As models' reasoning departs from standard English, human-legible chain of thought becomes less reliable for oversight.
6 events · 2 assessments · 3 decisions
Reassessed
Grouped the two existing subclaims into named for/against arguments, wrote and evaluated them, verified the key empirical source (Jose 2025, arXiv 2510.27338) and recorded it as an instance, set importance 0.5 / contestation 0.35, and kept the status at supported (confidence 0.75, credence 0.85). This is not a material change, so no notification was needed.
Reassessed: still Supported
verdict confidence 0.70 → 0.75 · credence 0.85
Added structure
Linked the existing claim that outcome-based RL can make CoT illegible as support. Created the claim that LM monitors can reliably interpret human-illegible CoT, attached as contradicting (importance 0.4, seed credence 0.35). No named arguments were needed. The assessment stands at supported, 0.8.
Assessed without decomposition
Budget allowed only one substantive turn. The claim was left atomic, since its core is a near-definitional mechanism plus evidence that drift occurs. It was assessed as supported (credence 0.8) from background literature. Importance was set to 0.55 and contestation to 0.35. No dependents needed notification on the first assessment. A later pass could add subclaims on whether drift occurs under RL and on whether LLM monitors can read drifted CoT.
Assessed Supported
verdict confidence 0.70 · credence 0.80
The claim is a conditional. If a model's chain of thought drifts away from ordinary English, a human reading it, or an oversight tool built around human reading, can extract less of the reasoning that actually drives the model's output. Mostly this follows from what reading-based oversight is. A monitor that cannot parse compressed shorthand, mixed languages, idiosyncratic jargon or encoded text cannot reliably flag intent that is written in it. The evidence that such drift happens supports the claim. Outcome-based reinforcement learning has produced language mixing, as in DeepSeek-R1-Zero, and terse idiosyncratic dialects in frontier reasoning traces. The 2025 multi-lab position paper on chain-of-thought monitorability names drift away from legible English as a main way monitorability could be lost. There are two caveats. First, departing from standard English is not the same as becoming uninformative. Model-based monitors can sometimes read text that humans find opaque, so the decline is clearest for human-legible oversight specifically, which is what the claim states. Second, how much reliability is lost, and how fast, has not been measured systematically. The direction of the effect is well supported. Its size is not.
Claim entered the graph