Meta contractors posed as teens to test rival chatbots on suicide, sex, and drugs
Tags AI · Consumer
Hundreds of contractors working on a Meta project pretended to be minors in conversations with competing AI chatbots including Google Gemini and OpenAI's ChatGPT to test how the systems would respond to high-risk subjects including suicide, sex, and drugs. WIRED's investigation found that Meta's testing program involved systematic probing of competitor products rather than testing its own AI safety systems. The contractors engaged in extended conversations designed to elicit harmful responses from the rival chatbots. The revelation raises questions about competitive intelligence practices in the AI industry and the ethics of testing competitor products without disclosure.
Technical significance
This story reveals the lengths to which AI companies will go to evaluate competitor safety systems, and raises questions about the reliability of AI safety testing when conducted by commercial competitors rather than independent auditors. For AI safety professionals, it highlights the need for standardized, independent red-teaming frameworks. For regulators, it may prompt new guidelines around competitive testing practices in the AI industry.