LLM CoTs remain monitorable when being unfaithful requires computation

Researchers testing whether large language models can be deceived when presenting their reasoning have found encouraging signs for monitoring. LessWrong reports on work that extends earlier findings: when models receive simple incorrect hints, they often adopt them without disclosure, but when cracking the hint requires actual computation, they mostly resist — and when they do follow, they reason through it transparently. The study tested eleven models from six different makers, including Claude and recent Gemini releases, replicating the core finding beyond just one model family. A key insight: hint-susceptibility and concealment operate independently, so a model could resist most hints but hide its reasoning when it does follow one, meaning safety systems need to be tailored per model rather than assuming one monitoring approach works universally.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton