OpenAI Contractors Read Live ChatGPT Conversations

The Next Web (14 September 2026) and The Decoder report that OpenAI is paying hundreds of contractors, under the internal name Project Lily, to read real ChatGPT conversations and rate the replies. Tom's Hardware carried the same 404 Media investigation.

This is investigative reporting from 404 Media, later picked up by those outlets. It is not an OpenAI press release confirming every claim.

Reviewers do not see usernames. They can see a user memories summary above the prompt, which can show prior use and location cues.

OpenAI says a Privacy Filter model strips personal information first. The company's own page for that filter says it can make mistakes, miss uncommon identifiers, and under-redact when context is limited.

The Decoder, citing 404 Media, says contractors are recruited through Crossing Hurdles and paid through Mercor. Reviewers rate replies on a scale of 1 to 7 and are tasked with cutting excessive flattery and human-like behavior.

Improve the model for everyone is on by default for free, Plus, and Pro. It is off for Enterprise, Business, and Edu. Switching it off applies to new conversations, not old ones.

OpenAI did not answer where users are told humans may read chats. After publication it pointed to a help page. Whether that buried FAQ is enough notice is disputed in the coverage, and the GDPR reading is still being argued.

The Court of Justice, in EDPS v SRB last September, held that a controller's duty to inform applies at collection from the controller's point of view. Italy's regulator previously fined OpenAI EUR 15 million, nine million of that for processing without an adequate legal basis, and ordered six months of public information advertising.

Anthropic confirmed human review for users who enable its improve-model setting, after stripping account IDs. Google Gemini notes that humans may review some saved chats.

Related OpenAI product notes include ChatGPT Ads at a $1 billion run rate, ChatGPT for Finance with Morgan Stanley, and the ChatGPT Pro pause amid Astra demand.

If you use consumer ChatGPT and do not want a contractor reading the thread, turn off Improve the model for everyone and use temporary chat for anything sensitive. Do not paste bank, health, or identity documents into consumer ChatGPT while that setting is on.