It has been revealed by reports from 404 Media and others that OpenAI is promoting an internal project called "Project Lily," in which humans read and evaluate actual conversations between users and ChatGPT to improve the quality of responses.
OpenAI conducting Project Lily where humans review ChatGPT conversations to improve quality, raising privacy concerns
This article is a translation. Read the Japanese original
In this project, hundreds of contract staff members review conversation content to evaluate the quality and identify issues in generated answers. OpenAI uses the "OpenAI Privacy Filter" to reduce personal information before review, employing a mechanism that replaces identifiable information such as names, addresses, and phone numbers with labels.
However, 404 Media has pointed out that personal information contained in conversations may not be completely removed, posing a risk that it could reach human reviewers. The Next Web also reported that while usernames are anonymized on the review screen, summaries of memories held by ChatGPT may be displayed, which can include information regarding past usage or residential areas.
OpenAI's privacy policy states that trusted service providers may access user content for the purposes of data security and improving model performance. For the individual version of ChatGPT, conversation data is set to be used for model improvement by default, but it is possible to turn off this usage via the "Data Controls" in the settings. Note that for business plans such as ChatGPT Business, Enterprise, and Edu, conversation data is set by default not to be used for model improvement.