OpenAI employs hundreds of contractors to read real ChatGPT user prompts and rate chatbot responses as part of 'Project Lily,' despite most users being unaware their conversations may be reviewed by humans. The prompts sometimes contain sensitive personal information, presenting privacy risks, though OpenAI attempts to filter such details. This human review process is a critical but overlooked component of model improvement, distinct from the company's public safety measures.