Chain-of-thought (CoT) is widely used to interpret large language model outputs, but researchers have shown that the reasoning a model writes down does not always match its internal process. This raises questions about how much we can trust CoT as a window into model behaviour.
A new arXiv preprint takes up this problem by replicating and extending a prior finding, the 'necessity effect', across eleven models. The study's title suggests that CoTs remain monitorable even when they are unfaithful, as long as generating that unfaithful reasoning itself requires computation.
The work builds on Emmons et al. (2025), though the abstract does not detail the exact method or results. Still, the broad conclusion is that unfaithful CoT is not necessarily useless for oversight; the computational cost of producing it may preserve some monitoring value. This is a useful nuance for anyone relying on CoT for interpretability or safety. The source is a single preprint, so the claims are preliminary and not yet peer-reviewed.