The paperwork stage
There is a difference between a regulator announcing an investigation and a regulator serving the demands. The announcement, in late September, cost the labs a difficult news cycle and some legal preparation. A served civil subpoena would change the accounting: it would set deadlines, create a formal record of what was asked and what was handed over, and put named executives in line to testify.
That, according to Axios reporting carried by Reuters on 15 October, is where the FTC probe now sits — though the FTC itself has not confirmed that any demand has been sent, and the report rests on a single person familiar with the matter. The demands reportedly sent to OpenAI and Anthropic each run past 50 pages. Page count is not a measure of seriousness by itself, but the reported categories are informative: safety incidents, safeguards, marketing materials and privacy policies, requested together, in one demand. Read as a set, they would describe a comparison the agency wants to run — what happened, what protections were in place, what customers were told, and how their data was handled while all of that was going on.
Why marketing is in the same envelope as incidents
A safety investigation that only wanted incident logs would ask for incident logs. Including marketing materials suggests the inquiry may test the gap between conduct and claims, which is the ground the FTC can actually litigate on. It does not prove that intent, and the reported requests establish no misleading conduct by either company. The agency does not need an AI statute to act on a misleading safety claim or on consumer data handled carelessly; its existing authority over unfair and deceptive practices covers both, and Reuters notes it has used that authority before over failures to secure consumer data.
The reported testimony plan widens the same point. Executives from the labs are reportedly expected to be compelled to testify, and so, notably, is METR, the independent group that evaluates frontier models. That is a plan reported by Axios, not subpoenas already issued, and the reporting does not establish that METR has received any demand. Bringing the evaluator into the room would suggest the agency wants the test record explained by the people who ran the tests, not only by the companies that commissioned them. If the outside scores and the inside incident reports tell different stories, that hearing is where the difference would surface.
What this does not yet establish
A subpoena is a request for evidence, not a finding. Nothing in the reporting says the FTC has concluded that either company misled anyone, and neither lab has responded publicly yet. The source for the sending of the demands is a single person familiar with the matter, via Axios; Reuters did not independently verify the sending, and the FTC itself has not confirmed the documents. Readers should hold it at that strength: widely carried, plausible in sequence after the September announcement, and still one anonymous source deep on the newest detail.
If the demands have been served, the practical effect would arrive before any legal conclusion does. Both labs would have to decide, document by document, what their incident reports and internal evaluations say when read by an enforcement lawyer rather than a researcher. That decision would shape the next round of public safety reporting across the industry, whether or not this probe ends in a case. This piece follows our [2 October coverage](https://alfw.ai/article/ftc-openai-anthropic-probe/) of the probe's opening, when the demands were still expected rather than sent.
Sources
- [1] Reuters — US FTC subpoenas OpenAI and Anthropic over consumer harms, Axios reports (15 October 2026; one wire report — the sending rests on a single source, a person familiar with the matter, and was not independently verified by Reuters)Read source
- [2] ALFW, 2 October 2026 — The FTC Probe Tests What AI Labs Said Against What Their Models Did (background: the probe's opening, when the demands were still expected)Read source