An AI can be steered in silence.
The FTC wants that banned.
On July 7, the Federal Trade Commission opened public comment on a policy statement declaring that AI companies which steer their systems’ answers toward undisclosed objectives — rather than what a user actually asked — may be committing deception under federal law. The first regulatory acknowledgment that an AI’s answer was never guaranteed to be neutral, and the first proposed cost for not admitting it.
An AI system can get every fact right and still not be answering you honestly. Not by lying about a name or a number — by choosing, quietly, which priority wins when your question and someone else’s objective disagree, and never mentioning that a choice was made at all. On July 7, the Federal Trade Commission proposed a name for that mechanism, and for the first time, a federal cost for hiding it: a company that steers its AI’s outputs toward objectives other than what a user asked for or reasonably expects — without disclosing that trade-off — may be committing deception under Section 5 of the FTC Act.
The proposed policy statement doesn’t ask an AI to be right. It asks the company behind it to admit when “right” wasn’t the only thing being optimized for. If a firm tunes its model to satisfy a compliance mandate, a brand-safety rule, or any priority that sits above the plain question a user typed — and says nothing about it — the Commission’s theory treats the silence as the violation, not the tuning.
Comment on the proposal closes July 31. Read narrowly, it hands every AI company a compliance path that costs almost nothing — say, in writing, what the system is actually built to prioritize.
What the proposal actually reaches
Nondisclosure is the offense, not the steering itself. Steering a model toward a disclosed priority is lawful under the proposal. The violation is a gap between what a user reasonably expects — an accurate answer — and what the system is actually built to deliver, left unstated.
Compliance with another law is not a shield. The proposal treats state-mandated output changes the same as any other undisclosed steering: if a user isn’t told an answer was shaped by something other than accuracy, the source of that shaping doesn’t excuse it.
Disclosure is the entire remedy on offer. Nothing in the proposal requires a company to change what its model optimizes for. It requires the company to say so, in language plain enough that a reasonable user could tell the difference.
What it doesn’t reach at all
A company can clear this bar completely — disclose its priorities in full, invite no deception claim whatsoever — and still leave a real business invisible inside its own answers. Disclosure and recognition are not the same failure. A system can be perfectly honest about what it optimizes for and still misclassify a firm’s authority, miss its corroborating signals, or default to whichever competitor happens to be structured so the system can read it. That gap has a name already, and it predates this proposal by years: Digital Derangement Syndrome™, the recognition failure inside the systems that now decide who gets cited before a person ever asks.
The FTC is regulating one half of the problem — the half where a system hides, by omission, what it’s actually optimizing for. The other half is structural, not legal: a system can be fully transparent about its priorities and still never see you at all. Identity Architecture™ is built for that second half — machine-readable structure that lets an inference system classify a firm’s authority correctly, whatever the system discloses about the rest of its priorities.
Eighteen editions in, the pattern holds: every new rule written for this era arrives narrower than the problem it names. The FTC just proposed making it illegal to hide that an AI’s answer has an agenda. It said nothing about the far more common failure — a system with no hidden agenda at all, that simply never learned you exist.
The Federal Trade Commission published a proposed policy statement on the suppression of accuracy in artificial intelligence systems on July 7, 2026, opening public comment through July 31 — Federal Register.
The Commission’s own announcement frames the theory: undisclosed steering of AI outputs toward objectives other than what a user requests may violate Section 5 of the FTC Act — FTC press release.
Legal analysis of the proposal’s scope and compliance path — Spencer Fane.