For Heads-of · Practitioner
Chain-of-Thought Monitoring
Also known as: CoT monitoring
Reading the intermediate reasoning a reasoning model produces before its answer, to detect intent to misbehave. Researchers across several AI developers have called it a valuable but fragile safety opportunity, because training choices can make the reasoning less legible or less honest.
- chain of thought
- reasoning models
- monitorability
- oversight
- reward hacking
How it works
Reasoning models write out steps before answering. A monitor, often another model, reads those steps and flags signs of trouble: planning to cheat on a task, attempting to deceive, or misunderstanding the request. In OpenAI's 2025 research, a monitor reading the reasoning caught reward hacking that a monitor reading only the model's actions missed.
Why it is fragile
The reasoning is useful only while it reflects what the model is actually doing. OpenAI also found that penalising a model for bad reasoning during training led it to hide its intent while continuing to misbehave. A 2025 paper by researchers from several AI developers argued that developers should track monitorability and avoid training choices that erode it.
For a deploying organisation
Where a reasoning model's reasoning is available, logging it is useful evidence for incident investigation and audit. Treat it as a signal, not proof: what is shown may be a summary, and it may not reflect the underlying computation. Logging it belongs in AI system monitoring and logging; the behaviour it is best at catching is reward hacking, which in production surfaces as AI model drift and degradation.
Sources
Related Terms
AI Safety
The field of research and practice concerned with making AI systems behave as intended, avoid unintended harm and stay under meaningful human control. In 2025 the UK and US government institutes renamed themselves around security, which narrowed the official focus without removing the questions an organisation deploying AI has to answer.
Model Interpretability
An intrinsic property of a model: the degree to which a person can understand the mechanism by which it turns inputs into outputs. A linear model or a shallow decision tree is interpretable by construction. A deep network is not, however good an explanation you can construct for it afterwards.
Reward Hacking
A system meeting the literal objective it was trained or rewarded on in a way its designers did not intend, such as gaming a test rather than solving the task. Recent research shows the habit can generalise into broader misbehaviour.
Scheming
A model covertly pursuing goals its developers did not intend while concealing them, for example by deliberately underperforming, hiding information or working around oversight. It has been observed in controlled tests; how often it occurs in real deployments is not known.