Overview
- Reports published Monday and Tuesday revealed an internal OpenAI program codenamed Project Lily that uses hundreds of contractors to read, summarize and rate real ChatGPT conversations to improve response quality.
- Reviewers sometimes see full chat transcripts and a 'user memories summary' that can include location, job and other personal context because OpenAI’s automated Privacy Filter can fail to remove identifiers.
- OpenAI says it withholds usernames and runs chats through the Privacy Filter but confirms the filter can miss sensitive details and that humans may review some content.
- Users of Free, Plus and Pro accounts are opted in by default to data use for model improvement and can toggle off the 'Improve the model for everyone' setting to stop future training use, but that setting does not remove past conversations from training pipelines.
- Contractors are hired through third-party staffing platforms reportedly called Crossing Hurdles and Mercor, paid more than $50 an hour, and score multiple AI outputs on a 1-to-7 scale while regulators and privacy advocates weigh oversight given ChatGPT’s large user base.