An internal program called Project Lily uses human reviewers to analyze live ChatGPT interactions to determine if the model's tone is too sycophantic or weird. These workers earn more than $50 an hour to summarize user intent and rate four AI responses on a scale of 1 to 7. Reviewers have access to a contractor dashboard with "user memories summary" data that can expose a person's location and profession when privacy filters fail to catch unique identifiers.

OpenAI enables data collection by default for 900 million users, with the opt-out located within the Data Controls section of the settings. A 404 Media report indicates that the company lacks the clear disclaimers used by competitors Google and Anthropic for similar human review processes. The current opt-out system applies only to future conversations and does not purge previously reviewed data.