Internally, OpenAI has introduced a program called "Project Lily," in which around 100 contractors review real conversations from users of its ChatGPT chatbot. These workers assess the quality of the responses generated by the AI, though they are not given detailed information about the nature of the review process. Contractors access a dashboard where they choose tasks, read anonymized user prompts, and sometimes receive a summary of the user's previous interactions, including their location. They evaluate several responses from ChatGPT, assign scores from 1 to 7, and explain their reasoning in writing. Each contractor summarizes the user's intent in one sentence, compares four different responses, and identifies at least three aspects that either align or do not align with expected behavior. The scoring is based on how well the responses match expected standards, with special caution given to responses involving medical, legal, or financial topics that lack proper sources. Before reaching the contractors, the user conversations go through a filtering model known as the Privacy Filter, which is designed to remove any personal information. However, OpenAI acknowledges that this tool can occasionally allow rare personal identifiers to remain or misjudge the removal of such data. The company did not respond to questions about whether users are explicitly informed about this review process. On its website, OpenAI only mentions human review in the context of reported or potentially harmful content, not for this specific purpose. Within ChatGPT's settings, there is an option labeled "Improve the model for everyone," which can be turned off. However, this option is set to opt-out by default on the Free, Plus, and Pro plans, while it is disabled entirely on the Enterprise, Business, and Edu plans. This setting does not affect conversations that have already occurred. OpenAI is not the only company using such a system. Anthropic, another AI company, confirmed to 404 Media that it employs a similar human review process to refine future responses from its chatbot, Claude. The only difference is that Anthropic allows users to disable this feature, unlike OpenAI. In the case of OpenAI, the user's identifier is separated from their email address, ensuring that the contractors reviewing the conversations do not have access to personal details beyond what is necessary. The contractors involved in this process are recruited through a firm called Crossing Hurdles, which directs them to Mercor, the company responsible for their pay. A worker interviewed by 404 Media, based in North America, reported earning over $50 per hour, a rate higher than what is typical in content moderation or data annotation roles. However, Mercor recently faced scrutiny after Meta ended its collaboration with the company in April due to a data breach that affected Mercor.