Meta directed hundreds of contractors to impersonate minors and feed harmful prompts into rival A.I. chatbots, according to a Wired investigation published Monday, in a covert benchmarking operation that was active inside the company as recently as April 21, 2026. The project, internally code-named Cannes and managed through the contractor Covalen, treated OpenAI, Google and Character.AI as test subjects without their knowledge.

The scale is the story. A single round of testing in August 2025 pushed more than 45,000 prompts through ChatGPT, Gemini and Character.AI from accounts presenting themselves as children. Wired reviewed one spreadsheet of 3,748 prompts: hundreds concerned suicide and self-harm, hundreds covered eating disorders, and at least 239 involved sex or romance. A separate file logged the dummy infrastructure, fake names, throwaway emails, passwords and birth dates engineered to clear age gates.

An internal Covalen document described the operation as “comprehensive A.I. safety benchmarking” producing “critical datasets for model comparison and compliance.” A Meta spokesperson didn’t dispute the work, saying “testing and benchmarking chatbot responses to help ensure safe and age-appropriate experiences is a responsible, industry-standard practice” and adding that “the company does not use competitor benchmarking to train its own models.” Covalen didn’t respond to Wired. OpenAI said it was looking into the matter. Character.AI, which closed open-ended chat to under-18 users in late 2025, said the conduct violated “our Terms of Service.” All three targeted platforms prohibit exactly this kind of testing.

Two lawyers who reviewed examples for Wired said the material didn’t cross into illegal obscenity or child sexual abuse material. Rumman Chowdhury, chief executive of Humane Intelligence, was less forgiving. She called the use of “dummy accounts masquerading as children” something “outside what is usually described as industry standard evaluation,” and described the broader practice as a “governance gray zone where safety becomes a convenient cover for anticompetitive practices.”

The timing was almost too clean. On the same day the Wired story landed, the House passed the Kids Internet and Digital Safety Act by 267-117, attaching new A.I. chatbot safeguards for minors. Representative Brett Guthrie, the Energy and Commerce Committee chairman, called it “the most comprehensive, impactful children’s online safety package Congress has considered.” Senator Maria Cantwell, looking at the same bill from the other chamber, said the House version “has gutted many of the key provisions” of an earlier Senate measure.

Two decades after the 1998 COPPA framework, Washington is again writing rules for child safety online while the largest platform under scrutiny was already running its own shadow audit of competitors, calling that audit safety work, and meeting no one in the room who could verify the claim.

Sources