Hundreds of eyes hidden behind ChatGPT: OpenAI hires hundreds of contractors to read user conversations to train models.

Beating AI News Flash: Internal materials obtained by 404 Media show that OpenAI is using a project codenamed "Project Lily" to have hundreds of outsourced workers read real user conversations with ChatGPT in order to train and improve its models. Reviewers cannot see usernames, but may see full conversations. Some tasks also include users' "memory summaries," including what they have discussed before, where they may live, and other information.

These people first summarize what the user is trying to do, then compare multiple responses generated by ChatGPT and rate them on a scale of 1 to 7. The review focuses on whether facts are correct, whether there is excessive sycophancy toward users, and whether ChatGPT presents itself too much like a real person.

OpenAI first uses Privacy Filter to remove personal information such as names and contact details, but the company also acknowledges that this model may miss uncommon identity information and vague private information. Some sensitive content may still appear before reviewers.

Not all ChatGPT chats are read by humans. For individual Free, Plus, and Pro versions, "improve the model for everyone" is enabled by default; after users turn it off, new chats will no longer be used for training; Business, Enterprise, and Edu are not used for training by default.