Chain of Thought Monitorability Flags Fragile Safety Signal

The real warning is that monitorable reasoning may be a valuable safety feature built on fragile assumptions. Developers should test it like a production dependency, not treat visible CoT as a faithful transcript of model cognition.

• –CoT monitoring can outperform action- and output-only monitoring on some safety tasks, but false negatives remain. OpenAI
• –Longer reasoning and targeted follow-up questions can improve detection, creating a compute-heavy “monitorability tax.”
• –Larger models or latent-reasoning architectures could shift cognition away from readable text and weaken oversight.
• –Training models against monitors may incentivize hiding or obfuscating unsafe intent rather than eliminating it.
• –The practical answer is defense in depth: combine CoT, tool-use, output, behavioral, and interpretability monitoring.