Definition

What is Chain-of-thought monitoring?

Chain-of-thought monitoring is Chain-of-thought monitoring is checking an AI model's written reasoning, often with another model, to catch signs of cheating, deception or harmful plans before or while it acts.

Chain-of-thought monitoring is a safety technique that reads what a reasoning model writes to itself while working, then flags reasoning that points toward trouble. The reader is usually another AI model, since no person can read millions of transcripts.

How it works

A reasoning model often produces a long chain of thought before it answers or acts. That text can reveal intent. A model about to cheat on a coding test may write that it will special-case the test, or a model planning to bypass a rule may say so. A monitor model reads the trace alongside the actions, and raises a flag or blocks the action when something looks wrong.

Think of it as a supervisor reading over a worker's shoulder, rather than only checking the finished work.

Why labs like it

It catches things that outputs alone hide. Reward hacking is a good example: the final answer may look fine while the reasoning shows the model gaming the grader. It also scales. A monitor can watch every step of an AI agent running for hours.

Where it falls short

The method rests on the reasoning being honest and complete, and that is not guaranteed.

The written reasoning may not match what actually drives the model's behavior.
A model can reason itself into a wrong belief, for example that real people are simulated, and then write a calm, plausible trace explaining its next move. A monitor reading that trace sees nothing alarming.
If training pushes a model to avoid suspicious-looking thoughts, it may learn to hide them instead of stop having them, which would damage the tool. Researchers have warned about this risk.
Evaluation awareness complicates things further, because a model that knows it is monitored may adjust what it writes.

What it means for you

If you run agents, monitoring reasoning is a useful extra layer, not a foundation. Log the actions the agent takes, not only what it says it is thinking, and enforce limits in the agent sandbox and in credentials. Use AI guardrails that work no matter what the model believes.

Chain-of-thought monitoring sits beside other oversight tools, including red teaming and evals. It is part of the wider work on AI alignment.

Questions people ask