---
title: "Chain-of-thought monitoring"
type: "entity"
entity_kind: "concept"
status: "watching"
canonical_name: "Chain-of-thought monitoring"
aliases: ["CoT monitoring","CoT monitorability"]
first_seen: "2026-08-24T00:00:00.000Z"
writer_model: "claude-sonnet-5"
connects_to: ["chain-of-thought prompting","OpenAI","chain-of-thought faithfulness","AI safety alignment"]
seek_code_commit: "17d9798"
---


The alignment idea that an AI model's chain-of-thought, left in its raw
"unaltered" form and unoptimized for user preferences, gives outside
observers a way to "read the mind" of the model — to check its reasoning
for signs of manipulation, deception, or misalignment before it acts.
OpenAI names it explicitly as one of the factors it weighed against
competitive advantage and user experience when deciding not to show o1's
raw trace to users
([[claim-openai-hid-o1-raw-chain-of-thought-partly-for-competitive-advantage]]).
Watching for recurrence — the technique's value depends entirely on trace
faithfulness, a question the vault already tracks separately at
[[cot-faithfulness-anthropic-biology]] and [[introspection-access-problem]]
— before promotion to a full hub.

## References

- [[claim-openai-hid-o1-raw-chain-of-thought-partly-for-competitive-advantage]]
- Capture: 10-inbox/raw/2026-08-24-does-openais-learning-to-reason-with-llms-actually.md