The FTC Says Complying With State AI Laws Can Still Be Illegal

The Federal Trade Commission has drawn a line that lands squarely on every company shipping a chatbot in the United States: if you quietly reshape what your AI tells people, saying a state law made you do it will not save you.

That warning sits inside a proposed policy statement the agency published in the Federal Register on July 7, titled “Suppression of Accuracy in Artificial Intelligence Systems.” It is open for public comment, and the window is nearly shut: the FTC says filings must arrive on or before Friday, July 31. The document is short for a federal filing, but its logic reaches into the core promise every AI vendor makes when it markets a model as helpful, accurate, and on your side.

What the FTC actually said

The agency’s argument runs through Section 5 of the FTC Act, the nearly ninety-year-old ban on “unfair or deceptive acts or practices” that the Commission uses against everything from fake supplement claims to hidden subscription traps. Nothing about that authority stops at software. AI products, the Commission writes, “are not exempt from section 5’s long-established, and generally applicable reach.”

Here is the chain the FTC builds. AI companies have spent years telling the public — directly in marketing and implicitly in the way the products behave — that their systems try to give the best, most accurate answer they can. Consumers believe it. They now lean on chatbots for research, medical questions, financial decisions, and legal or personal advice. Because of those representations, the agency argues, people have a reasonable expectation that an AI is trying to be truthful, not steering them toward some undisclosed goal.

So when a company secretly tunes a model to serve a different objective — and hides that the accuracy the user expected has been traded away — the FTC’s position is that the company may be deceiving its customers. The motive does not matter. As the statement puts it, that can be a violation “even if the deceptive steering is done in an effort to comply with State laws.”

The state-law collision at the center

That last clause is the part with teeth, and it is aimed at a specific fight.

The Commission was directed to write this statement by Executive Order 14365, “Ensuring a National Policy Framework for Artificial Intelligence,” which President Trump signed on December 11, 2025. The order’s stated goal is a single “minimally burdensome national policy framework for AI—not 50 discordant State ones,” and it told the FTC to address how state laws that force changes to accurate AI outputs collide with federal law. The backdrop is a wave of state activity: the White House has counted more than 1,000 AI bills introduced across state legislatures.

The FTC names Colorado’s Artificial Intelligence Act as its lead example. The agency reads the law — in both its original and its recently revised form — as pressuring companies to avoid outputs that could produce a “disparate impact,” even holding developers liable for how customers use their products. The concern the Commission spells out is that a firm might quietly suppress accuracy to dodge that liability, then keep the change hidden so users never learn the answers were shaped by something other than the facts.

The Commission is blunt about the escape hatch companies might reach for. When Congress wrote the deception ban, it “did not provide State-law safe harbors,” the statement says — so even where a state’s rules shape how a product works, the company still has to satisfy Section 5. A regulator telling businesses that “we made you do it” is not a defense is the whole point of the filing.

Then comes the legal move that AI general counsels will be studying closely. The FTC concedes its own statute “does not expressly preempt State law.” But it argues that “State law is impliedly preempted to the extent it conflicts with a Federal regulatory scheme.” A state rule that, in the agency’s telling, requires a company to deceive its own customers “obviously conflicts” with Section 5’s purpose. In other words: the FTC is positioning federal deception law as a shield companies can raise against state mandates — and a sword it can turn on companies that comply with those mandates the wrong way.

One more detail matters for how aggressively this could be enforced. Under the Commission’s long-standing deception standard, it does not need to show that everyone was misled — courts have accepted that even a “significant minority” of consumers, sometimes as small as roughly 10 percent, can be enough. For a product used by tens of millions of people, that is a low bar to clear.

Why this is really about trust

Strip away the politics and this is a consumer-trust problem, which is where it gets uncomfortable for anyone who uses these tools.

When a model is steered, you usually cannot tell. The answer still arrives in the same confident tone, formatted the same way, with the same air of having searched everything. The FTC leans on that exact gap. It cites how people actually use AI now — teenagers running homework through chatbots, adults asking for financial advice, patients describing symptoms — and points to research finding that the overwhelming majority of users do not fact-check what a model tells them. If the output has been bent and nobody says so, the person on the other end has no way to correct for it.

TECHi has tracked how quickly this reliance has spread into high-stakes work, from AI tools moving into junior analyst seats on Wall Street to enterprises buying agents that act on their behalf. The more decisions ride on a model’s answer, the more a hidden thumb on the scale actually costs. That is the harm the FTC is trying to name.

The one exit the FTC leaves open

The statement is not a ban on building models with rules, guardrails, or refusals — the kind of safety limits vendors already layer onto consumer chatbots. What the agency demands is honesty about them.

A company can, the FTC says, “clearly and conspicuously disclose that its systems are designed to produce outputs that prioritize certain objectives over what users request and otherwise expect.” Do that well enough and the deception problem falls away, because you have reset what a reasonable person should expect.

The catch is the standard. This is not a line buried in the terms of service. The disclosure has to be prominent and persistent, the agency warns, and a one-time notice later hidden in fine print is unlikely to be enough. The further your model’s real behavior drifts from what users would naturally assume, the louder you have to say so. For product teams, that turns a legal footnote into a design question: how do you tell a user, in the moment, that the answer they are reading has been shaped by something other than accuracy — without torching their trust in the product?

The part everyone will argue about

None of this is settled, and the fight over it will be loud.

The statement is openly political in its framing. It grows out of the administration’s push against what its executive orders call “woke AI,” and it repeatedly frames output-shaping as the insertion of “ideological bias” or “equity” objectives. That framing invites obvious objections. “Accuracy” is not always a clean, agreed-upon fact — reasonable people dispute what a correct answer even is on contested questions, and critics will argue the FTC is smuggling a viewpoint into a consumer-protection rule. There are live First Amendment questions on the other side, too, about how far the government can go in dictating what a private company’s model must or must not say. The agency notes its statement is not meant to force AI systems to produce clearly illegal content, and does not stop companies from blocking security abuse.

There is also the plain question of reach. This is a proposed policy statement, not a final rule or a court ruling. It signals how the current Commission intends to enforce Section 5; it does not, by itself, strike down Colorado’s law. Whether “implied preemption” holds up is something judges, not press releases, will decide. Governments restricting or shaping AI is not new — TECHi has covered moves from China pulling AI tools out of banks and state agencies to Washington’s own national AI strategy aimed at out-competing rivals. What is new here is a US regulator telling companies that following a state law could itself be the violation.

What to do before July 31

For anyone building or buying AI, the practical response is not to panic — it is to look at your disclosures.

If you ship a model, this is the moment to inventory where and how you shape outputs, and whether users are told in a way that would survive the “clear and conspicuous” test the FTC just described. If you buy AI for a business, it is a fair question to put to vendors: what are you steering, and where do you say so? And if you simply use these tools, the quieter lesson stands on its own — a fluent, confident answer is not the same as an accurate one, and even the regulators now assume some of what you read has been shaped by hands you cannot see. The comment window closes July 31, and the companies with the most to lose are already writing.