OpenAI has recruited hundreds of third-party contractors to manually review real user dialogues with ChatGPT. The work forms part of project Lily and gives reviewers direct access to prompts and responses that may include personal, medical, or other confidential details.
Reviewers are tasked with briefly summarizing each user query, comparing four model-generated answers, and scoring them on a scale from one to seven. They also flag issues such as excessive sycophancy, attempts by the model to appear human, and inappropriate use of emojis.
Although account names are masked and conversations pass through a Privacy Filter, OpenAI confirmed to 404 Media that the filter does not always remove all personal data. Reviewers may additionally see blocks of user memories containing prior conversation context and approximate location information.
Contractors are hired through external firms including Crossing Hurdles and Mercor, with some specialists earning more than $50 per hour. A similar human-evaluation process is used by Anthropic for conversations with Claude.
Under OpenAI policy, chats from personal plans may be used to improve models. Users can disable this by navigating to Settings → Data Controls and turning off the “Improve the model for everyone” option. The setting applies account-wide but does not retroactively remove previously processed data. Temporary chats are excluded from training, though they may still be reviewed for abuse detection.
