OpenAI's Project Lily Revealed: Humans Review ChatGPT Chats, Raising Privacy Concerns

A 404 Media report reveals that OpenAI's internal Project Lily uses humans to review ChatGPT user chats. These workers, called "prompt reviewers," check reply quality and flag issues like AI-speak, lecturing tone, or flattery. Replies cannot use human-like phrasing or make up personal experiences.

Reviewers earn over $50 per hour, but describe the work as repetitive, with guidelines that keep changing and often contradict each other. Chats are anonymized, but OpenAI admits filters can still miss personal data, especially in short conversations. Reviewers also see a "user memory summary" with questions, interests, and sometimes location info.

Project Lily only flags obvious errors, not factual accuracy. Deleted chats cannot be fully erased since data may already be collected. OpenAI later updated its help page to add opt-out options, but still does not mention human review of chats. Google Gemini and Anthropic disclose human review; Perplexity remains unclear.