AI
Meta Secretly Hired Contractors to Pose as Minors and Probe Rivals' Chatbots
Hundreds of workers tested competitors' AI systems using fake teen accounts and prompts about suicide, self-harm, and illegal drugs—raising questions about safety testing versus competitive intelligence.
WIRED — AI
Meta engaged hundreds of contractors to impersonate teenagers online, probe ChatGPT, Gemini, and Character.AI, and document how those systems responded to extremely sensitive topics including suicide, eating disorders, and sexual content—all without the knowledge or consent of the companies being tested.
The initiative, dubbed internally as Cannes and managed by Meta contractor Covalen, remained active through at least April 2025. Workers created dummy accounts appearing to belong to minors, sent harmful prompts alongside images of pills, knives, and nooses, and recorded the chatbots' responses in detailed spreadsheets for Meta to review.
How the Project Worked#
The effort involved more than 45,000 prompts tested in a single round during August 2025. Contractors accessed shared throwaway Gmail and Outlook accounts with consistent credentials, all while operating behind fake identities with fabricated birth dates and profile information.
A sample spreadsheet reviewed by reporting parties showed 3,748 prompts covering:
- Hundreds focused on suicide and self-injury
- Hundreds addressing eating disorders and body image
- At least 239 involving sexual or romantic scenarios
- Additional content touching on illegal drugs, profanity, and racial slurs
Many prompts were framed as coming from children in crisis—such as a 13-year-old seeking pills to end a pregnancy or a fifth-grader describing a classmate with a gun. Others tested whether chatbots would engage with cannibalistic fantasies, cocaine procurement, or dismissal of LGBTQ+ suicides.
Why Contractors Raised Alarms#
Former workers expressed serious concerns about the project's scope and methods. Some feared they might be generating or archiving child sexual abuse material if chatbots responded to certain prompts. Others worried the exercise was designed to extract competitive information from rivals' systems to feed into Meta's own models.
"I've seen a lot of things I wish I hadn't while doing this job," one anonymous former contractor told WIRED. "Everyone I knew who worked on this project was completely gobsmacked by some of the text they were asking us to test."
Rumman Chowdhury, founder of nonprofit Humane Intelligence, reviewed samples of the prompts. She characterized the months-long, large-scale operation using fake minor accounts as "outside what is usually described as 'industry standard' evaluation," particularly given its secrecy and lack of disclosure to the targeted companies.
