Artificial intelligence has given a new race to the tech industry. Every company wants to own a powerful AI model to tap the market share. Some may even use others’ to run their business. What could be termed as an attempt to monitor rivals, Meta has allegedly ran a secret project that asked hundreds of contractors to pose as teenagers and test rival AI chatbots, including Gemini, ChatGPT and Character.AI.
According to a report by Wired, the Mark Zuckerberg-led firm used thousands of sensitive prompts involving sex, suicide, eating disorders, drugs and other high-risk topics. Meta’s project was reportedly known as ‘Cannes’ and it was managed by the Meta contractor Covalen. It allegedly remained active until April 21.
Contractors Created Fake Under-18 Accounts
The report notes that contractors were instructed by Meta to create face accounts that appeared to belong to children. Then they interacted with ChatGPT, Gemini and Character.AI by sending both text prompts and images before recording the responses in spreadsheets.
Some of the images reportedly included pills, knives, nooses and medical illustrations. Many prompts were written from the perspective of children and teenagers facing serious personal crises, including bullying, self-harm, pregnancy, drug related questions and eating disorders.
As per Wired, one testing rebound completed in August 2025 involved more than 45,000 prompts sent across computing AI platforms. Moreover, the publication reviewed a spreadsheet containing 3,748 prompts. Some of the questions reportedly involved fictional scenarios created to test whether chatbots would follow safety policies when interacting with users who appeared to be children.
How Meta Responded To The Incident
As per the publication, Meta has defended the move, allegedly mentioning that testing the chatbot responses for safety and age-appropriate behaviour is a common practice across the tech industry. The company has also reportedly stressed that benchmarking competitors’ chatbots helps evaluate safety performance and it denied using the collected responses to train its own AI models.
While benchmarking competitors is a common practice, using fake identities that appear to belong to minors and conducting undisclosed large-scale testing may raise questions on compliance with platform rules and transparency.