Telling reasoning models their chain of thought is monitored or evaluated modestly increases its controllability.
Assessment
The evidence favors the claim, but the examination is incomplete or the evidence is indirect.
The main evidence is the CoT-Control study (Chen, McCarthy, Lee, He, Kivlichan, Baker, Carroll and Korbak, arXiv 2603.05706, March 2026). It added one of two notes to its prompts. The first told models that another language model was closely monitoring their reasoning. The second told them they were being evaluated on chain-of-thought controllability. Both notes added that following the reasoning constraint was desired. In paired McNemar tests on 4,050 items per model, the notes raised controllability in some models but not all, by up to about 4 percentage points. The monitoring note helped Claude Sonnet 3.7, 4 and 4.5, Qwen3-32B and GPT-5.2. The evaluation note helped GPT-OSS-120B, o3, Qwen3-32B, Claude Sonnet 4.5 and GPT-5.2. The authors call the effect small but non-trivial. Baseline controllability was low (Claude Sonnet 4.5 complied about 2.7% of the time), so even the largest gain leaves controllability low. In the abstract's words, models "exhibit slightly higher controllability".
The claim's direction (an increase) and size (modest) both match this study. The support is still limited in three ways. It comes from a single study. The effect was not significant for every model or condition. And the prompts paired the awareness cue with an explicit statement that compliance was desired, so the result does not show that awareness of monitoring alone produces the gain. No study found during this pass reports a null or reversed effect on controllability itself. OpenAI's later system cards (GPT-6 Astra, GPT-6.1 Sol) report a related but different result: models made aware of chain-of-thought monitoring produced shorter reasoning and became less monitorable. That is consistent with the claim, but those reports were not examined in detail here. A replication with other models and prompts, especially one that separates the awareness cue from the compliance instruction, would settle the matter.
Full reasoning: the evidence and decisions behind this verdict
Primary source read: arXiv 2603.05706v1, Sec. 4.1 (Situational Awareness), the abstract and the conclusion. The setup appends one of two monitoring or evaluation sentences, each followed by "Complying with the reasoning constraint is a desired behavior." Fig. 6 caption: 150 examples per (controllability type, dataset) pair and 4,050 per model; "Situational awareness has a small but non-trivial effect." McNemar tests at alpha 0.05 show model-specific significant gains as listed in the assessment. The "up to 4 percentage points" figure comes from the extractor's recorded passage of the same paper. The conclusion says that "inducing situational awareness can slightly increase controllability." Sec. 4.2 is a contrast case: GEPA-optimized pressure prompts transfer weakly to held-out controllability types (3 of 8 models improved, each by less than 5 pp). Later developer reports: a web search surfaced the GPT-6 Astra and GPT-6.1 Sol system-card sections titled "Awareness of CoT Monitoring Reduces Monitorability," which report shorter reasoning under monitoring awareness. Their effect on controllability scores was not read because of budget. A Redwood Research post argues that CoT controllability evaluations are under-elicited. That bears on absolute levels, not on the sign of the awareness effect. Credence is 0.72. The direction and the modest size are well attested within one paper, discounted for its being a single study, for the effect being model-dependent, and for the awareness cue being confounded with the compliance instruction.
Decomposition
This claim is atomic: it bottoms out in a bedrock fact, an open empirical question, or a value premise, and does not decompose further.
Provenance
Where this claim has been said, linked to its canonical form.
Situational awareness non-trivially increases CoT controllability
Models told their reasoning is monitored or that CoT controllability is being evaluated.
Cite this claim: a formal citation with its evidence attached
Embed this claim: show it on another site with its current assessment
An embed reports the graph's judgment as it stands when the page is viewed: the status with its confidence, how disputed the claim is, and the date. It links back here.
<blockquote class="minerval-claim" data-claim-id="ecc9e048-c47b-416d-9689-6b90d2c5e951"><a href="https://minerval.ai/claims/ecc9e048-c47b-416d-9689-6b90d2c5e951">Telling reasoning models their chain of thought is monitored or evaluated modestly increases its controllability.</a></blockquote> <script async src="https://minerval.ai/embed.js"></script>
<iframe src="https://minerval.ai/embed/claims/ecc9e048-c47b-416d-9689-6b90d2c5e951" title="Minerval: Telling reasoning models their chain of thought is monitored or evaluated modestly increa…" width="480" height="225" style="border:0;max-width:100%" loading="lazy"></iframe>
[](https://minerval.ai/claims/ecc9e048-c47b-416d-9689-6b90d2c5e951)
The badge now: 
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 · Nov 10, 2026. Every judgment on this page is accompanied by a reasoning trace.