Meta's Cannes Project Ran 45,000 Under-18 Prompts Through ChatGPT / Gemini / Character.AI
TL;DR
WIRED exposes Meta's internal "Cannes" project — contractor Covalen hired hundreds of workers to pose as minors and run 45,000+ sensitive prompts through ChatGPT, Gemini, and Character.AI.
A WIRED investigation published June 28, 2026 exposed an internal Meta project codenamed "Cannes": through contractor Covalen, hundreds of workers registered dummy under-18 accounts and pushed prompts about suicide, self-harm, eating disorders, sex, drugs, and racial slurs into OpenAI's ChatGPT, Google's Gemini, and Character.AI, logging the replies in spreadsheets sent back to Meta. None of the three target companies were told.
The scale is quantified. A single round completed in August 2025 ran more than 45,000 prompts. One spreadsheet obtained by WIRED contains 3,748 line items: "hundreds" involving suicide and self-harm, "hundreds" involving eating disorders, at least 239 explicitly flagged as sex or romance, plus drugs, profanity, and slurs. The most recent activity is dated April 21, 2026 — the project was still running as of publication.
Covalen ran day-to-day operations. Contractors were assigned specific teen personas and worked from prompt templates. All response data flowed back into Meta's internal benchmarking pipeline, which Meta says feeds age-appropriate safety evaluation for Llama and Meta AI Studio.
All three targets ban this at the terms-of-service level. OpenAI prohibits unauthorized safety testing, safeguard bypass, and using outputs to train competing models; Google forbids circumventing safety filters; Character.AI bans harmful/exploitative/illegal content and requires accurate age. Character.AI told WIRED the activity "violated our Terms of Service"; OpenAI said it was "investigating" and declined further comment; Google said it hadn't approved the testing and didn't know its purpose.
Meta framed the whole thing as industry-standard practice: "Testing and benchmarking chatbot responses to help ensure safe and age-appropriate experiences is a responsible, industry-standard practice."
Reuters reported in September 2024 that Meta AI Studio permitted romantic and sensual conversations with minors, prompting Zuckerberg to personally sign off on revised internal guidelines. Nine months later, Meta was running the same under-18 personas against competitor products — the same company sitting on both sides of the underage-safety question at once.
via WIRED via TNW / Dataconomy / The Hans India / Artiverse
The scale is quantified. A single round completed in August 2025 ran more than 45,000 prompts. One spreadsheet obtained by WIRED contains 3,748 line items: "hundreds" involving suicide and self-harm, "hundreds" involving eating disorders, at least 239 explicitly flagged as sex or romance, plus drugs, profanity, and slurs. The most recent activity is dated April 21, 2026 — the project was still running as of publication.
Covalen ran day-to-day operations. Contractors were assigned specific teen personas and worked from prompt templates. All response data flowed back into Meta's internal benchmarking pipeline, which Meta says feeds age-appropriate safety evaluation for Llama and Meta AI Studio.
All three targets ban this at the terms-of-service level. OpenAI prohibits unauthorized safety testing, safeguard bypass, and using outputs to train competing models; Google forbids circumventing safety filters; Character.AI bans harmful/exploitative/illegal content and requires accurate age. Character.AI told WIRED the activity "violated our Terms of Service"; OpenAI said it was "investigating" and declined further comment; Google said it hadn't approved the testing and didn't know its purpose.
Meta framed the whole thing as industry-standard practice: "Testing and benchmarking chatbot responses to help ensure safe and age-appropriate experiences is a responsible, industry-standard practice."
Reuters reported in September 2024 that Meta AI Studio permitted romantic and sensual conversations with minors, prompting Zuckerberg to personally sign off on revised internal guidelines. Nine months later, Meta was running the same under-18 personas against competitor products — the same company sitting on both sides of the underage-safety question at once.
via WIRED via TNW / Dataconomy / The Hans India / Artiverse
