Dramatic investigation: OpenAI employees read your chats with the bot

An exclusive investigation by 404Media reveals that OpenAI employs hundreds of contractors at high wages to read and rate personal conversations of ChatGPT users, including medical information and trade secrets, as part of a secret project called 'Project Lily'. The company admits its automatic filtering is not immune to errors.

An exclusive investigation by 404Media reveals that OpenAI employs hundreds of contractors at high wages (over $50 per hour) to read and rate personal conversations of ChatGPT users, as part of a secret project called 'Project Lily'. The exposed information includes medical records, trade secrets, and sensitive code files submitted for model training. Internal documents include training guides and communication channels where workers select real conversations to read. Although usernames are hidden, the prompts themselves contain identifying details, geographic locations, and memory summaries. The official goal is to improve model behavior. The company admits its automatic privacy filter is not immune to errors and may miss unusual identifiers. The investigation notes that this practice is common in the industry, and that both Anthropic and Google employ a similar approach. The article includes a guide to disabling the default setting that allows exposure.