OpenAI has hired contractors to review ChatGPT user conversations as part of Project Lily, aiming to enhance response quality.
OpenAI has reportedly enlisted hundreds of external contractors to review actual conversations from ChatGPT users as part of a project called Project Lily, according to findings shared by 404 Media.
Documents related to the initiative reveal detailed reviewer instructions, internal communications, and evaluation systems intended to enhance the quality of responses generated by ChatGPT.
The contractors are tasked with analyzing entire conversations rather than just individual messages, which sometimes include sensitive personal data. A source involved expressed concern that users might not be aware of this scrutiny, suggesting, "I don’t think they realize that some contractor is analyzing these conversations somewhere."
This review initiative is distinct from the monitoring procedures OpenAI has publicly disclosed, which focus on identifying potential threats. Project Lily operates independently from these acknowledged checks.
Review Process Overview
Contractors access tasks via a panel where they review user inquiries. Each reviewer follows a structured process: they read the request, summarize the user's intent, and evaluate responses generated by ChatGPT. For instance, one task involved a user seeking to rephrase a Slack message to make it more inviting.
Reviewers assess four different responses, marking segments that meet or fail the training model's criteria. They must justify their selections and rate the responses on a scale from 1 to 7, with 1 being unusable and 7 indicating excellence. Even informative responses can receive low scores if they are too lengthy or convoluted.
The documents did not specify which model was being trained, leaving open whether it pertains to an existing or future version.
Response Evaluation Criteria
Review guidelines specify that acceptable answers should accurately capture the user's intent while maintaining clarity and a friendly tone. Reviewers are instructed to avoid responses that appear overly subservient or mimic human emotions, focusing instead on helpfulness and authenticity.
They are also tasked with eliminating excessive "AI-speak" and inappropriate emoji usage. Context is critical; for example, mentioning a tree in an Arbor Day discussion is appropriate, while references to skulls in sensitive topics are not.
Reviewers are not responsible for fact-checking but may flag inaccuracies, especially in responses to medical, legal, or financial queries that lack citations.
Contractor Awareness and Privacy Measures
User inquiries are anonymized to protect identities, but reviewers sometimes encounter personal data. They are instructed to escalate any tasks that involve sensitive information or potential safety risks. OpenAI acknowledged that their Privacy Filter, which aims to remove personal data, is not flawless and can sometimes miss or misinterpret identifiers.
404 Media reached out to OpenAI for clarification on whether users are informed about the potential for their requests to be reviewed, but did not receive a response. While OpenAI's privacy policy mentions that flagged content may be reviewed by humans, this is typically related to security concerns.
Contractor Profile and Compensation
One contractor reported earning over $50 per hour for this work, which they found through a recruitment firm called Crossing Hurdles. Job descriptions indicated responsibilities related to evaluating AI responses, although specific mentions of OpenAI or ChatGPT were absent.
Feedback from contractors suggests that the work can be monotonous, with frequently changing guidelines. Other AI companies, such as Anthropic, also employ human reviewers for similar purposes, emphasizing the trend in the industry.
Expert Opinions on Human Review
Experts have weighed in on the implications of human review in AI interactions. Michal Luria from the Center for Democracy & Technology highlighted the potential for false intimacy and privacy perceptions in chatbot interfaces. Sarah T. Roberts from UCLA likened the situation to the revelation of hidden mechanisms in The Wizard of Oz, stating that constant human intervention is necessary for effective operation.
Notably, the hourly pay for these contractors is significantly higher than that of typical social media moderators, raising questions about the valuation of essential human labor in AI development.
In April 2026, OpenAI introduced the Privacy Filter, designed to help cleanse confidential information from user conversations with ChatGPT.
