The Chatbot's Hidden Boss Is FTC Territory

The FTC just declared that any AI company steering outputs toward hidden objectives — safety refusals, political filters, equity patches, state-law compliance — without telling users is likely committing consumer fraud under Section 5, with comments closing July 31 and Colorado's AI liability law directly in the crosshairs.

What's new: On July 7, the Federal Trade Commission published a proposed policy statement in the Federal Register — "Policy Statement Concerning the Suppression of Accuracy in Artificial Intelligence Systems" (91 FR 41638). The Commission voted 2-0. Comments are due July 31 at regulations.gov, Docket No. FTC-2026-0859.

Why it matters now: The same week the Meta Oversight Board reported that frontier models refused political-criticism requests 34% of the time for authoritarian jurisdictions versus 14% for democracies — all queried from Australia — the FTC put a federal price tag on undisclosed steering. Labs already shape what users can say and learn. Washington just said hiding that agenda is fraud.

The FTC's theory:

• AI companies market systems as problem-solvers that maximize user objectives.
• Consumers reasonably expect truthful, accurate outputs — not hidden secondary agendas.
• Steering away from that expectation without disclosure is material misrepresentation.
• Motives don't matter: profit, ideology, and state-law compliance all face the same test.

As the statement puts it: "Consumers have no basis to believe that AI systems aim to produce outputs that are distorted by undisclosed ideological objectives."

The 91% problem: The FTC cites data from a major AI developer showing consumers accept outputs without fact-checking over 90% of the time — even though systems only strive for, not guarantee, accuracy. If users overwhelmingly trust the answer, the Commission argues, hiding who's really steering it is deception.

Colorado in the crosshairs: The statement names Colorado's revised Artificial Intelligence Act (S.B. 26-189, enacted May 14, 2026), which can hold AI companies liable for discriminatory outcomes caused by customers' use of their products. The FTC predicts labs will suppress accuracy to embed "equity" and avoid state liability — then fail to disclose the tradeoff.

The Commission's preemption line is blunt: "A State law that requires an AI firm to deceive its consumers obviously conflicts with section 5's express purpose of protecting consumers from such conduct."

What's NOT covered:

• Ordinary hallucinations from technical limits — not deception on their own.
• Blocking illegal content or preventing cyberattacks.
• Steering users explicitly request — like creative fiction.

The safe harbor: Clear, conspicuous, persistent disclosure that the system prioritizes objectives beyond what users asked for. Burying it in terms of service won't work. The FTC warns that "a one-time disclosure subsequently hidden away in fine print" is doubtful to suffice.

The federal-state collision: The FTC argues complying with Colorado-style statutes is not a Section 5 defense. State laws forcing output changes are impliedly preempted where they conflict with federal consumer protection. Enterprise vendors operating across jurisdictions now face a disclosure architecture problem, not just a safety tuning question.

The Brisbane proof point: Thursday's Meta Oversight Board audit tested 10 models from six providers — Anthropic, DeepSeek, Google, Meta, OpenAI, and xAI — and found models "more than twice as likely" to refuse criticism of repressive regimes. Gemini cited Thai lèse-majesté from Brisbane. DeepSeek cited Saudi speech laws from Australia. The Board warned these refusals have "the practical effect of extending the long arm of restrictive governments across borders to limit speech in free countries." None of that steering was disclosed to users at query time.

Who gets hit: Foundational model labs. Enterprise deployers wiring third-party AI into customer-facing products. Financial institutions whose consumers increasingly turn to AI for advice — the FTC explicitly cites banking survey data showing a majority of consumers now use AI for financial guidance.

Convina's view: The FTC just did what labs refused to: name the second boss in the chatbot. For two years enterprises bought "helpful, accurate assistants" while vendors quietly layered geopolitical refusals, equity patches, and safety filters nobody disclosed. That era ends July 31 unless the comment period blunts it. The fix isn't stopping steering — it's labeling it. Any procurement team that can't explain whose objectives their model actually optimizes is about to inherit Section 5 exposure.