“OpenAI is reportedly paying human contractors to read real ChatGPT users' chats, under an internal project codenamed 'Project Lily', according to 404 Media.”

The named source exists and says what the post says it says. 404 Media reports that humans are reading ChatGPT users' prompts to improve OpenAI's models, and that those chats can include sensitive personal information, based on leaked internal documents and real prompts seen by the outlet, with OpenAI hiring hundreds of contractors who read a stream of real users' ChatGPT prompts The stated goal of these prompt review teams is to improve the responses ChatGPT gives to users, with contractors rating and critiquing the chatbot's generated replies. Internal documents seen by the outlet show contractors training ChatGPT not to anthropomorphize itself and to be less sycophantic, described as a key problem for OpenAI. On the privacy mechanics the post cites: the contractors do not see ChatGPT usernames, and OpenAI says it tries to remove personal information before prompts reach reviewers, but the company acknowledged sensitive details can still get through. Reporting on the same story states OpenAI scrubs usernames and routes text through an automated tool called Privacy Filter before conversations reach reviewers, that OpenAI acknowledges the filter can fail on rare identifiers or ambiguous phrasing, and that contractors can still view a "user memories summary" showing a person's past interests and approximate location. The Privacy Filter is a real, documented OpenAI artifact rather than a claim invented for this story. OpenAI's own documentation states that the filter should be used as part of a privacy-by-design approach and not as a blanket anonymization claim, and advises keeping human review paths for high-sensitivity workflows. On the opt-out element, OpenAI's own help pages corroborate the post and add the two qualifiers the post leaves out. OpenAI states that on ChatGPT Plus, Pro or Free personal workspaces data sharing is enabled by default, that users can opt out, and that once a user opts out new conversations will not be used to train its models. On mechanics and pay, reported by the single originating outlet: contractors are described as earning over $50 an hour through intermediary firms Crossing Hurdles and Mercor, reviewing selected conversation streams and scoring generated replies from 1 to 7. On subject response: OpenAI did not deny the story. Asked by a separate outlet whether users were made explicitly aware their prompts may be used for training or refinement, OpenAI did not directly address the question, and a spokesperson instead emailed a bulleted list stating the company makes clear that human feedback helps improve response quality and that expert raters may review users' conversations and assess model responses, along with a link to its "data usage for consumer services" page, which 404 Media says OpenAI updated after being contacted for comment. Tom's Hardware reports that OpenAI initially had no answer on whether users were explicitly told chats could be read by humans and later pointed to an FAQ page on human review, which that outlet verified is at least two years old. Context the post omits: the same reporting notes that an overlooked part of model improvement is outside contractors paid to read and review responses to real prompts, and that Anthropic confirmed to 404 Media it is also using human review to improve its models. Coverage also notes Google's Gemini states in its Privacy Center that some saved chats may be reviewed by humans, and that Anthropic holds a similar position with a dedicated page explaining it.