Comments on the FTC’s proposed policy statement on “Suppression of Accuracy in Artificial Intelligence Systems” close at 11:59 p.m. tonight, ending a 30-day window that the Commission opened on a 2-0 vote and closing the first public phase of an ambitious attempt to rewire Section 5 deception doctrine into a federal preemption tool against state AI laws.

The statement, published July 7 in the Federal Register as FR Doc. 2026-13628, applies the agency’s three-part deception test to undisclosed output steering. If a developer tunes a model’s answers, whether to satisfy an internal viewpoint policy or to comply with a state statute, and doesn’t conspicuously disclose it, the FTC’s theory is that consumers, who accept AI outputs without fact-checking more than 90% of the time, are being materially misled. Hallucinations traceable to technical limits are carved out. Conspicuous disclosure is offered as an affirmative defense.

The document singles out Colorado’s Artificial Intelligence Act as the paradigmatic offender, arguing that its disparate-impact framework pressures developers to alter accurate outputs. It then reaches for the harder move: state laws that “coerce” such changes are, in the FTC’s reading, “impliedly preempted to the extent [they] conflict with a federal regulatory scheme.”

That language doesn’t appear by accident. Executive Order 14365, signed by President Trump on December 11, 2025, instructed the agency to confront state laws requiring alteration of “truthful outputs of AI models,” and the policy statement is the operational deliverable.

The critical ambiguity, flagged by TechTimes, is that the framing draws no line between ideological tuning and safety tuning. RLHF content-moderation layers sit inside the same net as viewpoint edits. Roughly 40 comments were docketed as of the Federal Register’s tally; the real test comes when a Colorado-law challenge finally puts the theory in front of a judge.

Sources

Sources