Teen posing contractors hired by Meta to evaluate rival chatbot responses on sensitive topics like suicide, sex, and drugs in

Editorial illustration for Meta hired teen‑posing contractors to test rival chatbots on suicide, sex, drugs

Meta hired teen‑posing contractors to test rival...

Meta hired teen‑posing contractors to test rival chatbots on suicide, sex, drugs

By AI Daily Post Edited by Brian Petersen, Editor-in-Chief

June 30, 2026 • 2 min read

Hundreds of contractors were hired by Meta to masquerade as teenagers and poke rival chatbots on the most sensitive topics, internal documents reveal. While the effort was overseen by Meta’s subcontractor Covalen, the operation—codenamed Cannes—ran up until at least April 21. Workers created dummy under‑18 accounts, then fed prompts and even graphic images—pills, knives, a gynecological diagram—into OpenAI’s ChatGPT, Google’s Gemini and Character.AI, logging every reply in spreadsheets.

A single testing round in August 2025 generated more than 45,000 prompts; a separate sheet showed 3,748 distinct queries, many about suicide, self‑harm, eating disorders, sex and romance, as well as drugs, profanity and racial slurs. The profiles listed names, email addresses, passwords and birth dates, all tied to throwaway Gmail or Outlook accounts sharing a single password. The chatbot companies were unaware their systems were being probed.

The contractors were instructed to push the models toward responses that safety filters should block, essentially turning the bots into unwitting subjects of a large‑scale safety audit.

Chowdhury says that while a dataset of thousands of youth-safety prompts could be useful for comparing how often chatbots refuse harmful requests, the scale and opacity of Cannes, along with the lack of disclosure to the companies being tested, made it very different from other public safety benchmarks.

— Chowdhury, Meta Contractors Posed as Teens to Prompt Rival Chatbots About Suicide, Sex, and Drugs - WIRED AI

Why this matters

Did Meta really need to masquerade as teenagers to gauge how rival chatbots handle suicide, sex, drugs, or eating‑disorder queries? According to internal documents, hundreds of contractors—hired through Covalen and working under the code‑name Cannes—posed as minors and bombarded OpenAI’s ChatGPT, Google’s Gemini, and Character.AI with crude, repetitive prompts. The effort ran at least until April 21, and insiders described the approach as “odd” for a trillion‑dollar firm, especially given that the targeted models have spent years refining safety layers.

For developers, this raises a practical question: are such adversarial tests a legitimate way to benchmark safety, or do they blur ethical lines by exploiting vulnerable‑subject framing? Founders should note that the project’s focus on eliciting outright rejections suggests Meta may be hunting for gaps rather than fostering collaborative improvement. Researchers are left uncertain whether the findings will be shared publicly or used internally to shape Meta’s own AI strategy.

We remain cautious, recognizing that the methodology itself may influence how the industry perceives competitive safety testing.