November 7, 2026 · The Decoder (Manuel Uth)
AI Auditors Catch Most Agent Misbehavior but Flood Reviewers With False Alarms, New Study Finds
My take: More and more businesses are giving AI agents permission to write code, send emails, or make decisions on their own, and the obvious question is: who checks that the agent isn't misbehaving? An independent study that replicated a METR evaluation tested whether other AIs (from OpenAI, Anthropic, and Google) could act as "auditors," reviewing thousands of work logs from other agents. The result: at a reasonable false-alarm rate (5%), the best models caught up to 98% of obvious cheating, but subtler cases slipped through, and the stricter they got to avoid missing anything, the more false alarms piled up. At 100,000 agent sessions a month, the study estimates you'd end up with nearly 3,000 false alerts for every 2,000 real ones.
This confirms something I always say: AI makes you more efficient, not smarter, and it definitely doesn't replace your judgment. One AI agent checking another helps, but it's no guarantee of anything, and handing all the oversight to another machine can leave you buried in alerts you don't know whether to take seriously.
For any business already using (or thinking about using) AI agents to automate tasks, the practical lesson is clear: human oversight is still part of the job, not a step you can skip. The tool saves you time, but the responsibility of checking what matters is still yours.
How closely are you actually reviewing what your AI tools do for you, or have you already handed over full control without looking?
Want to use these tools? See the unbiased reviews or back to the news.