OpenAI Hired Contractors To Review ChatGPT Chats but Improve Opt-Out Does Not Remove Earlier Eligible Chats

OpenAI hired contractors to read real ChatGPT conversations under an internal programme called Project Lily, while its 'Improve the model for everyone' setting does not automatically remove chats that were already eligible for model improvement when a user later opts out, according to reporting by 404 Media published on 14 September 2026.

The contractors assess ChatGPT's responses to real user prompts, and some assignments contain entire conversations rather than individual messages. OpenAI's current guidance says switching off the Improve setting prevents new conversations from being used to train its models.

The result is a difference between conversations that become eligible after a user opts out and those that were already eligible before the setting was changed.

The reporting does not establish that a contractor continues opening a particular user's old chat after the setting is switched off, but changing the setting later does not automatically withdraw previously eligible material from the improvement process.

For Free, Plus and Pro users, the Improve setting is enabled by default. Business, Enterprise and Edu accounts have it disabled by default.

How Project Lily Uses Human Review

404 Media reported that Project Lily contractors are given real ChatGPT prompts and asked to determine what the user is seeking before assessing ChatGPT's responses.

The review process involves three stages: reading the prompt, summarising the user's request and evaluating four generated responses. Contractors score the responses from one to seven and identify specific elements that work or fail.

The internal guidance covers issues including excessive emoji use, 'AI-speak', forced imitation of a user's writing style, sycophancy and responses that falsely suggest ChatGPT has personal experiences, such as 'As a chef, I...' or 'I know what that's like.'

The work is not limited to safety or abuse investigations. 404 Media reported seeing assignments containing whole conversations, while reviewers do not see the account username.

Privacy Filtering Does Not Remove Every Detail

OpenAI says conversations go through a version of its Privacy Filter before reaching contractors. OpenAI's documentation says the filter is designed to detect and remove personal information, but can miss uncommon identifiers and ambiguous private references.

The documentation also says the filter can over-redact or under-redact information when there is limited context. Removing direct identifiers therefore does not guarantee that every piece of personal information disappears from material sent for review.

Some Project Lily assignments also display a 'user memories summary' above the prompt, according to 404 Media. The summaries can describe what a user has previously used ChatGPT for and, in some cases, include information about where the user may live and other personal context.

That means a conversation can contain identifying details even when the account username is not visible to the reviewer.

404 Media asked OpenAI where users are told that their conversations may be reviewed by people to improve ChatGPT's responses. OpenAI did not directly answer that question. The company later pointed to existing help language concerning human review and updated its opt-out guidance.

Turning off Improve Only Covers New Chats

OpenAI's current help guidance says that when a personal-account user switches off 'Improve the model for everyone', new conversations will not be used to train its models. The wording does not describe the change as a retroactive withdrawal of conversations that were already eligible for model improvement.

Project Lily is part of the model-improvement work described in 404 Media's reporting. A conversation that was already eligible before the user changed the setting is therefore not automatically removed from that broader process when the opt-out is activated.

OpenAI separately says deleted conversations are generally removed from its systems within 30 days. Its policy makes an exception when content has already been de-identified and disassociated from an account after being allowed for model improvement.

The available reporting does not show that every older ChatGPT conversation is reviewed by contractors, nor does it establish that a contractor necessarily keeps opening a particular user's conversation after the user opts out.