Chain-of-thought monitoring depends on models externalizing legible reasoning, which future models may stop doing.
Structured and assessed
First pass. Linked two existing supported claims as supports (training degrades monitorability; non-English drift reduces legible-CoT reliability), two related links (necessity forces externalization; LLM monitors reading illegible CoT). Recorded Korbak et al. 2025 as affirming instance. Assessed supported, credence 0.85; no named arguments needed.
Assessed Supported
verdict confidence 0.80 · credence 0.85
The claim has two parts. The first, that chain-of-thought monitoring depends on models externalizing legible reasoning, is close to true by construction: a monitor that reads a reasoning trace can only catch what the trace shows, and evidence that oversight becomes less reliable as reasoning departs from standard English fits this. The dependence is softened somewhat if language-model monitors can read traces humans find illegible, but that remains unsettled. The second part claims only a possibility: that future models may stop externalizing legible reasoning. The evidence supports this. A 2025 position paper by Korbak and some forty co-authors across the major AI laboratories calls monitorability "fragile" and names the routes by which it could be lost: latent-reasoning architectures that reason in continuous activations, optimization pressure applied directly to traces, and drift away from natural language under outcome-based reinforcement learning. There is also evidence that current training practices can already degrade monitorability. Against a high likelihood of loss, tasks demanding long serial reasoning appear to force externalization in current transformer architectures, and deployed frontier models still reason in readable text. That bears on how probable and how soon the loss is, not on whether it is possible, so the claim as worded is well supported. How likely loss actually is remains the open question.
Claim entered the graph