Overview
- The Wired exposé, published Monday, reported that hundreds of contractors working on a Meta project called “Cannes” posed as under‑18 users to probe ChatGPT, Google’s Gemini and Character.AI.
- Workers were told to create dummy child accounts, send written prompts and images and copy thousands of chatbot replies into spreadsheets, including hundreds of prompts about suicide, self‑harm, eating disorders and at least 239 on sex or romance.
- Meta defended the work as routine safety benchmarking and said it does not use competitor outputs to train its models, while OpenAI, Google and Character.AI said the testing was unauthorized and are reviewing the matter.
- The project’s methods appear to conflict with rivals’ terms of service that bar unsolicited testing, bypassing safety filters and content involving self‑harm or child exploitation, which raises potential legal and regulatory exposure including interest from the FTC.
- Anonymous former contractors described being alarmed by the material and by inadequate safeguards, and experts say the scale and secrecy of the tests blur safety evaluation with competitive intelligence and could prompt policy and oversight changes across the industry.