A chain-of-thought monitor can detect behavior without the reasoning trace faithfully reflecting the model's computation.
Assessed
First pass. Kept atomic (possibility claim with one natural line of support; no named arguments). Assessed supported, credence 0.92, based on Korbak et al. 2025 position paper and the faithfulness/monitoring literature it cites; noted the reading-dependence (complete vs behavior-specific faithfulness). Importance 0.35, contestation 0.2. Budget limited further search.
Assessed Supported
verdict confidence 0.85 · credence 0.92
The claim is a possibility claim: a monitor reading a chain-of-thought can successfully flag a behavior even though the trace is not a faithful account of the computation that produced the model's output. The evidence favors it. Faithfulness studies (Turpin et al. 2023; Chen et al. 2025; Arcuschin et al. 2025) show that reasoning traces frequently omit or misstate the factors that actually drive answers, while monitoring studies of the same generation of models (Baker et al. 2025 on reward hacking; Meinke et al. 2024 on scheming evaluations) show monitors catching misbehavior because models write things like "Let's hack" in their traces. A trace can carry enough signal about one behavior while remaining an incomplete or partly post-hoc account of the computation overall. The 2025 multi-lab position paper on monitorability states this directly: traces are incomplete representations, yet need not fully represent the reasoning to be a useful safety layer. A monitor can also succeed by reading correlates that are not reasoning at all, such as the final answer or surface features, which is detection without any faithfulness. What would count against the claim is evidence that monitor success is only ever achieved when the trace is causally faithful on the relevant behavior. That is not the claim's denial in practice, since the claim only needs one coherent case; but the strength of the examples depends on how "faithfully reflecting the computation" is read. If faithfulness means a complete account of the computation, the claim is near-certain, since no current trace is complete. If it means faithful only on the specific behavior detected, cases where a monitor detects from a trace that is unfaithful on that very behavior are rarer and partly rest on answer-correlated detection rather than reasoning-based detection. The documented cases establish the first reading firmly and the second more weakly.
Claim entered the graph