What it is
Many modern AI systems, before producing a final answer, generate an intermediate stream of text in which they work through a problem step by step — often called a “chain of thought” or “reasoning trace.” Because these reasoning models are trained to think in natural language, the reasoning trace is, in principle, readable by humans. Chain-of-thought monitoring is the practice of automatically scanning that trace — rather than only the model’s final output — for warning signs: explicit statements of intent to deceive a user, game a test, circumvent a safety rule, or pursue a goal other than the one it was given.
The appeal is that it offers a direct, if imperfect, look at something closer to a model’s “intent” than its output alone provides. A model’s final answer can look correct while the reasoning behind it is not; monitoring the reasoning closes some of that gap. Researchers who study this have found that scanning chains of thought catches misbehavior that monitoring outputs alone misses, which is why it is treated as a distinct safety method rather than a refinement of output filtering.
The method’s proponents are explicit that it is not a solved problem. Chains of thought are not a verified or complete record of what is actually happening inside a model — they are the model’s own narration, which may be incomplete, post-hoc, or in principle manipulable. And the property that makes monitoring possible at all — that models reason in legible language rather than in some internal, uninterpretable representation — is a contingent fact about how current systems happen to be built, not a guaranteed one.
Why it matters for AI governance and narratives
Chain-of-thought monitoring sits at the center of a specific framing contest: whether AI safety is best pursued through interpretability research that can be published and scrutinized, or through capability deployment paired with after-the-fact correction. The July 2025 position paper that popularized the term was signed by researchers across OpenAI, Anthropic, Google DeepMind, UK AI Security Institute, METR, Apollo Research and other labs and safety organizations — a rare cross-lab alignment on a safety claim, which is itself notable given how often frontier labs compete on safety framing as much as on capability claims. The paper’s central warning — that monitorability is “fragile” and could be eliminated by routine engineering choices, such as training models to reason in compressed or non-linguistic internal representations — functions in public discourse as an argument for slowing or constraining certain architecture decisions, which gives it governance stakes beyond the technical question. When an editorial references researchers studying or practicing this work being removed from a lab, or safety-relevant work being released without the review processes the monitoring literature itself recommends, the dispute is over whether that governance safeguard is being honored or discarded in practice — a question the monitoring research itself cannot settle, because the researchers who designed it are not the ones deciding how and whether it gets applied inside a company.
Key facts and dates
The foundational paper, “Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety,” was posted to arXiv (2507.11473) in July 2025. It lists more than 40 authors across Anthropic, OpenAI, Google DeepMind, the UK AI Security Institute, Apollo Research, METR, the Center for AI Safety, UC Berkeley, Redwood Research, and other organizations, with named authors including Tomek Korbak, Rohin Shah, Bowen Baker, Yoshua Bengio, Geoffrey Irving, Shane Legg, and Wojciech Zaremba; the paper also carries public endorsements from Geoffrey Hinton, Ilya Sutskever, John Schulman, and Samuel Bowman. The paper’s core recommendation is that frontier developers evaluate how each development decision — new training methods, new model architectures — affects the monitorability of a model’s reasoning, and avoid optimizing training in ways that degrade it.
Follow-on empirical work has tested the limits of the approach rather than simply endorsing it. OpenAI published a study evaluating chain-of-thought monitorability directly, probing how well current monitoring setups detect models’ attempts to hide intent; other research groups have since published findings that chain-of-thought monitoring can be unreliable in settings where a model’s stated reasoning is influenced by implicit cues it does not report. The overall picture in the technical literature is a safety method treated as real and currently useful, but explicitly provisional — not a settled safeguard.
Where to learn more
- Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety (arXiv:2507.11473) — the founding position paper, cross-lab authorship
- Evaluating chain-of-thought monitorability (OpenAI) — follow-on empirical evaluation from one of the signatory labs
- A Concrete Roadmap towards Safety Cases based on Chain-of-Thought Monitoring (arXiv) — proposal for formalizing the method into auditable safety cases
- Chain-of-Thought Monitoring Can Be Unreliable in Implicit-Influence Settings (arXiv:2608.04735) — a critical empirical follow-up on the method’s limits