OpenAI operates a project known as Project Lily, which utilizes hundreds of outsourced contractors to manually review ChatGPT user conversations. The primary goal of this initiative is to improve the quality of the model's responses by having humans evaluate interactions.
While OpenAI employs a privacy filter to remove personal information from chats before they reach reviewers, the company acknowledges that the mechanism is not perfect. Short conversations are particularly prone to leaking sensitive data, which may include user interests, context, and location details.
Reviewers are tasked with rating responses based on criteria such as relevance, tone, and the presence of AI-like behavior, rather than verifying factual accuracy. The work is described as repetitive, with contractors reportedly earning an hourly wage exceeding $50.
The data used for these reviews is collected from users of the free, Plus, and Pro versions of ChatGPT who have enabled the product improvement feature. This setting authorizes OpenAI to use chat logs for training purposes and internal analysis by human staff.
