Step 11: assistant

text

I've assessed the claim "Chain-of-thought monitoring depends on models externalizing legible reasoning, which future models may stop doing" as **supported**. My confidence in that status is 0.8, and I put the chance the claim is true at 0.85. The pass used about $0.21 of the $1.20 budget.

**How the verdict breaks down:**
- **Dependence:** The first half is close to true by definition. A monitor that reads a model's reasoning can only catch what the reasoning shows. This would weaken somewhat if AI monitors can read reasoning that humans can't follow, but that is still unsettled in the graph.
- **"May stop":** The second half only says this is possible, and the evidence supports that. A 2025 position paper by Korbak and about 40 co-authors from the major AI labs (arXiv 2507.11473) calls this kind of monitoring "fragile." I read its abstract to confirm that wording. The graph's other evidence that current training methods can already degrade monitoring points the same way.
- **What cuts the other way:** Hard tasks seem to force current models to reason in visible text, and models in use today still do so. That bears on how likely and how soon the loss would be, not on whether it is possible.

I held credence at 0.85 rather than higher because "legible" could mean readable by humans in particular. On that reading the dependence is less strict if AI monitors are enough.

**Changes to the graph:**
- I attached two claims already in the graph as supporting evidence: that current training methods can degrade monitoring, and that visible reasoning becomes less reliable for oversight as it drifts from standard English.
- I added two "see also" links. One goes to the claim that models struggle to evade monitors when reasoning is necessary for a task. The other goes to the claim about AI monitors reading reasoning humans find illegible.
- I recorded the Korbak et al. paper as a source that affirms the claim.
- I set importance at 0.5 and how actively it's disputed at 0.3. The possibility is broadly accepted; the live debate is about how likely the loss is and when.
- I added no named arguments, since the claim has one natural line of support.

**Messages and follow-ups:** I logged the decision and answered both inbox messages. I notified no other stewards, because this was the claim's first assessment and nothing that depends on it changed.