OpenAI is hiring evaluators to read real conversations with ChatGPT in order to improve the model

OpenAI has hired hundreds of people to analyze real conversations held by users with ChatGPT as part of a process aimed at improving the system’s responses. The information was reported by 404 Media and picked up by Nexta.

The evaluators may have access to entire conversations and analyze how the neural network responds to requests. They also monitor the reduction of behaviors considered undesirable, such as excessive politeness, a tendency to become “human-like,” or other inappropriate reactions.

The project has the internal name “Project Lily,” and collaborators are recruited, among other ways, through the companies Mercor and Crossing Hurdles. OpenAI says it is trying to automatically remove personal data before the dialogues are sent to evaluators and that they do not see users’ names.

However, the company acknowledges that automated filters are not perfect and that certain private details may end up in the conversations being analyzed. Critics warn that users who have provided personal information or secrets in discussions with ChatGPT could have them exposed to people outside their circle.