What 404 Media found
OpenAI is paying hundreds of contractors to read real ChatGPT prompts and rate the chatbot’s answers, 404 Media reported on 14 September. The work runs under the internal codename Project Lily, which appears in training materials the outlet reviewed. The documents do not say which model the ratings help train.
404 Media said it examined instruction guides, Slack channels, rating systems and examples of real user prompts, with some material marked “Confidential & Proprietary”. One contractor said the job paid more than $50 an hour. Workers are recruited through a firm called Crossing Hurdles and paid through Mercor, according to the report.
What reviewers see
Reviewers do not see usernames. But prompts can arrive with a “user memories summary” drawn from the person’s earlier use of ChatGPT, which sometimes includes where they are, 404 Media reported.
OpenAI says personal information is stripped out before prompts reach reviewers, by a model it calls Privacy Filter. Its own documentation is candid about the limits: “Like all models, Privacy Filter can make mistakes. It can miss uncommon identifiers or ambiguous private references.”
On by default
Whether a conversation can be used this way depends on a setting called “Improve the model for everyone”. It is switched on by default for Free, Plus and Pro accounts and has to be turned off by the user, while it is off by default on Enterprise, Business and Edu plans, according to 404 Media.
OpenAI declined to say whether users had been explicitly told that people, not only software, may read their chats. After the story ran, it pointed 404 Media to a section of its website saying humans review content to “improve model performance”.
The European question
In Europe, the disclosure may matter as much as the practice. The Next Web points to the EU Court of Justice’s September 2025 ruling in EDPS v SRB, which it reads as requiring a company to tell people at the point of collection that their data will go to third parties, whether or not those third parties can identify them. It said OpenAI had not specified where it disclosed the contractor review to users.
Why it matters now
Human feedback is how chatbots are tuned, and OpenAI’s own website acknowledges that people review content. What 404 Media adds is the scale, the use of outside staffing firms, and the memory summaries that travel with each prompt.
It also lands in a week when data handling is under pressure across the sector. Mathematician Andreas Thom has accused OpenAI of misleading him about his private chats, and large enterprise customers have restricted their use of Anthropic’s models over how long it keeps usage logs.
What to watch
Whether a European data protection authority opens an inquiry, and whether OpenAI changes the default for consumer accounts or adds a clearer notice that contractors may read conversations.
