Meta used hundreds of contracted workers to create fake under-18 accounts and probe the safety limits of rival AI chatbots, according to a WIRED investigation published on June 28, 2026.
The operation, internally called “Project Cannes” and managed through contractor Covalen, instructed workers to open throwaway email addresses and register accounts listing their age as under 18. They then sent prompts on sensitive topics — including suicide, self-harm, eating disorders, drugs, and sexual content — to OpenAI’s ChatGPT, Google’s Gemini, and Character.AI, logging responses in spreadsheets for later review.
A single testing round completed in August 2025 generated more than 45,000 prompts, with hundreds focused on self-harm and suicide. According to internal documents reviewed by WIRED, the project remained active as recently as April 21, 2026.
None of the three companies knew the tests were happening. Character.AI said the operation violated its terms of service. OpenAI launched an investigation into the accounts involved. Google said it had not approved the tests.
Meta defended the project. “Testing and benchmarking chatbot responses to help ensure safe and age-appropriate experiences is a responsible, industry-standard practice,” a spokesperson told WIRED.
Researchers have challenged that framing. Experts described the fake minor accounts as falling outside standard evaluation methods — a “governance gray zone,” in the words of one specialist, where safety justifications can overlap with competitive intelligence gathering.
The disclosure arrives as regulators in the United States and Europe tighten oversight of how AI systems interact with minors. Character.AI already faces litigation in the United States over alleged harms to teenagers, heightening scrutiny of the methods AI companies use when studying how rivals handle sensitive content.