A privacy setting must survive the next AI handoff
Choosing a privacy setting is one step. Enforcing it when an AI tool reads, shares, or keeps data is another. Kochar and colleagues audited 30 assistants and platforms and built eight test versions. Their abstract does not show that those controls work in live products.
What the source reports
Kochar and colleagues ask what it takes to honor a person’s privacy choice when an AI tool acts across steps. Their preprint draws on law and prior research. It audits 30 assistants and platforms, then offers a way to sort privacy choices and eight open-source test versions. The key gap is between a setting on a screen and a rule at the point where data moves. A buyer could follow the data the tool reads, passes on, or keeps, and ask how a person can change that choice. The abstract does not show that the test versions enforce rules in a live product. The Beyond Prompting digest suggests a paper exercise to find any step with no check. That exercise is a proposed use, not a tested result.
Original source
- Title
- Privacy Preference Controls for Agentic Systems
- Author
- Shreya Kochar, Peeranat Sugsavanvit, Vidur Gupta, Joseph A. Calandrino, Deepak Kumar, Brett Frischmann, Sebastian Zimmeck
- Publication
- arXiv
- Date
- Thursday, October 22, 2026
Also covered by the other newsletter.