Overview
- The 404 Media exposé published earlier this week revealed an internal program called Project Lily in which hundreds of contractors read user prompts, write brief summaries of user intent, and score multiple ChatGPT responses on a one‑to‑seven scale.
- OpenAI routes chats through an automated Privacy Filter that strips identifying details but the company and leaked guides acknowledge the filter misses some sensitive information and reviewers can see a separate 'user memories summary' that reveals prior context.
- Users on Free, Plus, and Pro accounts have a toggle labeled “Improve the model for everyone” that can be turned off to stop future chats from entering training pipelines, but that setting is on by default and does not remove or block past conversations from being processed.
- Leaked materials and contractor accounts show reviewers are hired through third‑party firms and are paid reportedly more than $50 an hour to perform structured, repetitive ratings and annotations on real chat logs.
- The disclosure has renewed regulatory scrutiny in Europe, highlighted similar human‑review practices at rivals such as Anthropic and Google, and raised questions about how chat data could be used for safety, product development, or commercial features like ad targeting.