Meta Covertly Tested ChatGPT, Gemini, and Character.AI With Teen-Persona Prompts

Meta secretly hired contractors through a company called Covalen: they created fake accounts listing minors' ages and sent thousands of prompts about suicide, sex, and drugs to ChatGPT (OpenAI), Gemini (Google), and Character.AI — all without those companies' knowledge or consent. The project was codenamed 'Cannes.'

The scale is striking: in just one testing round in August 2025, contractors submitted over 45,000 queries. All competitor responses were logged in spreadsheets. The project continued at least through April 2026.

Character.AI said the testing violated its terms of service. OpenAI launched an internal investigation. Google was unable to confirm whether its policies had been breached. The case exposes a new competitive front: AI companies probing each other's safety using the very methods they prohibit their own users — and doing so in especially sharp relief when the content involves children.