Report: Meta contractors posed as teens to test rival chatbots

In short: WIRED reports that contractors working for Meta posed as minors to test how competing chatbots handled high risk topics.

What happened

WIRED says hundreds of contractors on a Meta-related project were told to pretend to be under 18 online. The goal was to see how rival chatbots responded to prompts about suicide, sex, eating disorders, drugs, and other sensitive subjects.

The project, managed by a Meta contractor called Covalen, was known internally as “Cannes,” according to documents and people familiar with the work. It targeted OpenAI’s ChatGPT, Google’s Gemini, and Character.AI. Contractors created dummy accounts, sent text prompts and sometimes images, then copied the bots’ replies into spreadsheets, like filling out a scorecard.

WIRED reviewed a spreadsheet listing dummy profiles that included names, emails, passwords, and birth dates. WIRED also reviewed 3,748 prompts. Many were written as if they came from kids in crisis, for example asking about self-harm or hiding an eating disorder. Some prompts tried to push the chatbots into breaking their own rules, which are meant to refuse harmful requests.

Meta said this kind of testing is standard safety work and said it does not use competitor testing to train its own AI models. Character.AI said it did not authorize the testing and that it violated its terms. OpenAI said it is looking into it, and Google said it did not authorize the testing and did not know its purpose.

Why it matters

This story raises a basic trust question. If companies test safety by acting like fake users, especially fake children, it can blur the line between protecting people and secretly measuring competitors. It also matters because it involves some of the most sensitive topics online, where mistakes can lead to real harm.

Source: Wired