthogonal to the FTC Act issues we address here”).
[6] Id. at 41,638.
[7] 15 U.S.C. § 45(a)(2); 15 U.S.C. § 44 (defining “corporation”).
[8] Cf. AI Policy Statement, supra note 1, at 41,639 n.11-12 and accompanying text.
[9] See AI Policy Statement, supra note 1, at 41,639 (citing Exec. Order No. 14365, Ensuring a National Policy Framework for Artificial Intelligence, 90 Fed. Reg. 58,499, 58,499 (Dec. 11, 2025)).
[10] Rice v. Santa Fe Elevator Corp., 331 U.S. 218, 230 (1947); see also Arizona v. United States, 567 U.S. 387 (2012); cf. Jones v. Google, LLC, 56 F.4th 735 (9th Cir. 2022) (holding that state-law claims were neither expressly nor impliedly preempted by the Children’s Online Privacy Protection Act).
[11] Geier v. American Honda Motor Co., 529 U.S. 861, 869-72 (2000).
[12] See infra Section III.B; see also Va. Uranium, Inc. v. Warren, 587 U.S. 761, 778 (2019) (plurality opinion) (quoting CSX Transp., Inc. v. Easterwood, 507 U.S. 658, 664 (1993)); Caleb Nelson, Preemption, 86 Va. L. Rev. 225, 231-32 (2000) (arguing against obstacle preemption).
[13] Much of this section draws from earlier ICLE work, including Int’l Ctr. for L. & Econ., ICLE Comments to FTC Regarding Technology Platform Censorship (May 21, 2025), https://laweconcenter.org/resources/icle-comments-to-ftc-regarding-technology-platform-censorship; Memorandum of Amicus Curiae International Center for Law & Economics in Support of Defendant’s Motion for Summary Judgment, State ex rel. Yost v. Google LLC, No. 21-CV-H-06-0274 (Ohio Ct. Com. Pl. Delaware Cnty. Jan. 30, 2024), https://laweconcenter.org/wp-content/uploads/2024/01/2024-1-30-ICLE-Amicus-Curiae-Brief.pdf; and Ben Sperry, Knowledge and Decisions in the Information Age: The Law & Economics of Regulating Misinformation on Social-Media Platforms, 59 Gonz. L. Rev. 319 (2024). For further discussion, see Brief of Amicus Curiae International Center for Law & Economics in Support of Appellee, State ex rel. Yost v. Google LLC, No. 25 CAE 08-0070 (Ohio Ct. App. 5th Dist. Dec. 19, 2025), https://laweconcenter.org/wp-content/uploads/2025/12/ICLE-Amicus-Yost-v.-Google-Court-of-Appeals.pdf; Brief of Amicus Curiae International Center for Law & Economics, Moody v. NetChoice, LLC, Nos. 22-277 & 22-555 (U.S. Dec. 4, 2023), https://laweconcenter.org/wp-content/uploads/2023/12/Intl-Ctr-for-Law-and-Econ-Amicus-12.4.231148722.12.pdf; Brief of Amicus Curiae International Center for Law & Economics, Murthy v. Missouri, No. 23-411 (U.S. Feb. 9, 2024), https://laweconcenter.org/wp-content/uploads/2024/02/Murthy-v.-Missouri-Intl-Center-for-Law-Econ.-Am.-Br.-2-9-24-pm-FINAL.pdf.
[14] See, e.g., John Stuart Mill, On Liberty ch. 2 (1859); John Milton, Areopagitica (1644).
[15] See Exec. Order No. 14365, supra note 9; see also Kevin Schaul, Are ChatGPT and Other AI Chatbots Politically Biased? We Tested Them., Wash. Post (June 29, 2026), https://www.washingtonpost.com/technology/interactive/2026/06/24/are-ai-chatbots-like-chatgpt-politically-biased-we-tested-them.
[16] See Mill, supra note 14, ch. 2.
[17] See Thomas Jefferson, First Inaugural Address (Mar. 4, 1801), https://avalon.law.yale.edu/19th_century/jefinau1.asp.
[18] See Abrams v. United States, 250 U.S. 616, 630 (1919) (Holmes, J., dissenting) (“[T]ime has upset many fighting faiths,” and “the ultimate good desired is better reached by free trade in ideas—that the best test of truth is the power of the thought to get itself accepted in the competition of the market…. That at any rate is the theory of our Constitution.”)
[19] See David Schultz, Marketplace of Ideas, Free Speech Ctr., https://firstamendment.mtsu.edu/article/marketplace-of-ideas (last updated July 9, 2024).
[20] Manhattan Cmty. Access Corp. v. Halleck, 587 U.S. 802, 804 (2019).
[21] Moody v. NetChoice, LLC, 603 U.S. 707 (2024).
[22] Id. at 717.
[23] Id. at 709-10.
[24] Id. at 737-38.
[25] Id. at 733.
[26] AI Policy Statement, supra note 1, at 41,641.
[27] Moody v. NetChoice, LLC, 603 U.S. 707, 745 (2024) (Barrett, J., concurring).
[28] Cf. id. at 716 (“To the extent that social-media platforms create expressive products, they receive the First Amendment’s protection.”).
[29] Zhang v. Baidu.com Inc., 10 F. Supp. 3d 433 (S.D.N.Y. 2014).
[30] Id. at 438.
[31] Id.
[32] See e-ventures Worldwide, LLC v. Google, Inc., No. 2:14-cv-646-FtM-PAM-CM, 2017 WL 2210029, at *4 (M.D. Fla. Feb. 8, 2017); Langdon v. Google, Inc., 474 F. Supp. 2d 622, 629-30 (D. Del. 2007).
[33] Search King, Inc. v. Google Tech., Inc., No. CIV-02-1457-M, 2003 WL 21464568 (W.D. Okla. May 27, 2003)..
[34] Id. at *4.
[35] Cf. Moody, 603 U.S. at 732 (“[T]he government cannot get its way just by asserting an interest in improving, or better balancing, the marketplace of ideas.”).
[36] Martin v. City of Struthers, 319 U.S. 141, 143 (1943).
[37] Thomas v. Collins, 323 U.S. 516, 534 (1945) (“That there was restriction upon Thomas’ right to speak and the rights of the workers to hear what he had to say, there can be no doubt.”)
[38] Martin, 319 U.S. at 145–47 (“Freedom to distribute information to every citizen wherever he desires to receive it is so clearly vital to the preservation of a free society that, putting aside reasonable police and health regulations of time and manner of distribution, it must be fully preserved.”) (praising pamphlets as a means of “espousing various causes”).
[39] Stanley v. Georgia, 394 U.S. 557, 564 (1969) (“It is now well established that the Constitution protects the right to receive information and ideas.”); Smith v. California, 361 U.S. 147, 153 (1959) (finding that the ordinance “tends to impose a severe limitation on the public’s access to constitutionally protected matter”).
[40] Va. State Bd. of Pharmacy v. Va. Citizens Consumer Council, Inc., 425 U.S. 748, 763–64 (1976) (explaining that a consumer’s interest in the free flow of commercial information “may be as keen, if not keener by far, than his interest in the day’s most urgent political debate,” and that society may also have a strong interest in such information).
[41] Brown v. Ent. Merchs. Ass’n, 564 U.S. 786, 794 (2011) (holding that minors receive substantial First Amendment protection and that the government may bar their access to protected material only in “relatively narrow and well-defined circumstances,” including when regulating violent video games).
[42] Packingham v. North Carolina, 582 U.S. 98, 108 (2017) (“[T]o foreclose access to social media altogether is to prevent the user from engaging in the legitimate exercise of First Amendment rights,” including access to “the world of ideas.”).
[43] See AI Policy Statement, supra note 1, at 41,641 n.45 (“The Commission at this time takes no position on whether the practices discussed in this statement may also be unfair under the FTC Act.”).
[44] FTC Policy Statement on Unfairness, appended to International Harvester Co., 104 F.T.C. 949, 1070 (1984) (Dec. 17, 1980), https://www.ftc.gov/legal-library/browse/ftc-policy-statement-unfairness.
[45] Cf. AI Policy Statement, supra note 1, at 41,641 (“A company could be tempted, for example, to abuse consumer trust by training a model surreptitiously to produce ideologically motivated distortions in a response to a factual question, such as to correct what the developer believes are ‘historical injustices’ in the facts.”).
[46] United States v. Alvarez, 567 U.S. 709 (2012).
[47] Id. at 717 (quoting United States v. Stevens, 559 U.S. 460, 468 (2010)).
[48] Id. at 718.
[49] See Chaplinsky v. New Hampshire, 315 U.S. 568, 571–72 (1942) (“There are certain well-defined and narrowly limited classes of speech, the prevention and punishment of which has never been thought to raise any Constitutional problem.”).
[50] See Alvarez, 567 U.S. at 718–22; see also Pittsburgh Press Co. v. Pittsburgh Commission on Human Relations, 413 U.S. 376, 389 (1973).
[51] Alvarez, 567 U.S. at 713.
[52] Id. at 715.
[53] See, e.g., Catherine Thorbecke & Clare Duffy, Google Halts AI Tool’s Ability to Produce Images of People After Backlash, CNN (Feb. 22, 2024), https://www.cnn.com/2024/02/22/tech/google-gemini-ai-image-generator.
[54] Alvarez, 567 U.S. at 727–28.
[55] See AI Policy Statement, supra note 1, at 41,640.
[56] For more on the interaction among Section 5 of the Federal Trade Commission Act, antitrust law, and content moderation, see Comments of the Program on Economics & Privacy, George Mason University Antonin Scalia Law School, Re: Request for Public Comment Regarding Technology Platform Censorship (May 21, 2025), https://masonlec.org/wp-content/uploads/2025/05/PEP-Comment-FTC-Censorship_Final-Filed-5.21.2025.pdf; Int’l Ctr. for L. & Econ., ICLE Comments to FTC Regarding Technology Platform Censorship, supra note 13; Daniel J. Gilman & Ben Sperry, Is There an Empty Set at the Intersection of Antitrust and Content Moderation?, Concurrences (Nov. 3, 2025), https://laweconcenter.org/resources/is-there-an-empty-set-at-the-intersection-of-antitrust-and-content-moderation.
[57] See AI Policy Statement, supra note 1, at 41,639 n.11 (citing White House, Fact Sheet: President Donald J. Trump Ensures a National Policy Framework for Artificial Intelligence (Dec. 11, 2025), https://www.whitehouse.gov/fact-sheets/2025/12/fact-sheet-president-donald-j-trump-ensures-a-national-policy-framework-for-artificial-intelligence).
[58] Id. at 41,639 n.12.
[59] See, e.g., Bolger v. Youngs Drug Prods. Corp., 463 U.S. 60 (1983); Cent. Hudson Gas & Elec. Corp. v. Pub. Serv. Comm’n, 447 U.S. 557 (1980).
[60] FTC Policy Statement on Deception, appended to Cliffdale Assocs., Inc., 103 F.T.C. 110, 174 (1984) (Oct. 14, 1983), https://www.ftc.gov/system/files/documents/public_statements/410531/831014deceptionstmt.pdf.
[61] See, e.g., Donaldson v. Read Magazine, Inc., 333 U.S. 178, 190 (1948) (recognizing that the government’s power “to protect people against fraud” has “always been recognized in this country and is firmly established”).
[62] At common law, fraud generally requires (1) a material misrepresentation or omission where there is a duty to disclose, (2) intent to induce reliance, (3) knowledge of the statement’s falsity or misleading nature, (4) justifiable reliance, and (5) resulting injury. See, e.g., Mandarin Trading Ltd. v. Wildenstein, 919 N.Y.S.2d 465, 469 (N.Y. 2011); Kostryckyj v. Pentron Lab. Techs., LLC, 52 A.3d 333, 338-39 (Pa. Super. Ct. 2012); Masingill v. EMC Corp., 870 N.E.2d 81, 88 (Mass. 2007). Likewise, restrictions on deceptive or misleading commercial speech have long been held consistent with the First Amendment. See Va. State Bd. of Pharmacy v. Va. Citizens Consumer Council, Inc., 425 U.S. 748, 771-72 (1976) (“Obviously, much commercial speech is not provably false, or even wholly false, but only deceptive or misleading. We foresee no obstacle to a State’s dealing effectively with this problem. The First Amendment, as we construe it today does not prohibit the State from insuring that the stream of commercial information flow cleanly as well as freely.”).
[63] See Alvarez, 567 U.S. at 719 (“Even when considering some instances of defamation and fraud . . . the Court has been careful to instruct that falsity alone may not suffice to bring the speech outside the First Amendment. The statement must be a knowing or reckless falsehood.”). Thus, even in areas such as defamation and fraud, the First Amendment limits liability for false speech absent the requisite level of fault.
[64] Illinois ex rel. Madigan v. Telemarketing Assocs., Inc., 538 U.S. 600, 617 (2003).
[65] See, e.g., Schaumburg v. Citizens for a Better Env’t, 444 U.S. 620 (1980); Sec’y of State of Md. v. Joseph H. Munson Co., 467 U.S. 947 (1984); Riley v. Nat’l Fed’n of the Blind of N.C., Inc., 487 U.S. 781 (1988).
[66] Madigan, 538 U.S. at 620.
[67] FTC Policy Statement on Deception, supra note 60.
[68] Id.
[69] Id.
[70] Id.
[71] See supra note 2 and accompanying text.
[72] See AI Policy Statement, supra note 1, at 41,640 nn.35–36 and accompanying text.
[73] OpenAI, Why Language Models Hallucinate (Sept. 5, 2025), https://openai.com/index/why-language-models-hallucinate.
[74] OpenAI, Terms of Use (Jan. 1, 2026), https://openai.com/policies/row-terms-of-use.
[75] Anthropic, Consumer Terms of Service (Oct. 8, 2025), https://www.anthropic.com/legal/consumer-terms.
[76] DeepSeek, Terms of Use (Mar. 27, 2026), https://cdn.deepseek.com/policies/en-US/deepseek-terms-of-use.html.
[77] Gab AI, GAB AI Service Terms of Service (Apr. 8, 2025), https://gab.ai/terms-of-service.
[78] xAI, Terms of Service—Consumer (June 26, 2026), https://x.ai/legal/terms-of-service.
[79] Google, Generative AI Additional Terms of Service (Aug. 9, 2023), https://policies.google.com/terms/generative-ai.
[80] AI Policy Statement, supra note 1, at 41,640 n.35.
[81] Cf. FTC Policy Statement on Deception, supra note 60 (“The Commission generally will not pursue cases involving obviously exaggerated or puffing representations, i.e., those that ordinary consumers do not take seriously.”).
[82] AI Policy Statement, supra note 1, at 41,641.
[83] See, e.g., California v. ARC America Corp., 490 U.S. 93, 101 (1989); Rice, 331 U.S. at 230; Arizona, 567 U.S. 387. The Supreme Court has refined its preemption doctrine—grounded principally in the Supremacy Clause, U.S. Const. art. VI, cl. 2, and, to some extent, the Necessary and Proper Clause, id. art. I, § 8, cl. 18—since McCulloch v. Maryland, 17 U.S. (4 Wheat.) 316 (1819).
[84] See AI Policy Statement, supra note 1, at 41,639 (citing Exec. Order No. 14365, Ensuring a National Policy Framework for Artificial Intelligence, 90 Fed. Reg. 58,499, 58,499 (Dec. 11, 2025)).
[85] Id. at 41,638.
[86] Id. at 41,641.
[87] ARC America, 490 U.S. at 101 (citing Hillsborough County v. Automated Medical Laboratories, Inc., 471 U.S. 707, 716 (1985)); see also Wyeth v. Levine, 555 U.S. 555 (2009); Rice, 331 U.S. at 230.
[88] Schneidewind v. ANR Pipeline Co., 485 U.S. 293 (1988).
[89] Id. at 300; see also, e.g., Rice, 331 U.S. at 230; Arizona v. United States, 567 U.S. 387 (2012).
[90] Rice, 331 U.S. at 230; cf. Jones v. Google, LLC, 56 F.4th 735 (9th Cir. 2022) (holding that the Children’s Online Privacy Protection Act neither expressly nor impliedly preempted the state-law claims).
[91] Walsh v. Ford Motor Co., 807 F.2d 1000, 1013–14 (D.C. Cir. 1986) (“Congress intended the application of state law, except as expressly modified by Magnuson-Moss, in section 110(d) breach of warranty actions.”); Deadwyler v. Volkswagen of America, Inc., 748 F. Supp. 1146, 1150 (W.D.N.C. 1990) (explaining that, in enacting the Magnuson-Moss amendments to the Federal Trade Commission Act, “Congress did not intend to replace state law except in those instances[] where it explicitly said so”).
[92] California v. ARC America Corp., 490 U.S. 93 (1989).
[93] See id. at 101.
[94] Schneidewind, 485 U.S. at 300; see also Florida Lime & Avocado Growers, Inc. v. Paul, 373 U.S. 132, 142–43 (1963); Freightliner Corp. v. Myrick, 514 U.S. 280, 287 (1995).
[95] PLIVA, Inc. v. Mensing, 564 U.S. 604 (2011).
[96] Id. at 618 (quoting Freightliner, 514 U.S. at 287).
[97] AI Policy Statement, supra note 1, at 41,641.
[98] Id.
[99] Va. Uranium, 587 U.S. at 778 (plurality opinion) (quoting CSX Transportation, 507 U.S. at 664); see also Nelson, supra note 12, at 231–32 (arguing against obstacle preemption).
[100] Alvarez, 567 U.S. at 728.