If you have ever asked a chatbot about a contested topic and felt the answer was leaning on something other than the facts, the FTC has a new theory for you: what you sensed is real, and the people who built the model may be on the hook for it. On July 1, the agency proposed a policy statement that treats undisclosed "output steering" as consumer deception under Section 5 of the FTC Act. It is the boldest consumer protection play against AI in years, and I want to like it. I just cannot, because the theory underneath it does not survive contact with how AI systems are actually built.
The proposal, which follows Executive Order 14365, argues that consumers reasonably expect AI systems to strive for accurate responses and faithfully carry out requests. When a developer deliberately configures a model to advance objectives other than accuracy, whether ideological goals or a simple aversion to controversy, and does not tell anyone, the FTC calls that steering and treats it as a deceptive act.
The Fiction at the Center of It
Here is the problem. The FTC's framework assumes there is a clean baseline called "accuracy" that a model can be steered away from, like a car drifting off a road. In reality there is no road. Every frontier model is a bundle of thousands of design decisions, and each one shapes what counts as a good answer. Training data choices, reward modeling, safety filters, system prompts, temperature settings: all of it steers output in one direction or another.
The Colorado Artificial Intelligence Act angle makes the tension sharper. The FTC's statement points at the new state law, which can hold companies liable for discriminatory model outputs, and warns that even efforts to comply with it could compromise accuracy and trigger federal deception concerns. A developer can be sued by Colorado for outputs that discriminate and by the FTC for the very changes made to avoid that outcome. That is not a roadmap; it is a trap.
The Safety Tuning Catch-22
This is where the pushback gets loud, and it is coming from both wings of the political spectrum. The Center for Democracy and Technology, the Cato Institute, the Electronic Frontier Foundation, and more than twenty state attorneys general have filed comments opposing the proposal. Their core objection is the "technical fiction" that safety tuning can be cleanly separated from steering.
To the FTC's credit, the statement tries to carve out safe zones. Ordinary hallucinations do not violate Section 5 on their own. Blocking illegal content and preventing cyberattacks are fine. What is left unaddressed is the enormous middle: the daily work of making a model refuse harmful requests, resist manipulation, and stay on the rails. That work changes outputs, by design.
- Undisclosed steering toward ideological objectives: deceptive.
- A one-time disclosure buried in fine print: not enough.
- Wrong answers caused by real technical limits such as hallucinations: not deceptive on their own.
- Blocking illegal content or stopping cyberattacks: not deceptive.
- Overstating how rarely a model makes mistakes: potentially deceptive.
The disclosure standard is admirably strict, at least in theory. Fine print in a terms-of-service page will not cut it, and a one-time notice that gets "subsequently hidden" fails too. The disclosure must clearly and conspicuously dispel the notion that the system is designed to give the best answer possible. I genuinely appreciate the FTC demanding that companies say what their models actually do. But here is the uncomfortable question: how many companies would run that disclosure and still have a product left to sell?
A Crusade That Picks Favorites
Then there is the enforcement record, which is where my skepticism hardens into something sharper. Chair Andrew Ferguson sent compliance letters to AI companies this month, and according to reporting by Tech Times, the FTC's own policy document names Anthropic more than half a dozen times while Grok, despite Elon Musk's admitted interventions in his chatbot, goes effectively unmentioned.
I am not going to litigate whether Anthropic deserves scrutiny. The point is optics, and the optics are terrible. When the agency writes a policy statement about hidden bias in AI, and the company that explicitly trains its model to refuse certain topics keeps appearing while the company whose owner openly admits to steering his chatbot gets a pass, the word "steering" starts to describe the regulator.
None of this means the FTC should abandon the problem. Undisclosed output manipulation is a real consumer issue, and the agency is right that there is no AI exemption from consumer protection law. But a policy built on the myth of an unsteered model, enforced with visible favorites, will produce either a regulatory mess or a political circus. Consumers deserve an honest answer about what their AI is doing. They also deserve rules that do not pretend neutrality is a switch companies simply forgot to turn on.
The FTC has the right instinct and the wrong framework. Fix the fiction, fix the favorites, and there might be a policy here worth keeping. As written, it is an accuracy crusade that cannot even define accuracy, enforced by an agency that cannot seem to keep its own thumb off the scale.
Comments