As models' reasoning departs from standard English, human-legible chain of thought becomes less reliable for oversight.
Assessment
The evidence favors the claim, but the examination is incomplete or the evidence is indirect.
Oversight that relies on reading a model's chain of thought works only to the extent the reader can interpret it, so the claim's direction follows closely from how such oversight operates. The live question is whether drift from standard English actually happens and whether automated monitors make up for what humans lose. On the first, the evidence is fairly direct: outcome-based reinforcement learning can make some reasoning models' chains of thought illegible, as seen in language mixing in DeepSeek-R1-Zero and in a 2025 study of 14 reasoning models (Jose, NeurIPS 2025). That study found illegible passages that models used to reach correct answers, with legibility getting worse on harder questions.
The main counter-consideration is that language-model monitors might still interpret traces that humans cannot. The same 2025 study reports that the illegible reasoning was opaque to AI monitors as well as to humans, which weakens this escape route. Not all drift is equally damaging, though: cosmetic language mixing may leave the content recoverable, so how much reliability is lost depends on the kind of drift. The claim is well supported in direction. How large the effect is remains open, and a controlled measurement of monitor recall against degree of linguistic drift would settle it.
Full reasoning: the evidence and decisions behind this verdict
Own model before search: the claim is close to analytic for human readers (less legible text, less reliable reading-based oversight), so the empirical cruxes are whether drift occurs and whether LLM monitors offset it. Evidence: the DeepSeek-R1 report (2025) documents language mixing in R1-Zero. Korbak et al. 2025, "Chain of Thought Monitorability: A New and Fragile Opportunity", names drift from legible English under scaled RL as a threat. Jose, "Reasoning Models Sometimes Output Illegible Chains of Thought" (arxiv.org/abs/2510.27338), covers 14 models and finds RL "often causes reasoning to become illegible to both humans and AI monitors". In that study, accuracy fell 53% when models were forced to use only the legible portions, and legibility degraded on harder questions. This moved me against the machine-monitor counterargument. One caveat from the same study: resampling showed no correlation between legibility and performance, so the relationship is nuanced. Search results also surfaced a 2026 preprint, "Length Penalties Make Chain-of-Thought Less Monitorable" (arXiv 2607.09786), which is consistent with compression-driven drift reducing monitorability. I did not read it in full. The two related graph claims, that current CoT is understandable by default and that capability reduces CoT reliability, are both rated supported and cohere with this verdict. What would change it: evidence that LLM monitors keep high recall on heavily drifted traces, or that drift so far is purely cosmetic.
Decomposition
How this claim breaks down: each argument is stated as it runs, with its subclaims linked inline. ↗︎ opens a subclaim; the map shows how they fit together.
Oversight that works by reading a chain of thought depends on the reader understanding it, and text that departs from standard English (language mixing, compressed or idiosyncratic tokens) is harder for a human reader to interpret. Because outcome-based reinforcement learning can make some reasoning models' chains of thought illegible, such departure is a real trajectory rather than a hypothetical, so reading-based oversight becomes less reliable as it proceeds.
- Does it follow?
- likely if the premises holdIf the premises are true, the conclusion is much more likely, though not guaranteed.constitution →
- Do the premises hold?
- 1 supported
Granting that drift occurs, the conclusion follows for human readers almost by construction. It falls short of certain only because some drift, such as cosmetic language mixing, may leave the reasoning recoverable. The argument rests on RL-induced illegibility in some reasoning models, which is currently assessed as supported and backed by a multi-model study.
In practice chain-of-thought oversight is increasingly done by language-model monitors rather than human readers. If language-model monitors can reliably interpret chains of thought that humans find illegible, then departure from standard English need not reduce the reliability of chain-of-thought oversight, only its human readability.
- Does it follow?
- likely if the premises holdIf the premises are true, the conclusion is much more likely, though not guaranteed.constitution →
- Do the premises hold?
- 1 inconclusive
Granting its premise, departure from English would cost human readability but not monitoring reliability, so the claim's practical force would largely fail. The argument depends entirely on LLM monitors reliably interpreting human-illegible chains of thought. That premise is unassessed, and a 2025 study of 14 models found illegible traces opaque to AI monitors as well, which weighs against it.
Provenance
Where this claim has been said, linked to its canonical form.
These results suggest that without explicit optimization for legibility, outcome-based RL naturally produces models with increasingly opaque reasoning processes, potentially undermining monitoring approaches.
Study of CoT legibility across 14 reasoning models; finds RL-induced illegibility affecting both human and AI monitors.
While we rely on human-legible CoT for training, studying situational awareness, and demonstrating clear evidence of misalignment, our ability to rely on this degrades as models continue to depart from reasoning in standard English.
Abstract of the anti-scheming paper by Apollo Research and OpenAI.
Assessment history
0 status changes over 2 assessments. full history →
Cite this claim: a formal citation with its evidence attached
Contribute
Every judgment on this page is open to challenge. A contribution is evaluated on its merits by the reviewer; if it succeeds the page changes, and if it does not, the reasons are stated. Either way the exchange becomes part of the claim’s public record.
The attention this claim received was paid for by a funded mandate. Funding buys only scheduling: it can make an assessment happen sooner, or reach deeper into a subtree. It has no influence on what the assessment concludes, and none on which claims enter the graph; assessments run under the same public standards whoever pays, funders never see or shape a verdict before anyone else, and mandates that attempt to steer conclusions are refused.
Created by extractor · Oct 30, 2026. Every judgment on this page is accompanied by a reasoning trace.