AINEWS
FTC's new AI target: accuracy trade-offs, not hallucinations
A draft policy treats undisclosed steering away from correct answers as a federal consumer deception case, with Colorado's AI law named in the crossfire.
A draft policy treats undisclosed steering away from correct answers as a federal consumer deception case, with Colorado's AI law named in the crossfire.
When AI companies started selling "best answer" accuracy as their core promise, they opened a trap door. The FTC has now stepped through it. On July 1, 2026, the agency released a proposed policy statement treating deliberate steering of an AI's output away from a correct answer, without telling users, as a potential deception case under Section 5 of the FTC Act, the agency's main consumer-protection authority.
The agency has a named target in mind: Colorado's Artificial Intelligence Act. The FTC argues that compliance with that state rule is not a defense if the result is a federal deception claim. If a company quietly chooses a non-accuracy objective over the "best answer" it could give, and tells users it is aiming for the right one, the agency now considers that a Section 5 problem.
That framing is a deliberate pivot from how AI regulation has been discussed so far. The proposed statement does not chase hallucinations, the kind of confident-but-wrong output that comes from a model's training limits, compute constraints, or data gaps. It targets what the FTC calls "suppression of accuracy," the choice not to give the best available answer in order to serve some other goal. The agency's policy document lists examples: steering outputs to advance a political or ideological objective, to align with a developer's view of "historical injustices," or to comply with a state rule like Colorado's that the FTC says pressures companies into altering outputs to avoid disparate-impact liability, the legal exposure that punishes outcomes falling harder on a protected group even without intent to discriminate.
The shift matters because it changes what counts as a deceptive trade practice in AI. Until now, the FTC's AI work has focused on standard consumer-protection problems: chatbots impersonating real people, AI-generated reviews, or vendors making unsupported capability claims. "Suppression of accuracy" recasts the test. The deception is not that the AI is broken. It is that the company told the user it was aiming for correctness while privately choosing something else.
The mechanism is the AI vendor's own marketing. The FTC's position assumes that phrases like "best, most accurate answer," which major model developers use in launch pitches and product pages, are consumer-facing representations. Once a company has told users it is optimizing for accuracy, the agency argues, optimizing for anything else becomes a Section 5 issue. The standard is disclosure: if a company wants to weight safety, politics, or state-law compliance over the most accurate answer it can give, it has to tell users that, in plain terms, the model is not necessarily aiming at the right answer.
That puts Colorado at the center of a federal preemption fight. Colorado's Artificial Intelligence Act requires developers and deployers of "high-risk" AI systems to take reasonable steps to avoid algorithmic discrimination. The FTC's draft says that pressure to alter outputs in response to such a state rule does not let a vendor off the hook federally. The proposed policy statement names the conflict directly, arguing that "compliance with state law does not excuse a deceptive act or practice." That is a signal that the FTC is willing to litigate the boundary between state AI rules and federal consumer-protection law, even if the question of which authority wins out has to be settled in court.
There is a real weakness in the FTC's frame, and it is worth naming. The agency's position assumes the vendor's "best answer" marketing is itself accurate, that the model could in fact give a more correct answer, and that the developer knows it. Smart AI labs and skeptical readers will contest both halves of that assumption. Modern frontier models do not have a single objectively "correct" output for most user queries, and the line between a safety-tuning adjustment and an "accuracy suppression" decision is not always clean. The FTC's framing treats that ambiguity as a problem the company has to resolve, on the record, in its product disclosures.
The political durability of the position is also live. A 2026 Supreme Court ruling allows the President to fire FTC commissioners at will, which changes the political exposure of any policy posture the agency takes. A single administration can now reshape the FTC's enforcement theory, including this one, without leaving fingerprints on the statute book.
The proposal is open for public comment through July 31, 2026. After that, the FTC can revise, finalize, or quietly drop the statement. None of those outcomes bind a court the way a formal rule would. Even so, the document carries weight today. It tells AI vendors that the marketing copy on their product pages is now evidence in a federal deception case, and it tells state legislatures like Colorado's that the FTC considers itself the final word on whether a model is telling the truth.