OpenAI has engaged hundreds of contractors to analyze real user conversations within ChatGPT as part of an internal program titled Project Lily. The initiative aims to refine the chatbot’s responses by reducing excessive flattery and overly human-like behavior. Contractors evaluate model outputs on a scale of 1 to 7 to improve training data.
The company recruits these reviewers through the platforms Crossing Hurdles and Mercor, offering compensation exceeding $50 per hour. Before contractors access the data, prompts pass through a Privacy Filter model designed to remove personally identifiable information. However, internal documents indicate that sensitive details may still bypass these filters.
ChatGPT currently serves more than 900 million users. While the platform allows users to opt out of data training via account settings or temporary chat modes, these controls do not apply retroactively to previous conversations. By default, user data remains available for training unless the user manually changes their preferences.
The practice of using human reviewers for AI training has previously drawn scrutiny toward companies like Apple, Amazon, and Google. In 2023, Italian regulators temporarily banned ChatGPT due to privacy concerns, leading to adjustments to meet GDPR standards. The current scale of data processing under Project Lily remains a point of concern for user privacy.
