The Hidden Agenda
A proposed FTC policy statement argues that AI companies secretly steering model outputs toward undisclosed objectives are deceiving their users under federal law.
Every major AI company claims that its models try to give users the best possible answer. The Federal Trade Commission has now asked a pointed follow-up: what if they don’t, and what if users have no way to know?
The assumption of impartiality
On July 1, the Federal Trade Commission proposed a policy statement titled “Suppression of Accuracy in Artificial Intelligence Systems.” The document builds its case on a straightforward observation: AI companies have spent years telling consumers that their products aim to produce the best possible output. OpenAI markets ChatGPT as a problem-solving tool. Anthropic describes Claude as a “thinking partner” that tackles challenges with answers “grounded in evidence.” xAI calls Grok a “truth-seeking AI companion.” These marketing claims, combined with the inherent nature of the product category itself, create what the FTC calls a reasonable consumer expectation that AI systems pursue accuracy as a primary objective.
The commission cites Anthropic’s own research to quantify the reliance: consumers accept AI outputs without independent fact-checking over ninety percent of the time. Under the FTC’s three-part deception test, that trust satisfies the materiality requirement, because consumers are making real decisions on the basis of representations about accuracy.
The statement draws a line between two kinds of inaccuracy. Hallucinations produced by technological limitations fall outside its scope. The deception theory targets design decisions that subordinate accuracy to undisclosed objectives. When a model produces a worse answer because its developers trained it to prioritize something other than the user’s stated goal, and the user has no way to detect that tradeoff, the commission treats that gap between promise and performance as a Section 5 violation.
The motive is irrelevant
The statement identifies four categories of motive that could drive undisclosed output steering: a company’s own ideological convictions, compliance with state anti-discrimination laws, capitulation to public pressure over politically inflammatory outputs, and internal employees’ political agendas. Under longstanding Section 5 precedent, all four receive identical treatment. A company’s reasons for deceiving consumers have never mattered to the deception analysis, and the FTC sees no reason to create an exception for AI.
Colorado’s revised Artificial Intelligence Act, enacted May 14, receives the most sustained attention. The law holds AI companies liable for discriminatory outcomes caused by their customers’ use of their products, which could incentivize developers to adjust outputs preemptively to reduce disparate-impact exposure. The FTC frames this as accuracy suppression in service of “so-called ‘equity,’” in the statement’s language, and argues that complying with a state law does not excuse deception under federal law.
Applying motive-indifference to model training extends Section 5 into new territory. Traditional deception enforcement targets marketing claims and product labels, contexts in which a company makes a specific statement that can be evaluated as true or false. The FTC’s proposed framework treats the model’s behavior itself as the representation. If a model consistently steers outputs toward undisclosed objectives, that behavioral pattern constitutes the deceptive act, regardless of whether any human at the company ever made a false verbal claim.
Honesty is the best policy
The statement offers one compliance pathway: disclosure. A company can steer its model’s outputs toward objectives other than accuracy, provided that it tells users clearly, conspicuously, and persistently that this steering occurs. The FTC specifies that a disclaimer buried in terms of service would be inadequate, that a one-time notification would likely fall short, and that the further a model’s actual behavior departs from consumer expectations, the more prominent the disclosure must become.