The FTC voted 2-0 on July 1 to propose a policy statement declaring that undisclosed steering of AI system outputs is a deceptive practice under Section 5 of the FTC Act, and that state AI bias laws are “impliedly preempted to the extent” they conflict with that federal deception regime. Published July 7 in the Federal Register as 91 FR 41638 (File No. P264200), the notice singles out Colorado’s revised AI Act, S.B. 26-189 § 6-1-1707, by name.

The doctrinal move is the news. The Commission argues consumers reasonably expect “truthful and accurate outputs,” and that steering models toward any other objective without clear and conspicuous disclosure constitutes a material misrepresentation. Per the Stanford Law analysis, motive is irrelevant: good-faith compliance with a state anti-discrimination statute doesn’t cure the deception. Colorado’s law, the FTC notes, exposes developers to liability for discriminatory outcomes downstream of customer use, exactly the kind of mandate that would push a company to steer outputs in the first place.

The trap is elegant. Comply with Colorado and disclose the steering, or comply with Colorado silently and face a federal Section 5 action. The burden of rebutting consumer expectation shifts to the company.

Chairman Andrew N. Ferguson framed the statement around ending “the subversion of AI systems for ideological ends,” language that ports directly from Executive Order 14365, signed December 11, 2025, on a national AI policy framework. That’s the actual mechanism: an executive order becomes an FTC interpretive doctrine becomes, once finalized, a preemption argument the Commission can litigate against state enforcement actions.

Comments close July 31 at Docket FTC-2026-0859. The venue for the fight after that’s federal court, on ground the agency has spent the summer surveying.

Sources

Sources