When AI acts for you, control points shape the failure
AI systems that write code or use websites can act between moments when a person checks their work. A review of 695 public incidents found 12 ways that human and system can fall out of step. It classifies reported problems; it does not measure how often all users face them.
What the source reports
Canwen Wang and coauthors studied 695 reports of AI systems acting on behalf of users. The reports came from two public collections. The team grouped failures into 12 types across stages of a person working with a system. They also noted how much the system could do before a person stepped in. Every type appeared at both low and high levels of independence, but the form of a failure changed as the system did more on its own. The same model doing the same job can create a different problem under different settings. The authors argue for controls between full handoff and doing everything by hand. Public incident reports show possible failure types, not the odds of failure for every system.
Original source
- Title
- When AI Agents Act on Our Behalf: A Taxonomy of Human-Agent Interaction Failure
- Author
- Canwen Wang, Qing Xiao, Haiyi Zhu, Hong Shen
- Publication
- arXiv
- Date
- Friday, October 30, 2026