OpenAI Hires Contractors to Review ChatGPT User Chats in Secret Project
A recent investigation reveals that OpenAI employs hundreds of contractors in a secret initiative called Project Nightly to review user prompts and chat histories, raising privacy concerns.
OpenAI has been employing hundreds of contractors to review user prompts and chat histories in a previously undisclosed initiative known as Project Nightly, according to a recent investigation by 404Media.
Inside Project Nightly
The report reveals that OpenAI utilizes contract workers to analyze prompts submitted to ChatGPT, aiming to improve the model's responses and reduce excessive flattery. While OpenAI's public privacy policy states that human reviews are generally reserved for safety incidents such as self-harm threats, internal guidelines leaked in the investigation show contractors evaluating full conversational contexts, including sensitive personal details.
Contractors are instructed to summarize user intent, evaluate model outputs for natural tone, and flag AI hallucinations or formatting issues, such as inappropriate emoji usage. Although usernames are stripped from the data, the prompts frequently contain identifiable personal information and snippets from the chatbot's memory feature.
"An excellent response must understand user intent, provide accurate and helpful information, and be written in a clear, natural, and appropriately pleasant style." — OpenAI internal contractor guide
Compensation and Industry Context
According to the investigation, contractors in North America are paid upwards of $50 per hour through specialized recruitment firms such as Mercor. While OpenAI noted that automated privacy filters are applied before data reaches contractors, the company acknowledges in support documentation that these filters may occasionally fail to redact sensitive details.
How to Opt Out
Users concerned about their data privacy can disable human review by adjusting their account settings. By navigating to ChatGPT settings and turning off the "Improve the model for everyone" option, users can prevent their future prompts from being reviewed by human contractors. However, this setting does not apply retroactively to past conversations.
