Openai's project lily reveals human review of ChatGPT user conversations

OpenAI has disclosed its internal Project Lily, which involves human reviewers examining ChatGPT user conversations to refine the model. The project, revealed by 404 Media, includes operational guidelines, Slack records, and anonymized chat samples. Reviewers, called 'prompt reviewers,' assess the quality of responses, checking for AI-specific phrasing, overly formal tones, and inappropriate content. They are instructed not to use personal anecdotes or fabricate experiences. One reviewer noted the work is repetitive but well-paid, with hourly rates exceeding $50. OpenAI admitted that anonymization processes may fail to fully protect user data, especially in short conversations. The project does not verify factual accuracy but flags obvious errors, indicating separate teams handle safety and performance assessments. Users are often unaware their chats are reviewed, raising concerns about privacy, especially for those sharing personal information. While most chatbots collect user data by default, ChatGPT’s enterprise and educational versions disable this by default. However, past data cannot be undone, and deleted content may still be stored in anonymized datasets.