FTC Proposes AI Accuracy Policy That Could Preempt State Laws Requiring Output Changes
The Federal Trade Commission has issued a proposed policy statement declaring that AI companies distorting their models' outputs to comply with state regulations could be committing deceptive practices under federal law. The statement specifically singles out Colorado's AI Act and invites public comment through July 31.
The U.S. Federal Trade Commission has entered what may become one of the defining legal battles of the AI era: the fight over who gets to decide what an AI model says. On July 1, the FTC issued a proposed policy statement asserting that AI companies which alter their models’ outputs to satisfy state laws — rather than to improve accuracy — could be engaging in deceptive practices prohibited by federal law.
The statement is not merely regulatory guidance. It is a shot across the bow of state-level AI legislation, framing the patchwork of state AI laws as a potential threat to the integrity of the AI systems American consumers and businesses depend on.
The Core Argument
The FTC’s proposed policy statement, officially titled “Policy Statement Concerning the Suppression of Accuracy in Artificial Intelligence Systems,” rests on Section 5 of the FTC Act — the century-old prohibition against “unfair or deceptive acts or practices” in commerce.
The agency’s reasoning proceeds in two steps. First, when an AI company represents to users that its system will provide accurate, objective, or expert-quality outputs, those representations create an implicit contract. Consumers who use an AI legal assistant, a medical information service, or a coding tool reasonably expect the outputs to reflect the model’s genuine best knowledge — not the results of political or commercial filtering applied after the fact.
Second, the FTC argues that state laws which require AI companies to alter “truthful outputs” — for example, by mandating that outputs be adjusted to achieve demographic parity — create pressure on companies to engage in exactly this kind of undisclosed output manipulation. An AI company that quietly modifies its model’s outputs to comply with a state mandate, without disclosing that modification to users, is in the FTC’s view deceiving consumers even if it is technically complying with state law.
The statement’s conclusion: such conduct may constitute a deceptive practice under Section 5, and because that federal prohibition exists, state laws that conflict with it are “impliedly preempted.”
Colorado in the Crosshairs
The statement directly names Colorado’s Artificial Intelligence Act as an example of state legislation that could trigger this analysis. Colorado’s law, one of the more expansive state AI statutes enacted in recent years, requires AI developers and deployers to take reasonable care to protect consumers from algorithmic discrimination in high-stakes decisions including lending, hiring, and healthcare.
Critics of the FTC’s position argue that Colorado’s law is about fairness, not accuracy — that requiring an AI system to not discriminate by race in lending decisions is not the same as requiring it to produce inaccurate outputs. Supporters of the FTC approach counter that compliance with disparate impact requirements often requires actively modifying model outputs in ways that can reduce accuracy for individual users, and that those modifications are almost never disclosed.
The specific tension is real and not easily resolved. A model trained on historical lending data may encode historical discriminatory patterns; correcting for those patterns to achieve demographic parity in outcomes requires intervention in the output distribution. Whether that intervention suppresses “accuracy” or corrects for biased training data depends heavily on one’s definition of what an AI system is supposed to do — predict the past, or serve users equitably.
An Executive Order at the Root
The FTC did not arrive at this policy on its own initiative. The statement was issued pursuant to Executive Order 14365, signed by President Trump on December 11, 2025, which directed the FTC to clarify how Section 5 of the FTC Act applies to AI models and specifically to address “how state laws requiring alterations to the accurate outputs of AI models can conflict with federal law.”
That framing — “accurate outputs” — embedded a significant assumption in the FTC’s mandate from the start. By defining the problem as state laws interfering with accuracy, the executive order predisposed the FTC toward a preemption analysis. The Commission voted 2-0 in favor of issuing the proposed statement, indicating unanimous support among the agency’s current membership.
The broader context is a Trump administration that has been consistently skeptical of state-level AI regulation, particularly as it applies to large language models. The White House has pursued voluntary frameworks for frontier AI development — agreeing with major AI companies on pre-release access and safety evaluations — while pushing back against mandatory state requirements it views as overreach.
The Preemption Doctrine and Its Limits
Federal preemption of state law is a well-established constitutional doctrine, but its application to AI is novel and contested. For preemption to work, a federal law or regulation must affirmatively occupy a field, or a specific state law must conflict with federal requirements. The FTC’s proposed statement asserts “implied preemption” — the idea that a state law can be preempted not because Congress explicitly said so, but because the state requirement conflicts with a federal regulatory scheme.
The statement is not a rule. It has not gone through notice-and-comment rulemaking, it has not been enacted by Congress, and it would not automatically invalidate state laws. Its legal weight would depend on whether courts find it persuasive as an interpretation of Section 5 and existing implied preemption doctrine.
Legal scholars are split. Some argue that the FTC’s proposed statement is a plausible extension of consumer protection law to AI. Others contend it is an aggressive overreach that uses an anti-deception provision to nullify democratically enacted state civil rights protections — and that a Commission with only two voting members approving such a sweeping position lacks the democratic legitimacy the move requires.
Comment Period and What Comes Next
The FTC is accepting public comments through July 31, 2026, via Regulations.gov (Docket FTC-2026-0859). The breadth of the proposed statement means stakeholders ranging from state attorneys general to civil rights organizations to AI developers have strong incentives to weigh in.
Whatever the eventual legal resolution, the FTC’s move reshapes the debate. Before July 1, the dominant narrative around U.S. AI policy was of a fragmented landscape of state laws slowly accumulating toward something that resembled federal regulation without a federal bill. That narrative assumed states would, over time, fill the vacuum left by Congressional inaction.
The FTC’s statement challenges that assumption directly. If the agency’s preemption theory holds — either before courts or as a political deterrent — states may find their AI legislation blocked not by a new federal AI law but by the application of a 100-year-old consumer protection statute to a technology its drafters could never have imagined.
The battle over who governs AI in America is no longer only happening in Congress. The FTC just made that clear.
Public comments on the FTC’s proposed AI accuracy policy statement are due July 31, 2026. The full text is available in the Federal Register (Document 2026-13628).