Ruben Hassid · 6 min read · AI & Product
Privacy.
Read the original on ruben.substack.com →
This article traces the journey of a user prompt through AI systems like ChatGPT and Claude, revealing nine stages of data handling from encryption to potential human review and model training. It uses a fictional scenario of mathematicians solving the Navier-Stokes Millennium Prize problem to illustrate real privacy concerns around how AI companies use user-generated data. The piece highlights that de-identified user data is routinely used to improve AI models across all major providers.
Key Takeaways
- Every prompt sent to AI platforms like ChatGPT or Claude is stored as a record in a database, regardless of whether you're on a free or paid plan.
- AI models don't retain or share your specific chats with other users, but your conversations are stored and can be used as de-identified training data to improve future models.
- Automated safety classifiers review every conversation, and flagged chats can be retained for up to 2 years, with safety scores kept for up to 7 years.
- Human reviewers, including outside contractors, may read a sample of conversations for safety and quality purposes — making thumb ratings a potential signal to avoid on sensitive chats.
- All major LLM providers (OpenAI, Anthropic, Google) use de-identified user data to improve their models, making it critical to understand privacy settings before sharing sensitive information.
Excerpt
Tristan Buckmaster paid for his own ChatGPT ‘Codex’ subscription. He’s a mathematician at NYU. For a year, he and Levent Alpöge (who works at Anthropic, but did this on his own time) went after Navier–Stokes, a Millennium Prize problem. There are seven. Solve any of them and the Clay Institute hands you a million dollars. Grigori Perelman did it in 2003. Nobody has since. On August 15, they had their result using Codex. On September 1, OpenAI heard a rumor that someone had cracked it. It pointed ten thousand agents at the problem. 88 hours later, OpenAI had cracked it…