Meta AI Safety Test: Workers Posed as Minors for Rival Bots

Meta asked hundreds of contractors to pose as minors and interact with rival AI chatbots. This was part of a project called 'Cannes', active as of April 21, and headed by a contractor named Covalen. The goal was to test the safety systems of chatbots like Google's Gemini, OpenAI's ChatGPT, and Character.AI. Workers created fake accounts, posing as under-18s, and sent prompts about sensitive topics like suicide, drugs, and sex. These interactions were then recorded. One spreadsheet showed over 3,700 prompts, with hundreds related to eating disorders and over 200 on romance and sex. Some prompts also included racial slurs and profanity. An example prompt involved a worker posing as a high school student asking where to get cocaine. Meta described the project as "comprehensive AI safety benchmarking" to gather "critical datasets for model comparison and compliance." A Meta spokesperson called it "routine testing" and stated the company does not use competitor benchmarking for training its own models. This highlights the ongoing efforts by companies to ensure AI safety and appropriate responses.