FTC’s Proposed Policy Statement On AI Accuracy And Preemption Of State AI Laws

Earlier this month, the Federal Trade Commission (FTC) issued a proposed policy statement addressing what it calls the “suppression of accuracy” in artificial intelligence (AI) systems. Under the policy described in the proposed statement, the FTC would use Section 5 of the FTC Act, which prohibits “unfair or deceptive acts or practices,” to pursue AI companies that allegedly distort their systems’ outputs toward undisclosed objectives, including ideological agendas, rather than the objectives users request or reasonably expect. The FTC is seeking public comment on the statement, with comments due July 31, 2026.

The FTC’s proposal executes a directive in President Trump’s Executive Order 14365, which required the FTC to issue such a statement explaining how Section 5 would preempt state laws that “require alterations of the truthful outputs of AI models.” The executive order singled out Colorado’s since-revised Artificial Intelligence Act as an example of a state mandate that could “force AI models to produce false results.” The same accuracy-versus-ideology framing also appeared a few months earlier in Executive Order 14319, which barred federal agencies from procuring AI models that “sacrifice truthfulness and accuracy to ideological agendas.”

Assuming it is adopted, the proposal will create enforcement risk not just for developers of AI models, but also for downstream entities that create systems based on those models and for customers that fine-tune and deploy them. Any perceived bias, including ideological bias, of an AI system could form the predicate for an FTC inquiry.

The statement expressly addresses state AI laws (and in particular the revised Colorado statute) and asserts that undisclosed distortion of AI systems to accommodate state law requirements will not immunize them from federal scrutiny. Indeed, the statement suggests that those state laws may be preempted where they require AI companies to engage in conduct that would violate Section 5.

FTC’s “Distortion” Theory

The FTC’s proposed statement begins from the premise that AI companies have represented, both “explicitly and implicitly,” that their systems are designed to produce “output that faithfully and accurately achieves users’ stated objectives” and the objectives that “users expect,” subject to technological and resource constraints. In the FTC’s view, consumers therefore have a reasonable expectation that AI systems will aim to provide truthful and accurate outputs, not outputs “distorted by undisclosed ideological objectives.”

In the FTC’s view, an AI company thus may violate Section 5 if it steers a system’s output toward objectives that users did not request and would not reasonably expect. As an example, the FTC suggests that “[a] company could be tempted … 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.”

The FTC acknowledges that a company may avoid liability by making truthful, non-misleading representations about its model’s aims. But the proposed policy statement suggests that those disclosures would need to be sufficiently prominent and persistent to have an actual effect on the expectations of customers or others (like employees) on whom a company uses the AI system. A disclosure “buried in terms of service,” or a one-time disclosure later obscured by product marketing, may not suffice if the overall net impression remains that the system is designed to give the best or most accurate answer possible.

Significantly, the FTC distinguishes undisclosed distortions from ordinary AI “hallucinations.” In the commission’s view, hallucinations caused by technological limitations do not themselves raise Section 5 concerns, although misrepresentations about the likelihood or frequency of hallucinations may do so. The FTC’s focus is instead on intentional, design-based steering that causes a system to prioritize an undisclosed objective over the user’s own.

State AI Laws and the Colorado Example

As noted, the proposed policy statement devotes particular attention to state AI laws. Consistent with Executive Order 14365, it states that attempted compliance with state law does not excuse an undisclosed distortion of AI outputs and that, although the FTC Act does not expressly preempt state law, state laws may be impliedly preempted where they require conduct that conflicts with Section 5.

Colorado’s recently revised Artificial Intelligence Act is the FTC’s central example of a potentially preempted state statute. The proposed statement notes that the original version of the Colorado Artificial Intelligence Act imposed broad duties on AI companies to avoid outputs that could produce a disparate impact — including where the disparate impact arose from a customer’s own foreseeable use of the output. According to the FTC, Colorado’s revised law raises similar concerns because it potentially allows AI companies to be held liable for discriminatory outcomes caused by their customers’ use of AI products.

The FTC’s underlying concern appears to be that a developer facing this kind of state-law exposure may alter or constrain model outputs to reduce disparate-impact risk — for instance, in consequential decision-making contexts like employment, housing, lending, or health care. The commission suggests that if those output modifications cause a system to depart from users’ expected objectives without adequate disclosure, the company could be deceiving consumers, and there is no defense to deception in violation of Section 5 based on intent to comply with state law.

Some industry participants may welcome this preemption-oriented theory, since it could supply a federal-law argument against a growing patchwork of state AI requirements. At the same time — particularly until the FTC’s preemption theory is tested in court — the theory may create significant uncertainty for companies trying to reconcile federal and state compliance obligations, along with any internal responsible-AI commitments they may have in place. Companies creating models or systems used (or with effects) outside of the United States also have to consider the requirements of those jurisdictions.

Legal Authority and Potential Future Questions

Any enforcement action under the proposal, if adopted, would likely reflect the FTC’s longstanding three-part test for finding unlawful deception under Section 5: whether there is a representation, omission, or practice likely to mislead a reasonable consumer in a material way. Indeed, in the proposed statement, the FTC cites recent Section 5 enforcement actions alleging misleading claims about the accuracy, efficacy, or capabilities of AI products.

The central legal and practical questions will likely concern how far those principles can extend in this setting. Industry will likely argue that the proposal does not clearly define key terms — like “accuracy,” “truthfulness,” “objectivity,” “ideological objectives,” and “best output.” They may note that these concepts are especially hard to apply to general-purpose AI systems that must balance accuracy and helpfulness, safety and security, and user-specific context all at once — even more so as generative AI systems inherently may provide different outputs in response to the same prompt. Downstream entities that use these models may also ask how the policy applies to them. For example, an entity that fine-tunes a leading model for use by its customers arguably could face enforcement risk, even if the company cannot control many features of the underlying model.

The FTC’s proposed policy may also raise First Amendment questions if applied through enforcement actions. Industry participants are likely to argue that model design and output-selection decisions involve editorial judgments or speech-related choices, and that enforcing Section 5 against “ideological” steering risks chilling protected expression or penalizing particular views about safety or fairness. Industry participants may go further and frame the proposal’s imprecision as a constitutional problem in its own right. They could claim that standardless, capacious terms like “objectivity” are unconstitutionally vague because they give enforcers effectively unbridled discretion over what counts as an acceptable output. They may also argue that requiring companies to characterize their own model objectives in a government-prescribed way raises compelled-speech concerns. The FTC would likely respond that the proposed policy targets misleading commercial statements to which First Amendment protections do not attach, not protected editorial judgment as such.

The preemption discussion will be another flashpoint. States and advocates for more stringent state AI oversight will likely contest the FTC’s preemption theory. Even accepting the proposition that state laws that “require an AI firm to deceive consumers” are preempted, state laws requiring transparency, risk management, or anti-discrimination safeguards arguably can coexist with truthful disclosures to users. On the other hand, even if a state law does not actually conflict with federal law, it (or certain aspects of it) may be preempted where they pose an “obstacle” to the federal law’s purpose. The FTC might therefore argue that state AI laws like the Colorado law make undisclosed “distortion” (under the FTC’s broad conception of distortion) more likely, by pressuring AI developers to alter outputs in ways that are difficult to disclose fully or administer consistently.

Finally, the FTC’s proposed disclosure pathway may raise its own implementation questions. The FTC’s view that additional disclosures may be necessary rests on a premise about what leading AI systems have, “explicitly and implicitly,” represented to users about the goals of their systems and how those representations align with user expectations. But that premise is contestable, and many platforms may question whether Section 5 could plausibly require disclosures beyond those they have already made. The proposed statement also suggests that companies can shape consumer expectations through adequate disclosures. But it is not clear if it would suffice for example, for an AI platform to state that its systems’ outputs are designed to comply with applicable legal requirements, including anti-discrimination requirements. Developers and enterprise customers may seek further guidance on what must be disclosed, where disclosures must appear, how often they must be repeated at the point of use, how to explain technical model behavior without overwhelming users, and how disclosure obligations should account for downstream customization by customers. Although the proposal is directed primarily at companies that market AI systems, this last question makes it relevant to downstream customers as well.