Report reveals human access to ChatGPT chats in large review program

OpenAI is reportedly paying hundreds of contractors to read real ChatGPT conversations and evaluate the chatbot’s responses under an internal project called Lily. Reviewers do not see usernames, but they do see user memory summaries that can reveal past usage and approximate location. OpenAI says a Privacy Filter model removes personal data, though the company acknowledges it can miss identifiers or under‑redact information. When asked where OpenAI informs users that humans may read their chats, the company did not provide a clear answer. Europe’s top court ruled last year that companies must inform users at the moment data is collected, regardless of whether recipients can identify individuals. Italy has already fined OpenAI EUR 15 million for inadequate legal basis and ordered public information campaigns. The default setting that improves the model is switched on for most consumer accounts, meaning their chats may be reviewed unless they disable it. Meanwhile, OpenAI markets stronger privacy guarantees to enterprise customers, including zero data retention. Contractors say users likely have no idea humans are analyzing their conversations, even as OpenAI encourages them to link bank accounts.

Read the full story on TNW →