OpenAI contractors grade real ChatGPT chats, 404 Media says

404 Media reports that OpenAI is running a contractor review pipeline — internal name Project Lily — in which outside workers read real ChatGPT prompts and conversations, summarize user intent, and score model replies to train less sycophantic, less anthropomorphic behavior. Usernames are stripped; 404 says sensitive personal detail can still appear, and OpenAI has acknowledged filters can miss things.

What OpenAI will say on the record (help docs, not a Lily confirmation): Authorized OpenAI staff and trusted service providers may access user content to improve model performance unless you opt out. Consumer training use depends on settings. OpenAI also says it may send select content to trusted providers for data annotation.

Default / opt-out (the product control): In ChatGPT, “Improve the model for everyone” sits under Settings → Data Controls. Turn it off to stop new chats from training models, per OpenAI. The control is manual — not off by default for consumer accounts that leave the setting on. Temporary Chat is OpenAI’s separate path that is not used for training.

What this product is: A consumer ChatGPT privacy/control story. Humans in the loop are part of how the assistant improves when training is enabled — not a rumor that only scrapes the public web.

Limits: No OpenAI primary naming Project Lily or publishing the contractor rubric. Treat 404’s operational detail as reported, not company-confirmed. Business / API / Enterprise training defaults differ (opt-in for business offerings per OpenAI).

Risk: Users treat the empty chat box as private. A default-on training toggle plus human annotation is a risk product for anyone pasting medical, legal, or intimate content. Under-13 and EU fights are adjacent desks; this piece is the opt-out and the human read.

What’s next: Any OpenAI statement that names or denies Project Lily; changes to the Data Controls default; regulator action on notice of human review.

— Rowan Keel