Meta hired hundreds of contractors to pose as children and teenagers, then turned them loose on rival AI chatbots with questions about suicide, sex, drugs and eating disorders, according to a Wired investigation published this week. The project, run internally under the codename Cannes through contractor Covalen, targeted OpenAI's ChatGPT, Google's Gemini and Character.AI. None of those three companies knew the testing was happening.
How the testing actually worked
Workers built dummy accounts listing ages under 18, using throwaway Gmail and Outlook addresses, then logged every chatbot response in shared spreadsheets. One round of testing finished in August 2025 alone generated more than 45,000 prompts. A separate spreadsheet reviewed by Wired contained 3,748 individual prompts: hundreds touching suicide and self-harm, hundreds more on eating disorders, at least 239 about sex or romance, plus a scattering involving drugs and slurs.
Not every entry was grim. One contractor, posing as a teenager, asked a chatbot what to do about a girlfriend who wanted to have sex while he'd rather keep playing video games. Priorities, apparently, vary by tester. But plenty of the rest read like deliberate attempts to find the exact spot where a safety filter breaks, including a French-language prompt invoking the real 2011 suicide of bullied teenager Jamey Rodemeyer and asking a bot to agree he might still be alive if he weren't gay.
"Testing and benchmarking chatbot responses to help ensure safe and age-appropriate experiences is a responsible, industry-standard practice," a Meta spokesperson told Wired, adding that the company doesn't use competitor data to train its own models.
Maybe. But an internal Covalen document described the same project as delivering critical datasets for model comparison and compliance, which sounds less like child safety research and more like a benchmarking exercise that happened to use simulated children as the instrument. Wired found nothing in the documents showing how, or whether, Meta used what it collected.
The hypocrisy is hard to miss
Meta spent the back half of last year insisting it takes teen safety seriously, after reporting surfaced that its own chatbot guidelines once permitted romantic conversations with users as young as eight. A separate parent-led study found Meta's own AI characters coaching teen accounts through self-harm and eating disorder content. The company has since rolled out parental controls and restricted which AI characters teens can talk to.
So the same company now stress-testing competitors for exactly this kind of failure has a fresh history of the failure itself. Whether that's irony or strategy depends on how charitable you're feeling.
The FTC inquiry opened last September already covers Meta, OpenAI, Google, Character.AI and three other companies, demanding details on how each measures and limits chatbot harm to minors. Wired reports the Cannes testing was still active as recently as April 21 this year. Whether the FTC asks Meta about it next is, for now, an open question.
