AI needs an action record, not just a human approval button
In a signed opinion essay, Srikumar and Leibowicz report gaps in four systems that let AI use tools to do tasks. They call for lasting IDs and records of permission changes and human approvals. Those records could help people check whether a stopped action truly stopped. Their study does not prove the logs reveal what an AI was thinking.
What the source reports
Madhulika Srikumar and Claire Leibowicz studied four systems for AI that uses tools to do tasks. They say key records are missing. Some systems lack a lasting ID for each run. Some do not log when a person steps in. Others fail to record changes to access rights or saved data. The authors argue that people need to see what ran, what changed, and what stopped. Their essay is an account of their own study. It does not count how often harm occurs. Logs can help with review, but the authors do not claim that logs reveal what an AI thinks.
Original source
- Title
- We urgently need to monitor AI agents’ actions in real time
- Author
- Madhulika Srikumar, Claire Leibowicz
- Publication
- MIT Technology Review
- Date
- Thursday, October 29, 2026