The FTC Moves to Police How AI Models Are Steered

The Federal Trade Commission has floated a policy statement that would treat a hidden thumb on the scale of an AI model as a form of consumer deception. Published on July 1 and open for public comment until July 31, the proposed statement argues that companies which quietly steer their models toward undisclosed goals, rather than the answers a user asked for, may be committing a deceptive act under Section 5 of the FTC Act.

The reasoning runs like this. AI companies tell the public, sometimes outright and sometimes by implication, that their systems aim to give the most accurate and faithful answer they can. The FTC says consumers accept AI outputs without checking them more than 90 percent of the time, so people reasonably expect the machine is not secretly working toward some other objective. Break that expectation without telling anyone, the argument goes, and you have crossed into deception.

A safe harbor, and a target

The statement leaves a clear way out. A company can avoid liability by disclosing, plainly and up front, that its system is built to prioritize certain objectives over what a user might otherwise expect. The problem it is trying to reach is not steering as such, but steering that is concealed.

The politically loaded part is where the FTC points next. It singles out Colorado's Artificial Intelligence Act as an example of a state law that could push companies to suppress accurate outputs in order to dodge disparate-impact liability. On that basis the statement concludes such a law is impliedly preempted, at least where it collides with the federal scheme under Section 5. The move traces back to a December executive order that told the agency to address state laws requiring changes to the truthful outputs of AI models.

Why it matters

Two things are happening at once here, and it is worth keeping them apart. One is a genuine consumer-protection idea, that people deserve to know when a model has been tuned to favor some answers over others. The other is a federalism fight, with the commission using a policy statement to lean on state AI laws it views as overreaching. Legal analysts at firms including Reed Smith and Spencer Fane have flagged that a policy statement is not binding law, and that the preemption claim in particular would have to survive a court before it means much.

It also sits inside a busy season for AI rulemaking. Illinois just passed the strongest state AI safety law yet, and Washington is still assembling its own frontier-AI framework. The FTC's statement is one more front in the same underlying contest, over who gets to set the rules for these systems and how much they can shape what the models say. For now it is a proposal seeking comment, not a settled rule, and the thirty-day window will show how hard industry and the states push back.