OpenAI’s internal Project Lily reveals human reviewers assess ChatGPT chats for tone and relevance, earning over $50 hourly. While anonymized, privacy risks persist due to included user memory summaries containing location and interests. Most users opt into data collection by default, with deleted chats potentially retained in datasets. This human-in-the-loop approach, also used by Google and Anthropic, underscores that human review remains essential for AI model optimization despite algorithmic advances.