OpenAI employs hundreds of contractors for a project codenamed Project Lily to manually review ChatGPT conversation logs. These reviewers, known as prompt reviewers, assess the quality of responses based on specific guidelines, such as avoiding AI-speak, patronizing tones, and anthropomorphizing.
Reviewers receive pay exceeding $50 per hour for this task. The process involves evaluating anonymized chat data to improve model performance, though the guidelines provided to contractors are reportedly subject to frequent changes and inconsistencies.
While OpenAI applies an anonymization process to the logs, the company acknowledges that personal information may still be visible to reviewers, particularly in shorter conversations. The data provided to reviewers includes user memory summaries, which contain context, interests, and potentially location data.
Project Lily focuses on response quality rather than factual accuracy or safety checks, which are handled by separate teams. OpenAI notes that users can opt out of data collection for model training, although this setting does not apply retroactively to existing chat logs.
The practice of human review for AI training is common among major developers. Google Gemini and Anthropic explicitly disclose that human reviewers may access saved chats in their privacy documentation, while Perplexity's policy on human access remains unclear.
