OpenAI Uses Hundreds of Contractors to Review Real ChatGPT Conversations

OpenAI employs hundreds of third-party contractors to read real ChatGPT user prompts and full conversations. This practice is part of a broader effort to improve model performance and safety through human oversight.

According to reports from 404 Media, the operation is internally referred to as "Project Lily." Under this program, human reviewers analyze a stream of actual user interactions to identify model failures or "hallucinations." To protect user identity, OpenAI utilizes a "Privacy Filter" model designed to detect and redact personally identifiable information (PII) before the data reaches contractors. However, OpenAI has acknowledged that this automated system is not foolproof, stating, "Like all models, Privacy Filter can make mistakes."

The Role of Human Feedback

This human-led review is a critical component of Reinforcement Learning from Human Feedback (RLHF). Large Language Models (LLMs) do not learn in a vacuum; they require humans to grade, edit, and align AI responses with human preferences and safety guidelines. By reviewing real-world examples of how users interact with the bot, OpenAI can pinpoint specific areas where the model deviates from intended behavior or produces inaccurate information.

Privacy Implications

The revelation of Project Lily highlights a significant gap between user perception and technical reality. Many users interact with AI under the assumption that their conversations are private, yet sensitive data—ranging from corporate secrets to personal confessions—may be viewed by external contractors if the Privacy Filter fails to redact it. This creates a fundamental tension between the industry's need for human-led alignment to make AI safer and the user's right to data privacy.

How to Opt Out

Users who wish to prevent their data from being viewed by human reviewers can opt out of having their information used for training. Exercising this opt-out typically removes a user's conversations from the model improvement pipeline, effectively shielding them from the human review process used for RLHF.

What's Next

As AI companies scale their operations, the reliance on massive fleets of human contractors for data labeling and review remains a systemic necessity. The industry continues to grapple with how to maintain this essential feedback loop without compromising the privacy of the millions of people providing the raw data. This tension suggests that as models grow more complex, the demand for human-verified data will only increase, potentially expanding the scope of these review programs.