FTC AI Accuracy Policy Traps Builders Between Two Laws
The FTC's proposed AI accuracy policy statement, out July 1 and open for comment until July 31, 2026, says steering a model's output away from the correct answer without disclosure can be deception under Section 5. It was ordered by Trump's Executive Order 14365 and singles out Colorado's AI Act as impliedly preempted. The bigger signal: federal and state law now point AI teams in opposite directions over what a model is allowed to say. For PMs, the liability lands not on the labs but on the companies deploying these models into regulated decisions.
Ask yourself an uncomfortable question this week: if a regulator demanded it tomorrow, could you produce a written record of every place your product alters a model's raw output — and prove what you told users about each one? Most teams can't. Guardrails, fairness tuning, and safety filters get shipped as engineering decisions, not documented disclosures, and the FTC's draft just turned that gap into legal exposure. The concrete move: run a one-week output-modification audit. List every layer between the model's response and what the user sees, note why it exists, and flag which ones exist to satisfy a state law like Colorado's. That inventory is your first line of defense whether the fight ends up federal or state. To be fair to the FTC, the honest core of this is disclosure, not a ban on tuning — so the fix is often as simple as telling users, in plain language, that outputs are moderated and why. That's cheap insurance. The reason to move now: the comment period closes July 31, 2026, and once the statement is final, "we didn't realize our guardrails counted" stops being a defense. Preemption is still unlitigated, which means for the foreseeable future you're exposed to both regimes at once. The teams that document early will be the ones not scrambling when the first enforcement action names a deployer instead of a lab.
The FTC's July 1 draft policy statement says steering an AI model away from the correct answer without disclosing it can be deception under Section 5 of the FTC Act.
Picture a compliance lead at a small lending startup in Denver, two browser tabs open on the same screen. One is Colorado's Artificial Intelligence Act, which tells her the company's credit-scoring model must be tuned so it doesn't produce a disparate impact on protected groups. The other, published July 1, is a proposed FTC policy statement warning that altering a model's output to dodge that exact kind of state liability could be an illegal deceptive act under federal law.
Same model. Two governments. Opposite instructions.
This draft doesn't hand AI teams the clarity they've been begging for. It hands them a two-front legal war — and the public comment window closes July 31, 2026.
Two governments, one model, opposite orders
According to the FTC, the proposal — formally the "Policy Statement Concerning the Suppression of Accuracy in Artificial Intelligence Systems" — came out of Executive Order 14365, the Trump administration's national policy framework for AI signed in December. The order directed the agency to spell out how Section 5 of the FTC Act applies when developers alter model outputs to satisfy a state law.
The FTC's answer, stripped of the legalese: steering a model away from what it would otherwise output as the correct answer, without telling the user, can be deception. Chairman Andrew Ferguson framed the effort as protecting people from "the subversion of AI systems for ideological ends," and said the agency wants to hear from businesses and consumers about their experiences.
Now read that next to a state statute that requires you to adjust those same outputs. You start to see the shape of the trap. Comply with Denver, and you may have handed Washington its case.
Comply with Washington, and Denver's regulators are still holding a discrimination claim.
What "steering" means when a federal regulator says it
The draft lists the conduct that worries the FTC: modifying outputs to advance an ideological goal, bending answers under political or public pressure, and — the one that should make product teams put down their coffee — changing outputs specifically to avoid liability under a state law. The agency names Colorado's AI Act directly, arguing the law pressures companies to suppress accuracy to escape disparate-impact exposure, and concludes the state law is impliedly preempted to the extent it conflicts with Section 5.
That's the constitutional drama. The operational drama is quieter, and in some ways worse.
Most teams don't think of their safety filters, their RLHF tuning, their fairness constraints, or their content guardrails as "altering truthful output." They think of it as responsible engineering — the stuff you build so the model doesn't say something reckless to a real customer. The FTC just reframed a routine layer of the ML pipeline as a candidate for a Section 5 violation whenever it's undisclosed.
To be fair to the FTC, there is a genuine consumer harm buried under the politics. Someone who asks an AI a medical or financial question deserves to know whether the answer was shaped by something other than the model's honest best estimate. The stated target is nondisclosure, not tuning itself — a distinction the final statement will need to hold onto if it's going to survive contact with how software actually gets built.
The bind lands on the deployment team, not the lab
Here's the part that isn't in the press release. This risk doesn't settle on the frontier labs training foundation models. It settles on the thousands of companies deploying them: the fintech wrapping an LLM for underwriting, the HR platform screening résumés, the health app triaging symptoms at 2am.
Think of it like renting a commercial kitchen where one inspector demands every dish come out well-done and a second inspector cites you for altering the food from how it arrived at the door. You didn't write the recipe or build the oven. You're just the one standing at the pass when both inspectors walk in.
The stakes sharpen when you notice how unsettled this is. Colorado's law and the FTC policy carry different definitions, different enforcement mechanisms, and — the load-bearing detail — no court ruling on whether preemption actually holds. "Impliedly preempted" is the FTC's legal argument, not a judge's decision. Until that gets litigated, a deploying company sits exposed to both regimes at once, as of mid-2026, with lawyers on retainer and no safe harbor to point to.
And the map is about to get more crowded. Colorado is the named example, but it isn't the only state drafting AI accountability rules, each with its own idea of what a "fair" or "accurate" output should look like. A regulatory patchwork was already coming for AI.
What this draft does is turn the patchwork into an outright contradiction — federal and state pointing product teams in opposite directions over the same line of model behavior.
The harder question leadership is actually wrestling with
So the question on the whiteboard next quarter isn't "is our model accurate?" It's colder than that: can you prove, on demand, why your model said what it said — and show exactly what you disclosed to the user about it? Documentation and disclosure, the least glamorous items on any roadmap, are shaping up to be the only defensible ground when two regulators aim in opposite directions.
The deeper story here is about who ends up holding the risk in the AI supply chain. The labs ship the capability. The deployers absorb the liability.
That asymmetry was survivable when the hardest question was uptime. It reads very differently when it's the gap between a federal deception charge and a state discrimination suit — and the people caught in the middle are builders who never got a vote on either rule.
The comment period runs through July 31. If your product touches a regulated decision, this is not a memo to skim.
Get this in your inbox. AI Rundown Daily delivers original briefings every morning — free. Subscribe →
Frequently Asked Questions
It's a proposed policy statement, not a finalized rule, and it reflects the current administration's priorities. But the FTC has real Section 5 authority behind it, and policy statements shape how the agency chooses enforcement targets. The nondisclosure theory it advances is a plausible legal hook, not empty signaling. Treat it as a preview of enforcement posture rather than settled law you can ignore.
The cheapest and highest-leverage step is documentation, which is mostly labor, not spend. Inventory every place you modify model output and record why, then add plain-language disclosure that outputs are moderated. That combination directly addresses the FTC's stated concern, which is undisclosed steering, not tuning itself. Heavier costs come only if you're forced to reconcile a genuine conflict between a state mandate and the federal position, which is where legal counsel becomes unavoidable.
Nobody knows yet, and that uncertainty is the risk. The FTC argues Colorado's AI Act is impliedly preempted where it conflicts with Section 5, but that's an assertion no court has ruled on. Until a judge decides, you can be exposed to both a state disparate-impact claim and a federal deception theory over the same output. The defensible position is thorough documentation and disclosure that holds up under either regime.
