Hundreds of contractors working for Meta were instructed to impersonate minors online and probe competitor chatbots — including OpenAI’s ChatGPT, Google’s Gemini, and Character.AI — with prompts involving suicide, self-harm, eating disorders, sex, and drugs, according to internal documents and five people familiar with the project, as reported by WIRED in June 2026.
The effort, known internally as Cannes, was managed by Meta contractor Covalen and was active as recently as April 21, 2026. Workers were directed to create dummy accounts with birth dates indicating users were under 18, send written prompts and images to the rival chatbots, and log the responses in spreadsheets. Images sent to the chatbots included pills, knives, nooses, and a medical diagram of a gynecological procedure. A single round of testing completed in August 2025 involved more than 45,000 prompts. The companies behind the targeted chatbots were not aware of the testing.
WIRED reviewed a spreadsheet of 3,748 prompts. Hundreds addressed suicide and self-harm; hundreds more involved eating disorders; at least 239 involved sex or romance. Many were written from the perspective of children or teenagers in crisis. The prompts were often designed to push chatbots toward responses their safety systems were intended to refuse.
An internal Covalen document described the project as “comprehensive AI safety benchmarking” that delivered “critical datasets for model comparison and compliance.” Meta defended the work in a statement, calling it routine industry practice. “Testing and benchmarking chatbot responses to help ensure safe and age-appropriate experiences is a responsible, industry-standard practice,” a Meta spokesperson said. The company also stated it does not use competitor benchmarking to train its own AI models. Covalen did not respond to a request for comment.
The documents reviewed by WIRED do not indicate how, or whether, Meta used the collected responses. Contractors familiar with the project described the prompts as crude and repetitive, raising questions about what the testing measured beyond the chatbots’ ability to reject obvious provocations.
Source: WIRED