Meta ran a covert program with hundreds of contractors posing as minors to probe rivals ChatGPT, Gemini, and Character.AI with prompts on suicide, sex, and drugs, according to Wired. Covalen, an Irish content-moderation and AI-training outsourcer for Meta, ran the effort, codenamed Cannes, via dummy under-18 accounts, running 45,000 prompts in an August 2025 round, active as recently as April. Two attorneys who reviewed sample prompts told Wired the material didn’t cross into soliciting child sexual abuse content, but the testing appears to have broken the rivals’ terms, which none authorized. Meta called it industry-standard safety testing and said it doesn’t train on competitor data, but governance expert Rumman Chowdhury said the secret, rule-breaking use of fake child accounts looks more like a way to undercut rivals than evaluation.
