Can AI systems check each other? Researchers have proposals.
Could one AI system catch another making a bad choice? MIT Technology Review describes proposals to have systems predict each other and take turns doing and checking work. A person might settle disputes. Researchers also warn of shared tricks and failures on new tasks. These ideas are not proven safety controls.
What the source reports
Amit Katwala reports on proposals to make AI systems check each other. Here, the systems can carry out tasks rather than only answer questions. Researchers discuss having them predict what other systems will do. Another proposal lets systems take turns solving a task and checking the work. A person could step in when the checks disagree. That arrangement might reveal a problem, but it also creates new questions. Systems could work together to fool the check, and a method that works on one task might fail on another. Labs and markets may have incentives that make oversight harder. The reported ideas are proposals for control. They have not been shown to keep advanced AI systems safe in real-world use.
Original source
- Title
- How “reverse game theory” could keep AI agents in check
- Author
- Amit Katwala
- Publication
- MIT Technology Review
- Date
- Tuesday, November 3, 2026