On July 1, the FTC released a proposed policy statement addressing what it calls the "suppression of accuracy" in AI systems, and opened a comment period that runs through July 31, 2026. The statement takes the position that steering an AI model's output away from a correct answer, without telling users, can amount to deception under Section 5 of the FTC Act.

The Core Argument

The statement's starting point is that AI companies have spent years marketing their products, explicitly and implicitly, as tools designed to give the best, most reliable answer possible within technical and resource constraints. The statement cites language pulled directly from AI companies' own marketing as evidence of that representation. Under the FTC's standard deception test, a representation is actionable if it is likely to mislead a reasonable consumer in a way that affects their decisions. Therefore, if a company represents that it is giving consumers its best, most accurate answer, but manipulates the output behind the scenes to favor some other objective, the FTC will treat that as a Section 5 violation.

The statement is careful to distinguish ordinary hallucinations. Wrong output caused by a model's technical or resource limitations is not the target of this policy statement and does not, on its own, raise a Section 5 problem. The target is a deliberate design decision to prioritize some other objective over accuracy, without telling users that tradeoff exists. The statement's example is a company training a model to correct what its developers consider "historical injustices" in a factual answer. That is presented as a case where a company has decided, on its own initiative, to override an otherwise correct output and substitute a different one, while continuing to market the system as aiming for the best possible answer. 

The Colorado Angle

A significant portion of the statement is aimed at state AI laws, and Colorado's Artificial Intelligence Act gets singled out by name. The statement describes Colorado's law as one that could pressure companies into altering outputs to avoid disparate impact liability, potentially requiring a company to suppress a correct result in favor of one designed to avoid a disparate outcome for protected groups. Compliance with that kind of state mandate, the statement argues, is not a defense to a federal deception claim. Where the two conflict, the statement says Section 5 controls, leaving companies that alter their outputs to satisfy Colorado's law still exposed to FTC scrutiny.

What Counts as an Adequate Disclosure

The statement also attempts to explain how a company could avoid liability, and the standard it describes is fairly demanding. A company is allowed to prioritize objectives other than pure correctness, but only if it makes that clear and conspicuous to users. The statement specifically rules out burying a disclaimer in terms of service, and says a one-time disclosure that is later tucked away in fine print is unlikely to be sufficient. The agency makes clear that a disclaimer is not adequate unless it is prominent and unambiguous enough to actually change the impression a consumer would otherwise take away, and that a prominent misrepresentation is not cured by a smaller, later disclosure. The statement also says the degree of prominence required scales with how far the disclosure departs from what a user would otherwise expect, such that the more a company's chosen approach cuts against ordinary assumptions about how an AI tool operates, the more persistent and visible the disclosure must be.

Where This Fits

This is a proposed policy statement, not a rule, and it does not by itself create new legal obligations. It signals how the Commission intends to apply an existing framework, the same deception test that has governed FTC enforcement for decades, to a new category of product. It also traces directly back to Executive Order 14365 (December 2025), which instructed the Commission to address how state AI laws interact with Section 5, so this statement is executing a specific directive.

It's also worth noting the parallel to the earlier fight over alleged suppression of conservative viewpoints on social media platforms. There, too, the claim was that a group of tech companies was quietly favoring certain viewpoints, and the fix offered in some state and federal proposals was disclosure. In other words, a platform could set its own moderation rules, as long as it published them and applied them consistently. This statement follows that same structure, just applied to model outputs instead of moderation decisions.

It's also a good example of the dynamic Jeff Greenbaum flagged in his post on the Supreme Court's FTC removal decision. This statement passed on a 2-0 vote by the two Republican Commissioners at the direct instruction of an executive order. Commissioners who can be removed at will are likely to track the President's priorities more closely, which could result in less continuity that has historically persisted in terms of how laws and rules are applied.

Comments are open through July 31. If you have a view on how the line between addressing potential bias and producing accurate outputs should be drawn, or on what a compliant disclosure would actually need to look like, this is the moment to put it on the record.