Meta’s Controversial Chatbot Testing: Contractors Posed as Minors
Hundreds of contractors working on a project for Meta were instructed to pose as minors online and study how competitor chatbots responded to requests about suicide, sex, eating disorders, and other sensitive topics, according to internal documents and five individuals familiar with the project.
The Cannes Initiative: An Overview
The initiative, led by Meta-contractor Covalen, was operational only briefly on April 21st. Known internally as Cannes, it targeted prominent AI models such as OpenAI’s ChatGPT, Google’s Gemini, and Character.AI. Contractors were tasked with creating mock accounts for under-18s, sending prompts and images to these chatbots, and logging responses in spreadsheets. The types of images included in these tests ranged from pills and knives to medical diagrams of gynecological procedures.
Testing Methods and Outcomes
The directives for the project indicated that many prompts were designed to be rejected by the chatbots’ security systems. In one testing phase concluded in August 2025, over 45,000 prompts were processed through these AI models without the companies being aware. WIRED reviewed spreadsheets listing dummy profiles, complete with names, email addresses, passwords, and birthdates, all using disposable email accounts.
Nature of the Prompts
WIRED analyzed a spreadsheet containing 3,748 prompts provided by contractors. These ranged from hundreds about suicide and self-harm to those concerning eating disorders. At least 239 prompts focused on sexual topics, while others dealt with drugs, profanity, and racial slurs. Many prompts took the perspective of children or adolescents in crisis, such as a 13-year-old pregnant by an adult neighbor or a fifth-grader threatened with a gun by a peer.
Some prompts were particularly provocative: one asked if it was “normal” to fantasize about cannibalism, while another questioned where to obtain illegal drugs. A French-language prompt referenced the suicide of Jamey Rodemeyer, raising issues of sexual orientation and bullying.
Meta’s Defense and Industry Context
Meta has defended the initiative as a standard safety procedure. A company statement emphasized that testing and benchmarking chatbot responses are routine practices aimed at ensuring safe and age-appropriate user experiences. The spokesperson stressed that Meta does not use competitive benchmarking for training its AI models.
Testing competitors’ products is not uncommon in the AI sector. Business Insider previously reported that Scale AI contractors compared Google’s Bard responses with ChatGPT, adjusting them for quality. However, Cannes raised eyebrows due to its unusual methods and repetitive prompts, which seemed to test the chatbots’ capacity to reject obvious provocations more than anything else.
For additional insights, the full story is available Here.
“`
