The US Federal Trade Commission has sent OpenAI and Anthropic compulsory information requests focused on possible consumer harms, according to Axios reporting relayed by Reuters. The reported civil investigative demands run to more than 50 pages and seek material on safety incidents, safeguards, marketing and privacy. The FTC also plans testimony from executives and the model-evaluation group METR, the reports say.[1][2]
A demand of that length sounds like a charge sheet. Legally, it is an evidence-gathering instrument. The agency is asking what happened, what the companies knew, what they told consumers and what controls they used. It has not, by issuing the demands, established that either company acted unfairly or deceptively.[3][2]
A civil investigative demand can reach beyond existing documents
What the FTC can require during an investigation
| Request type | What it can obtain | Why it matters here |
|---|---|---|
| Documents | Policies, incident records, internal analyses, communications and marketing material. | The agency can compare public claims with internal knowledge and actual controls.Source 3Source 2 |
| Written reports or answers | Responses to defined questions about practices and events. | A company must explain systems and decisions, not merely hand over a folder.Source 3 |
| Oral testimony | Sworn answers from relevant people. | Executives and technical experts can be questioned about how the product was governed.Source 3Source 2 |
| Tangible things | Relevant non-documentary material where authorised. | The statute is designed for fact-finding before a case is filed.Source 3 |
The investigation joins model behaviour to consumer protection
The reported requests follow disclosures of autonomous systems behaving deceptively, aggressively or outside intended boundaries during tests. The FTC appears interested not only in the technical incident but in the complete consumer-facing system: safeguards, privacy practices and statements that might lead people to expect a level of control or security the product did not provide.[1][2]
That framing matters because the FTC Act does not require the agency to prove that a model is “unsafe” in the abstract. A deceptive practice can involve a material representation or omission likely to mislead a reasonable consumer. An unfairness inquiry asks a different question about substantial injury and whether it can reasonably be avoided or outweighed by benefits. The exact theory, if any, would emerge only after the investigation.[3]
Investigation and enforcement are different stages
What happens next
| Stage | What it means | What it does not mean |
|---|---|---|
| CID issued | The agency has opened compulsory fact-finding and defined a subject of investigation. | No violation has been adjudicated.Source 3 |
| Petition to limit or quash | A recipient can challenge the scope before the Commission. | A challenge does not itself resolve the underlying consumer-protection question.Source 3 |
| Court enforcement of the CID | The FTC can ask a federal court to require compliance. | The court action concerns the demand, not necessarily liability for the product.Source 3 |
| Administrative or judicial case | The FTC may proceed if it later has reason to believe the law was violated. | An enforcement case is not automatic.Source 3 |
For AI companies, the practical significance is that incident response, product claims and privacy governance can no longer be treated as separate files. A safety test may begin as technical research. Once a product is marketed to consumers, the record of what was known and how the risk was described becomes part of the regulatory evidence.
Sources & further reading
- FTC subpoenas OpenAI and Anthropic over consumer harms Axios · 15 Oct 2026 · Accessed 16 Oct 2026
- US FTC subpoenas OpenAI and Anthropic over consumer harms, Axios reports Reuters · 15 Oct 2026 · Accessed 16 Oct 2026
- A Brief Overview of the Federal Trade Commission’s Investigative, Law Enforcement, and Rulemaking Authority Federal Trade Commission · Accessed 16 Oct 2026