OpenAI used outside reviewers to read some real ChatGPT chats for training, report says

Internal materials obtained by 404 Media show that OpenAI has been running a project code-named Project Lily, under which hundreds of outsourced workers review real conversations between users and ChatGPT to help train and improve the model. Reviewers reportedly cannot see usernames, but they may be shown full chat logs. In some cases, they also receive a user “memory summary” that can include what the user discussed previously and where the person may live. The workers first summarize what the user is trying to do, then compare multiple ChatGPT responses and score them on a scale from 1 to 7. The review criteria include factual accuracy, whether the model is overly agreeable, and whether ChatGPT presents itself too much like a human. OpenAI uses a Privacy Filter to remove names, contact details and other personal information before the material is reviewed, but the company has acknowledged that the filter may miss unusual identifying details and vague private information. Not every ChatGPT conversation is read by humans. Free, Plus and Pro accounts are opted in by default for model improvement unless users turn that setting off, while Business, Enterprise and Edu are excluded from training by default.