Definition
Plain language
Reading what an AI 'thinks out loud' to catch it before it does something bad.
As stated in the literature
A safety practice in which a separate model or human reads a reasoning model's CoT trace before it acts, flagging plans for deception, sandbagging, or policy violations.
Also called: CoT monitoring
Why it matters: It's one of the few oversight tools that scales with capability, but only as long as models keep doing their reasoning in legible text.
For example, a watcher model reads a coding agent's scratchpad and pauses it when the plan mentions disabling logging before exfiltrating files.
Heard on the show
“The second is chain-of-thought monitoring, and the paper's evidence cuts both ways on it.”Episode 128 — How a Model Can Earn Full Reward and Still Resist Training