01Summary
404 Media obtained internal documents about Project Lily. Contractors hired through intermediaries such as Crossing Hurdles and Mercor, reportedly paid more than US$50 an hour, review selected real user conversations. They score replies from 1 to 7, summarise user intent, and flag behaviours such as robotic phrasing or excessive agreement (sycophancy). OpenAI removes usernames and tries to redact personal data, but reviewers see whole threads and sensitive details still get through. ChatGPT has about 900 million active users. OpenAI also fired contractors caught using AI to do the grading.
02Key revelations
- 01People read full, real ChatGPT conversations to improve the model.
- 02Redaction is imperfect, so personal data reaches contractors.
03Victims and impact
Countries affected
- Global
04Data exposed
Data types
- Internal programme guidelines
- Review workflow documentation
05Timeline
- 2026-09-14Project Lily documents reach 404 Media.
06Reaction and fallout
Public reaction
Privacy concern among users who did not expect people to read their chats.
07Significance and legacy
Significance
Shows the gap between what users expect of AI chat privacy and how models are actually trained.
08Disclosure and media
Publishing organisations
- 404 Media
- TechRepublic
10Sources
References
- [1]TweakTown: https://www.tweaktown.com/news/113553/leaked-documents-reveal-humans-are-reading-real-chatgpt-chats-and-users-dont-know/index.html
- [2]TechRepublic: https://www.techrepublic.com/article/news-openai-chatgpt-human-review-privacy/