OpenAI contractors are reading real ChatGPT chats to train its models

A ChatGPT conversation is not always a conversation between one person and a model. Hundreds of contractors are reading real user prompts so that OpenAI can train the chatbot on how it answered them.

What 404 Media found

On September 14, 2026, 404 Media reported, based on internal documents and real prompts it obtained, that OpenAI runs a human review program codenamed Project Lily. The contractors rate ChatGPT’s answers to real user prompts on a scale of 1 to 7. Part of the goal, according to the report, is to make the chatbot less sycophantic.

One worker told 404 Media they were recruited through Crossing Hurdles, paid through Mercor and earn more than $50 an hour. Asked whether users know that people read their chats, one contractor said: “No.”

What reviewers can see

Usernames are hidden from the reviewers. The review dashboard, however, sometimes includes a summary of the user’s ChatGPT memories, which can hint at where someone may live and what they have used ChatGPT for.

OpenAI says its Privacy Filter model strips personal information before prompts reach reviewers. OpenAI’s own documentation for that model says it does not catch everything.

How common it is, and the opt-out

Anthropic confirmed to 404 Media that it also uses human review, for users who have switched on its model-improvement setting.

In ChatGPT, the setting that controls training on conversations is under Settings, Data Controls, “Improve the model for everyone”. Turning it off only applies to future conversations. Temporary Chats are not a complete exception either: they may still be reviewed to monitor for abuse.

The practical takeaway is narrow but concrete. Anything typed into ChatGPT with model improvement switched on can end up in front of a human reviewer, sometimes together with a summary of the account’s memories, even though the reviewer does not see a name.