On July 1, 2026, the Federal Trade Commission issued a Proposed Policy Statement Concerning the Suppression of Accuracy in Artificial Intelligence Systems. The comment period closed July 31. It is a short document with an unusually large blast radius, and almost nobody outside the AI policy world noticed it.
Here is what it says, why it exists, and why the Electronic Frontier Foundation has asked the Commission to throw it away.
The theory
Section 5 of the FTC Act prohibits “unfair or deceptive acts or practices.” The deception prong is well-worn: if you represent something material to consumers, and the representation is false or misleading, and a reasonable consumer would rely on it, you have a Section 5 problem.
The policy statement applies that framework to AI outputs. The argument runs roughly:
- AI companies represent — explicitly or implicitly — that their systems provide accurate information.
- Some of those companies configure their systems to prioritize objectives other than accuracy.
- When they do that without telling users, the representation becomes misleading.
- Therefore, deliberately steering a model away from the most accurate answer, without disclosure, may be a deceptive practice.
Stated that abstractly, it is not obviously wrong. If a model is tuned to recommend its parent company’s products when asked a neutral question, that is a real deception, and it is exactly the kind of thing Section 5 exists for. Undisclosed commercial steering inside a system users treat as an oracle is a genuine consumer protection problem.
Where it came from
The statement was not generated by the Commission’s own agenda-setting. It was issued pursuant to Executive Order 14365, signed December 11, 2025, which directed the FTC to clarify how its deception framework applies to AI models and — this is the operative clause — to address potential conflicts between federal law and state laws requiring alterations to the accurate outputs of AI models.
That second directive is what turns a consumer-protection document into something else.
The problem
Read the enabling language again: state laws requiring alterations to the accurate outputs of AI models.
State AI laws in force or pending in 2026 do a range of things. Some require bias testing and mitigation in high-risk systems — hiring, lending, housing. Some require guardrails on outputs to minors. Some require disclosures, watermarking, or restrictions on synthetic media. Some require models not to produce certain categories of content.
Every one of those is, if you squint, a law that causes a model’s output to differ from whatever the raw weights would have produced. A framework that treats “deviation from the most accurate answer” as presumptively deceptive, while explicitly tasked with resolving conflicts with state law, is a framework built to characterize state AI safety compliance as a federal deception problem.
That is a preemption argument wearing consumer-protection clothing. And it arrives at a moment when the states are the only jurisdictions in the United States actually legislating on AI.
The second problem: “accuracy” is doing impossible work
The entire structure rests on there being a determinate “most accurate answer” from which a model can be shown to deviate. For a narrow factual question — what year did a treaty get signed — that is fine.
For most of what people actually ask these systems, it is not a coherent concept. Questions about contested history, medical judgment under uncertainty, legal risk, political causation, and social science do not have a single retrievable correct answer that a regulator can hold up against the model’s output. What they have is a distribution of defensible positions, and every model produces something by making choices about that distribution — choices baked in at pretraining, at fine-tuning, at RLHF, at system-prompt time.
A model that declines to give confident medical advice is “suppressing accuracy” under one reading and exercising appropriate epistemic humility under another. A model tuned to present multiple mainstream views on a contested question is either being balanced or being evasive, depending entirely on who is characterizing it.
An enforcement standard that cannot distinguish those cases is not an enforcement standard. It is discretion, and discretion applied to what models are allowed to say has a specific name.
EFF’s position
EFF has urged the FTC to withdraw the proposal. The core of the objection is that a government agency deciding which model outputs count as impermissibly “suppressed” is government involvement in editorial judgment about speech — and that the First Amendment problems with that do not disappear because the speaker is a matrix of weights rather than a newspaper.
There is a consistent principle underneath EFF’s position that is worth naming, because it also explains its opposition to the youth AI bills moving through the Senate: government mandates on how systems shape what users see are speech regulation, whether the stated goal is protecting children or ensuring accuracy.
What this has to do with privacy
Two things, and neither is obvious at first glance.
First, the enforcement mechanism is inspection. Determining whether a model was “configured to steer outputs away from the most accurate answer” is not something you can do from the outside. It requires visibility into training data, fine-tuning procedures, system prompts, RLHF reward models, and output filters. A Section 5 theory built on model internals implies a regulatory relationship in which those internals — including whatever personal data sits in the training corpus — become discoverable.
Second, it cuts against the direction privacy regulation is heading. The GDPR-style position, and the position taken by most state AI laws, is that automated systems making consequential decisions about people must be constrained, tested, and sometimes deliberately adjusted away from what the raw model would output — because the raw model reflects the biases in its training data. A federal framework that recasts those adjustments as deception puts the two regimes in direct opposition.
You cannot simultaneously require a hiring model to be de-biased and treat de-biasing as suppression of accuracy. Someone has to lose.
Where it stands
The comment period is closed. What the Commission does with the record is unknown, and a proposed policy statement is not a rule — it does not have independent legal force, and it is not directly reviewable in the way a rulemaking would be.
What it is, is a statement of enforcement intent. Companies read those and adjust. That is the point of publishing one.
If the Commission finalizes it as drafted, the practical effect is that any AI company facing a state law requiring output adjustments has a federal document to point at when arguing the state requirement puts it in conflict with Section 5. That argument does not have to win in court to be effective. It only has to make state compliance expensive enough to contest.
The pattern
Watch for the shape of this, because it will recur: a genuine consumer protection problem — undisclosed commercial steering in AI systems, which really does exist and really should be enforced against — used as the vehicle for an authority that mostly does something else.
The narrow version of this policy statement would be worth having. The version that arrived carries an executive directive about state law preemption inside it, and that is not a detail. It is the reason the document exists.
