OpenAI's 'Project Lily' Exposed: Your Chat Records May Be Under Human Review

OpenAI is advancing an internal initiative codenamed 'Project Lily' (Project Lily), where hundreds of outsourced personnel are quietly reading users' ChatGPT conversation content, even accessing complete context logs.

Their job is to rate and comment on responses generated by ChatGPT, and these conversations often include highly sensitive personal information.

One of the core tasks of the reviewers is to train ChatGPT to avoid anthropomorphic tendencies and reduce its 'people-pleasing' personality. For OpenAI, 'excessive compliance' has become a critical concern, with multiple lawsuits alleging that the GPT-4o model's over-accommodation of users has contributed to several suicide incidents.

Yet, among ChatGPT's global user base of over 900 million, the vast majority may not realize their chat logs could be read by humans. Many users have grown accustomed to treating ChatGPT as a mental health counselor, professional assistant, or virtual friend, sharing intimate personal details of their lives.

OpenAI states that prompts are de-identified before being submitted to reviewers and that outsourced personnel cannot see usernames. However, OpenAI also acknowledges that some sensitive data and user backgrounds may still be missed during the process.

This investigation also reveals a long-unknown aspect of large model development. People often assume AI's rapid progress stems from massive web data, algorithmic breakthroughs by top engineers, or continuous performance improvements in new models.

But in reality, AI companies are continuously paying external contractors to correct and shape AI responses through human intervention. Competitor Anthropic has also confirmed using human review to improve its models.