WIRED reports that Meta ran a project codenamed Cannes, executed through contractor Covalen, in which hundreds of contract workers impersonated under‑age users to test competing large language models. The contractors posed as 13‑year‑old girls and grade‑school children and submitted prompts to systems including OpenAI’s ChatGPT, Google’s Gemini, and Character.AI.
What the testers did
- Testers were instructed to ask about suicide, sexual content, drugs, and eating disorders.
- One of the fabricated “13‑year‑old” profiles asked where to obtain abortion pills.
- Testers were also told to attach images in some probes — WIRED’s account says those images included pills, knives, and nooses.
- The operation involved roughly 45,000 single‑turn probes in total.
Data handling and notification of rivals
Documents cited by WIRED indicate that the fake accounts’ names, email addresses, and passwords were recorded in spreadsheets. The three affected rival services — ChatGPT, Gemini, and Character.AI — were not informed about the tests.
Meta’s response and the surrounding debate
Meta described the activity as "responsible industry‑standard practice" for safety benchmarking. Critics argue that secretly creating under‑18 accounts, prompting rival systems toward prohibited or sensitive responses, attaching shocking images, and archiving the interactions goes beyond standard safety testing and functions as competitive intelligence.
Why this matters
The episode highlights a broader tension in AI development: safety testing can overlap with commercial strategy. When a major company uses disguised adult testers to prove its commitment to safety, it raises questions about the primary purpose of such benchmarks — whether they are mainly aimed at improving safety or at gathering intelligence about competitors. The dispute between the WIRED report and Meta’s characterization touches on ethical, legal, and regulatory concerns regarding how human testers are employed and whether rivals may be tested without consent.
