Inside ‘Project Lily’: The Humans Reading Your ChatGPT Chats
Key Points:
- OpenAI employs hundreds of contractors to review real users’ ChatGPT prompts, including sensitive personal information, to improve the chatbot’s responses; users are often unaware that humans may read their conversations.
- Contractors follow detailed guidelines to rate and critique ChatGPT’s replies, focusing on alignment with user intent, tone, factual accuracy, and avoiding anthropomorphism or sycophancy, under a project codenamed “Project Lily.”
- OpenAI anonymizes prompts and uses a Privacy Filter to remove personal data before review, but sensitive information can still slip through; users can opt out of data use for model improvement, though this setting is on by default.
- The contractors, paid over $50 an hour, are hired through intermediaries like Crossing Hurdles and Mercor, reflecting a broader industry practice where human reviewers play a crucial but often hidden role in refining AI models.
- Experts warn that human review is essential for chatbot safety but raises privacy concerns, as users may mistakenly believe their interactions are private one-on-one conversations, highlighting the tension between AI advancement and user transparency.