OpenAI has engaged hundreds of external contractors to review real-time ChatGPT user conversations as part of a project codenamed Project Lily. The initiative aims to improve model responses by having reviewers analyze user intent and evaluate AI performance based on specific quality guidelines.
Contractors access user queries through a dedicated interface where they assess four AI-generated responses. Reviewers rate these responses on a scale from 1 to 7, focusing on clarity, tone, and the absence of unnecessary AI-isms like excessive emojis or simulated emotions. Fact-checking is not part of the review process.
Although OpenAI uses a Privacy Filter to redact sensitive information, the tool has known limitations. Reviewers may still encounter personal data, including user memory summaries that reveal location and past interactions. Contractors are instructed to escalate tasks if they identify security risks or sensitive personal details.
The data collection process is linked to the 'improve the model for everyone' setting, which is enabled by default for Free, Plus, and Pro users. While Enterprise, Business, and Edu accounts have this setting off by default, users must manually opt out to prevent their future conversations from being reviewed by humans.
One contractor reported earning over $50 per hour for this work, recruited through the firm Crossing Hurdles. Other AI companies, including Anthropic and Google, also utilize human reviewers to improve their models, though they maintain different disclosure protocols regarding the use of human oversight in their systems.
