Meta paid contractors to pose as suicidal teens and probe rival chatbots for failure
meta used a covert contractor army to flood competitor models with extreme prompts involving suicide and sex. the data exposes how aggressive ai benchmarking can violate user safety norms and privacy.
internal documents show meta managed a secret project called cannes that hired contractors to pose as children and test rival ai safety filters. workers entered graphic scenarios involving suicide, eating disorders, and sexual requests into models from openai, google, and character.ai. this operation aimed to gather compliance data, but it involved simulating distressed minors and sharing sensitive images. meta defends the practice as responsible benchmarking necessary for safety.