The FTC's proposed policy on AI accuracy treats a deliberately adjusted model as a potential deception while excusing one that is simply wrong by accident. Public comments close July 31, and the statement already names Colorado's AI Act as impliedly preempted, on a 2-0 Commission vote. The argument is not really about whether these systems tell the truth; it is about who is allowed to touch the dial.

When the Guardrail Becomes the Fraud: The FTC's Quiet Redefinition of an Honest Machine

A proposed federal policy says an AI system that is accidentally wrong raises no concern, while one deliberately corrected may be deceiving you. Public comments close July 31.

By The Daily Reflection · July 20, 2026 · 10 min read

---

On July 1, the Federal Trade Commission proposed a policy statement with a title that sounds like housekeeping: "Suppression of Accuracy in Artificial Intelligence Systems."

It is not housekeeping. It is the first serious federal attempt to treat what an AI model says as a claim the company can be sued over, and it arrives with a specific target already named in the press release. Colorado's Artificial Intelligence Act, the FTC wrote, "appears to coerce companies into altering the output of their AI models," and is therefore "impliedly preempted to the extent it conflicts with a federal regulatory scheme."

Chairman Andrew Ferguson framed the request for comment plainly. The Commission, he said, wants to hear about "the subversion of AI systems for ideological ends," and the resulting policy will advance the president's goal of "expanding America's global dominance in artificial intelligence."

The vote authorizing all of this was 2-0.

Comments close on July 31. Almost everything written about the statement so far has been law firm client alerts explaining compliance exposure, which is useful if you are general counsel at a model developer and beside the point if you are anyone else. The interesting thing here is not the compliance risk. It is what the document quietly does to the word "accuracy."

---

What the Statement Actually Says

Section 5 of the FTC Act prohibits unfair or deceptive acts or practices. For a century it has policed the gap between what a company promises and what it delivers: the mattress that is not organic, the loan whose rate is not fixed, the supplement that does not do what the box says.

The proposed statement extends that logic one step further than it has ever gone. Its premise is that AI companies have made explicit and implicit representations that their systems produce accurate, objective, useful answers. Consumers rely on those representations. So if a company intentionally designs a system to pursue undisclosed objectives that differ from what users reasonably expect, that design choice can itself be deception, whether it shows up in the marketing or only in the model's behavior.

Note the shift. Earlier FTC action in this space went after false claims about AI, as in the 2025 order against DoNotPay over its "robot lawyer" advertising. That is ordinary consumer protection: the company said something untrue.

This is different. Here the product's own output becomes the representation. The company need not have lied in any brochure. It only needs to have tuned the model toward a goal the user did not know about.

The proposal lists the conduct it has in mind: modifying outputs to advance ideological goals, changing outputs in response to political or public pressure, and altering outputs to avoid liability under certain state laws.

That third item is the one doing the real work, and we will come back to it.

---

Accuracy Is Not What This Document Means by Accuracy

Here is the part almost no coverage has flagged, and it is the whole hinge of the thing.

The FTC explicitly carves out hallucinations. When a model states something false because of technological and resource limitations, the statement says, that does not by itself raise Section 5 concerns. Those errors come from the limits of the machine rather than a design decision, so they are not deceptive.

Now put the two halves together.

A model that tells you something false because it does not know any better is fine. A model that was deliberately adjusted, including toward being more careful, may be deceptive.

Read that again, because it is the opposite of how every other consumer protection regime in American law works. Ordinarily, intent aggravates. A company that ships a defective product by accident is in less trouble than one that shipped it knowingly. Here, intervention is the thing that creates exposure. The unedited machine is presumptively honest; the edited one has to justify itself.

A standard that punishes deliberate correction and excuses accidental falsehood is not a standard about accuracy. It is a standard about who is allowed to touch the dial.

You can see the practical consequence immediately. Every safety measure a lab applies after pretraining is, mechanically, output steering: refusing to give synthesis routes for nerve agents, declining to generate sexual content involving minors, adding hedges to medical answers, suppressing a confident answer in favor of "consult a professional." All of it is intentional deviation from what the raw model would have said.

Nobody at the FTC is proposing to prosecute a lab for refusing to help with a bioweapon. That is not the concern, and pretending otherwise would be dishonest. But the document draws its line at disclosure rather than at category of intervention, which means the legal question is never "was this steering good?" It is "did you tell the user you were steering?" Those are very different questions, and only one of them has anything to do with truth.

---

Colorado Is the Target and Implied Preemption Is the Weapon

The statement did not emerge from the Commission's own study of consumer harm. It was directed. Executive Order 14365, signed December 11, 2025, instructed the FTC to explain how Section 5 applies when developers alter model outputs in response to state law requirements. The order sets out a national policy framework for artificial intelligence, and the White House published it under a heading about eliminating state law obstruction of that framework, which says the quiet part before you reach the text.

Colorado's AI Act, enacted May 14, 2026, imposes potential liability on AI companies when their customers' use of a system produces discriminatory outcomes. The FTC's argument runs like this: that liability pressures developers to modify outputs; modified outputs may deceive consumers; therefore a state law compelling the modification conflicts with federal consumer protection law and is impliedly preempted.

It is an elegant piece of legal engineering, and it should worry people regardless of what they think about Colorado's statute.

Preemption normally requires Congress to legislate, or at minimum an agency to complete a rulemaking that carries the force of law. A policy statement does neither. It is not binding. It creates no new legal obligation. What it does is announce how the agency intends to read an existing statute, which means the doctrine gets built in a document that never has to survive a vote in either chamber.

A state law can now be squeezed into irrelevance by an agency's stated interpretation of a 1914 statute, at the direction of an executive order, on a 2-0 vote.

Whatever you think of the Colorado law, notice the machinery. It works exactly as well in the other direction. A future commission could declare that failing to steer outputs is the deception, and the same architecture would carry it.

---

The Disclosure Escape Hatch Gives the Game Away

The proposal offers companies a way out. Steering is curable through clear and conspicuous disclosure, and the FTC is specific that the disclosure has to be prominent enough to actually change consumer expectations. Burying it in terms of service will not do.

Sit with what that means. If a model developer posts a visible notice explaining the objectives its system pursues, the deception problem dissolves. The outputs do not change. The steering does not stop. Only the label changes.

So the remedy the statement contemplates is not a more accurate machine. It is a labeled one.

That is a defensible policy goal, and it is roughly what Europe is pursuing through the AI Act's transparency rules. But it is a strange fit with a document titled "Suppression of Accuracy," and it tells you the real anxiety is not that models are wrong. The anxiety is that models are shaped, that the shaping reflects somebody's values, and that users cannot see whose.

That anxiety is legitimate. It is also, as stated here, unfalsifiable in practice. Every model reflects choices: what data went in, what got filtered out, which answers were rewarded during training. There is no unshaped baseline to compare against, no neutral model sitting in a vault that the tuned one departs from. Asking a company to disclose its deviations from objectivity assumes an objective version exists, and it does not.

---

The Case for the Commission, Made Honestly

It would be easy to stop there, and unfair to, because the case for some version of this is not crazy.

Undisclosed steering is a real harm, and it does not require any ideological framing to see it. A model tuned to favor its parent company's products in shopping advice is deceiving you in a way that maps cleanly onto a century of consumer protection law. A financial chatbot nudged toward the sponsor's mutual funds is a kickback with a chat interface. A medical assistant tuned to reduce liability rather than to give the best answer is trading your outcomes for the company's legal exposure, silently. Those are ordinary frauds wearing new clothes, and Section 5 is exactly the right tool.

The FTC is also right that people trust these systems in a way the marketing invited and the technology has not earned. If a company advertises a research assistant and ships a persuasion engine, somebody should be able to sue.

The problem is not the theory of harm. It is the drafting. A statement aimed at self-dealing and covert commercial influence would be narrow, boring, and probably bipartisan. This one instead centers ideological subversion, names a state civil rights law as its example, and arrives with an executive order attached. It reads less like consumer protection that happens to touch politics and more like politics that happens to invoke consumer protection.

---

The Bottom Line

Strip away the procedural detail and one question is left sitting on the table: who gets to decide what an honest machine sounds like?

For three years the answer has been the labs, which nobody was happy about. The proposed answer now is the FTC, applying a deception statute written when the cutting edge of consumer fraud was patent medicine, at the direction of an executive order, with state legislatures cut out of the loop by a preemption theory that never passed through Congress.

The document's own logic gives away that this was never really about accuracy. Accidental falsehood is excused. Deliberate correction is suspect. A prominent label cures everything. Those three rules cannot be reconciled with a concern about whether AI systems tell the truth, but they reconcile perfectly with a concern about whether anyone is quietly putting a thumb on the scale, and about which thumbs are objectionable.

Comments close July 31. The docket is open to anyone, which in practice means it will be filled by the companies with the most at stake and the trade groups they fund, unless people who are not lawyers decide to file something too.

Whatever the final statement says, the precedent will outlast this administration. Somebody will always be interested in the dial. The only real question is whether we build the rules while pretending there is a neutral setting to return to.

---

The Daily Reflection cuts through the noise to find the stories that actually matter. Follow for thoughtful takes on politics, technology, and whatever's shaping our world.