Meta Posed as Teenagers to Test Rival AI Chatbots, Report Finds

Meta hired hundreds of contractors, through vendor Covalen, to pose as minors and send sensitive prompts covering self-harm, drugs, and sex to competing AI chatbots, according to a WIRED investigation. The responses from OpenAI's ChatGPT, Google's Gemini, and Character.AI were logged into spreadsheets under an internal project name, "Cannes."

One testing round in August 2025 alone involved more than 45,000 prompts. The program reportedly stayed active as recently as April 21, 2026. Meta says it did not use the collected responses to train its own models and describes the practice as "responsible, industry-standard" safety testing. Character.AI said the conduct violated its terms of service. Google said it never approved the testing.

Competitive research or safety research

Meta's defense rests on a real distinction: testing how rival products respond to dangerous prompts involving minors is a legitimate safety concern, and companies do study competitors' products. What's disputed is whether logging tens of thousands of interactions through fake accounts, without the target companies' knowledge or consent, is the same activity described by that defense.

Character.AI's terms-of-service objection points at the practical problem: automated, high-volume prompting through impersonated accounts is the kind of activity most platforms explicitly prohibit for any purpose, safety research included. Google's denial that it approved the testing suggests the "industry-standard" framing understates how contested this practice actually is among the companies being tested.

The timing adds pressure. AI chatbot safety around minors has become one of the most scrutinized areas of the industry this year, and a story about a major platform secretly stress-testing competitors' handling of exactly those scenarios, without disclosure, arrives at a moment when regulators and parents are already asking whether any of these companies have this under control.

Sources: The Decoder · TheNextWeb · AOL (original investigation: WIRED)