The Federal Trade Commission has demanded documents from OpenAI and Anthropic about how the two artificial intelligence developers handle safety failures and treat their users, according to people familiar with the matter. The requests, reported by Axios on October 15, run about 50 pages for each company and cover safety incidents, the protections built into their models, and consumer protection practices.
A civil investigative demand is not a finding of wrongdoing. It is an order for records that the commission can issue before it decides whether to bring a case, and regulators use them when they need internal documents, evaluation results and complaint logs that a company has no reason to hand over on its own. The people familiar with the matter said the commission is concentrating on whether either company may have broken consumer protection law.
The demands arrive at an awkward moment for both. OpenAI was sued on October 14 by the attorneys general of eight states, each filing in its own courts, over claims that ChatGPT misled users about its safety and put minors at risk. Iowa Attorney General Brenna Bird said the company described its chatbot as safe while it produced dangerous medical advice, encouraged self-harm and suicide, and presented itself as a false friend to children. OpenAI said in response that the product carries age-graded protections.
Both companies have made safety spending part of how they describe themselves, to customers and to regulators alike. The distance between a claim made in public and the records kept in private is what a consumer protection inquiry is built to measure, according to lawyers who work on such matters.
Consumer protection authority is narrower than it sounds and broader than companies sometimes assume. The commission does not need a new statute to examine whether a product description overstates what a system can do, whether a company discloses how it uses customer data, or whether the warnings it gives are adequate for the people most likely to be harmed. It can ask for the material that shows what executives believed at the time.
The unusual part of the requests is the volume. Demands of roughly 50 pages suggest a document review that stretches well past press statements and published safety policies, into incident reports, red-team exercises, escalation logs and the internal debates that followed. Companies that have organized their public accounts around safety research may find the internal record harder to reconcile.
Anthropic’s own disclosure to prospective investors on October 15 put safety spending next to its commercial results. The company told investors it booked $16.8 billion of revenue in the third quarter, up from $11.6 billion in the second and $4.7 billion in the first, and said it was profitable on an adjusted basis with positive cash flow. It has told investors it may begin its initial public offering roadshow as early as the week of November 9.
That combination, a safety-first public identity attached to a fast-growing business, is what the inquiry will test. A company heading toward a listing has an added reason to resolve regulatory questions early, and an added exposure if it cannot.
OpenAI’s difficulty is more immediate. The state suits filed on October 14 are the first coordinated legal challenge to the company over child safety, and they name the same conduct the commission is now asking about. Anything OpenAI produces for the states becomes material for the federal inquiry, and any inconsistency between the two proceedings will be difficult to manage in public.
For the industry, the significance is procedural. Washington has spent years debating how to govern artificial intelligence without passing a broad new law. The commission’s approach applies rules written for advertising and product safety to systems that were not imagined when those rules were drafted, and it does so through requests for documents rather than through legislation.
The practical consequences for both companies are cost and attention. Responding to a demand consumes legal resources, pulls engineers off product work to collect records, and creates a paper trail that later plaintiffs can seek. Companies that have faced similar inquiries say the internal review alone often changes how they write the policies they publish.
What the commission does with the material is not yet clear. It may close the inquiry without action, issue a warning, negotiate a consent order with compliance obligations, or sue. Each outcome carries a different cost, and none of them is quick. Cases built on documentary evidence typically take a year or more to resolve.
There is a further question about how the two companies will read the same request. They compete for the same enterprise customers and the same investors, and a regulatory finding against one would shape how buyers evaluate the other. Their lawyers will compare notes only to the extent the commission’s rules permit.
For now, the demands are the start of a process rather than a verdict. The commission has asked two companies that built their reputations on safety to prove, from their own files, what that reputation was worth. The answer will arrive slowly, most likely in the form of documents neither company was planning to publish.
