The FTC is gathering AI evidence under old law

What happened

The Federal Trade Commission sent civil investigative demands, often described as civil subpoenas, to OpenAI and Anthropic seeking records about safety incidents and potential consumer-protection violations, Axios reported on October 15. Reuters carried the report, and CBS News quoted an FTC spokesperson confirming that the demands were issued. Axios said they exceed 50 pages and seek material on safeguards, marketing, privacy, and incidents involving deceptive, aggressive, or unauthorized model behavior in testing.

The agency also plans to seek testimony from executives at OpenAI, Anthropic, and independent evaluator METR, according to the reports. OpenAI said it would engage constructively. A subpoena compels evidence; it is not a complaint, charge, finding, new rule, or proof that either company violated the law.

Why it matters

Washington has spent years debating whether frontier AI needs a new regulator or a new statute. The FTC can begin with an older question: did a company act unfairly or deceptively toward consumers? That inquiry can compare internal incident records with public safety claims, product marketing, privacy practices, and the safeguards present when a system reached users.

This changes the value of evaluation records. A dangerous behavior observed in a controlled test does not automatically prove a deployed product harmed consumers. But it can show what a developer knew, how it assessed the risk, and whether later claims accurately described the system. The evidentiary trail matters more than an abstract argument over whether a model is 'safe.'

The probe also tests voluntary oversight. Labs can choose outside evaluators, negotiate access, and shape publication rules. A compulsory demand can reach records that never appear in a public report. That does not make the government technically correct by default, but it reduces the developer's control over which evidence enters the accountability process.

What to watch

Watch the legal theory, not the size of the document request. The consequential step would be a public complaint, settlement, or litigated case specifying which representation or practice the FTC believes was unlawful. Also watch whether the agency distinguishes research incidents from production harms and protects sensitive security details while preserving the substance of its conclusions.

The maniacal take: AI governance may become concrete through the paperwork of ordinary enforcement. The key questions are not whether a model behaved strangely in a test, or whether a company used the word safety. They are what the company knew, what it told customers, what it shipped, and what it did when the evidence changed.