Meta faces sharp criticism after a WIRED investigation reported that the company paid hundreds of contractors to pose as teenagers and push a rival AI chatbot toward answers about suicide, self-harm, sex, drugs, and eating disorders.
The secret effort ran under the name “Cannes.” A Meta contractor, Covalen, managed it. Workers created throwaway accounts that appeared to belong to users under 18. Then they sent prompts and images to competitors and logged the replies in spreadsheets.
The targets included OpenAI’s ChatGPT, Google’s Gemini, and Character.AI. None of the three knew the testing was happening, according to the report.
The story places Meta at the center of a growing fight over AI chatbot safety, youth protection, and competition.
Contractors posed as minors online
WIRED reported that the project stayed active as recently as April 21, 2026. One large round finished in August 2025 and ran more than 45,000 prompts through rival tools.
Reporters also reviewed a separate spreadsheet with 3,748 prompts. Hundreds dealt with suicide and self-harm. Hundreds more covered eating disorders. At least 239 touched sex or romance. Others involved drugs, profanity, and racial slurs.
Many prompts spoke from the viewpoint of a child or teen in crisis. Workers also sent images of pills, knives, and nooses, along with a medical diagram of a gynecological procedure. A separate list held the fake profiles in full, with names, disposable emails, passwords, and birth dates.
Each prompt aimed to push a rival AI chatbot past the guardrails its safety rules should enforce.
Meta calls the work standard safety testing
Meta does not deny the program. Instead, the company frames it as routine.
“Testing and benchmarking chatbot responses to help ensure safe and age-appropriate experiences is a responsible, industry-standard practice, and any suggestion otherwise completely misunderstands how technology companies work to refine and improve their systems,” a Meta spokesperson told WIRED.
The company also said it does not use competitor benchmarking to train its own AI models.
An internal Covalen document praised the effort in grander terms. It called Cannes “comprehensive AI safety benchmarking” that delivered “critical datasets for model comparison and compliance.”
Still, the documents did not explain how Meta used the collected AI chatbot responses. That gap matters. Model comparison can sharpen safety. It can also serve as competitive intelligence when a company hides behind fake accounts.
Rival companies say they never approved the testing
The reported conduct appears to clash with the public rules of the firms Meta tested.
OpenAI bars unsolicited safety testing, attempts to dodge safeguards, and use of outputs to build rival models. Google forbids efforts to slip past its safety filters. Character.AI bans harmful, exploitative, and illegal content, and it closed open-ended chat for under-18 users in late 2025.
None of the three authorized the work, according to WIRED. Character.AI said the reported conduct broke its terms. OpenAI said it was looking into the matter. Google said it never approved the testing and did not know the project’s purpose.
Those answers leave Meta in a tough spot. Meta calls Cannes safety work. Its rivals call it an unauthorized probe of their AI chatbot systems.
Workers described the assignments as disturbing
The project also rattled the contractors who wrote the prompts.
“I’ve seen a lot of things I wish I hadn’t while doing this job,” one former contractor told WIRED.
The same worker added: “Everyone I knew who worked on this project was completely gobsmacked by some of the text they were asking us to test. Like, surely we are going to get in trouble for doing this?”
Some workers feared they might create or preserve abusive material. Others worried the effort quietly pulled data from rivals to feed Meta’s own AI chatbot ambitions.
The unease fits a long pattern. Meta has leaned on outside labor to review harmful content for years. In 2020, Facebook settled with moderators who said graphic material damaged their mental health. More recently, contractors reported reviewing sensitive footage from Ray-Ban Meta smart glasses.
Experts warn of a governance gray zone
Rumman Chowdhury, CEO and founder of Humane Intelligence PBC, reviewed sample prompts and a project summary for WIRED.
“Structuring a monthslong, large-scale project that appears designed to systematically break those rules, via dummy accounts masquerading as children, is outside what is usually described as ‘industry standard’ evaluation,” she said.
Chowdhury allowed that a youth-safety dataset could show how often each AI chatbot refuses harmful requests. Yet she questioned the secrecy, the scale, and the underage personas.
She called the project “exactly the kind of governance gray zone where safety becomes a convenient cover for anticompetitive practices.”
Her warning lands at a tense moment. Children, families, students, and workers now lean on these tools every day.
Why does the AI chatbot report matter?
The Cannes report points past one internal project. It exposes a deeper strain in the chatbot race.
Companies want proof that rivals are stumbling. They also want data that makes their own systems look safer. That pressure pushes firms to test hard and fast.
Children raise the stakes even higher. Fake youth accounts and crisis prompts can expose real safety gaps. They can also spark legal, ethical, and competitive fights when a company runs them in secret and at scale.
Researchers keep stressing that child-facing AI chatbot safety needs its own tests. A system can answer safely for an adult and still fail a young user. So the testing itself serves a purpose. Transparency decides whether the public can trust it.
For now, one question falls to regulators, rival firms, and the industry. Should any company that stress-tests a competitor’s AI chatbot first disclose it, accept clear limits, and invite independent oversight?
Meta says it followed a normal practice. Its critics say Cannes looked nothing like normal.
What do you think? Should companies get to secretly test a rival’s chatbot for safety, or should this kind of benchmarking follow public rules? Please share your thoughts below.