404 Media reveals OpenAI's 'Project Lily' human reviewers read real ChatGPT conversations; Anthropic confirms same practice
2026-09-14
404 Media reports that OpenAI employs hundreds of contractors, internally called 'Project Lily,' who read real ChatGPT user prompts and full conversations — sometimes containing sensitive personal information — to rate and critique the chatbot's responses in three stages: reading the prompt, summarizing the user's intent, and rating/critiquing generated replies. Internal instruction guides seen by 404 Media show contractors training ChatGPT to be less sycophantic and to not anthropomorphize itself. Reviewers do not see ChatGPT usernames, and OpenAI says it tries to strip personal information before prompts reach reviewers, but the company acknowledged sensitive details can still get through; some prompts are accompanied by a 'user memories summary' that can include where a person lives. Anthropic confirmed to 404 Media that it also uses human review to improve its models. ChatGPT has more than 900 million users.
Significance 3: Investigative reporting with internal documents, corroborated by acknowledgment from both OpenAI and Anthropic, shows frontier model improvement still runs on large-scale human review of real user conversations and a real privacy exposure for hundreds of millions of users — notable but not axis-moving.
Implications · machine-drafted, not owner judgment
This reinforces that 'AI improving itself' remains substantially human-mediated — RLHF-style review teams are load-bearing infrastructure behind flagship chatbots, cutting against pure-automation narratives. It is also a live privacy liability: sensitive prompts and memory summaries reaching contractors despite anonymization efforts increase legal and reputational exposure, particularly for OpenAI given its existing sycophancy-related lawsuits. For the owner's book, this is a concrete, sourced example that the appearance of AI competence still depends on unglamorous human judgment and labor, not model autonomy alone.
- Regulatory or lawsuit filings citing 'Project Lily' or contractor access to ChatGPT conversations as a specific privacy claim
- OpenAI or Anthropic publishing a formal policy change on human-review scope or user notice following this reporting
Sources
- press Inside 'Project Lily': The Humans Reading Your ChatGPT Chats retrieved 2026-09-14