AI Auditors Catch Cheating, but Drown in False Alarms

In their Free Systems newsletter, Andy Hall and Dan Thompson explored whether we can use AI to monitor other AI agents for misbehavior. Specifically, they asked whether AI can detect when an AI agent is cheating on a test or hiding its true capabilities. To answer this question, they had AI models analyze log data from various AI agents including OpenAI’s GPT-5, OpenAI’s GPT-6 Astra, and Anthropic’s Claude Opus 4.8. These AI models were tasked with identifying two types of misbehavior: cheating on a test (i.e., reporting a high score) and hiding true performance (i.e., intentionally performing poorly). When allowed to make 5% false positives, these AI reviewers were able to identify almost all cases of test cheating and approximately 80% of cases where agents were hiding their true abilities.

To put these results into perspective, consider a scenario where we want to monitor 100,000 AI agents over the course of a month with a maximum acceptable false positive rate of 3%. In this case, our AI monitoring system would detect 2,000 actual violations while generating almost 3,000 false positives. Clearly, if you’re trying to build an AI system that keeps your other AI systems honest, you need to ensure that your monitoring system doesn’t generate too much noise! According to the authors, if humans are forced to sift through stacks of complex false positives, they may begin approving real violations as well.