# OpenAI: what training choices do to chain-of-thought monitoring

> OpenAI measures how pretraining, architecture and reinforcement learning change whether a model's chain of thought stays readable to a monitor.

- **Severity:** Info: Research, tools, evals or lab claims with no direct exposure. Context, not an alert.
- **Category:** Lab safety, Research
- **Published by:** OpenAI
- **Disclosed:** 2026-11-10
- **Affects:** GPT-6 Astra
- **Research by:** Jeffrey G. Wang, Aiden Low, Zora Che, Holden Mui, Harshit Sikchi, OpenAI Alignment Research
- **Primary source:** https://alignment.openai.com/cot-monitorability-model-lifecycle/
- **Page:** https://theinjection.dev/items/openai-cot-monitorability-lifecycle/

OpenAI reports that a model's ability to reason without a visible chain of thought tracks pretraining perplexity, and that control over its own chain of thought rises during reinforcement learning and carries into models distilled from it.

## What happened

OpenAI's alignment research team published a study of chain-of-thought monitorability on 10 November 2026, written by Jeffrey G. Wang and colleagues. The post asks which stages of the model lifecycle, meaning pretraining, architecture, reinforcement learning and training on model-generated data, erode a monitor's ability to read a model's reasoning.

## How it works

Monitorability is measured through two proxies. No-CoT capability is the reasoning a model completes in a forward pass with no chain-of-thought or tool-call tokens, so a model that scores high does more work a monitor never sees. CoT controllability measures how well a model can add, remove or manipulate its chain of thought when prompted to. OpenAI reports that no-CoT capability improves smoothly as base-model perplexity falls, and that controllability rises substantially over reinforcement learning even with no process supervision beyond length constraints and final-answer grading.

## Who is affected

Anyone who treats a chain of thought as an oversight signal is the audience, including teams that read agent reasoning traces to catch misbehavior. OpenAI reports that controllability can be distilled: training a student on rollouts from a more controllable teacher produces a more controllable student.

## What to do

Treat chain-of-thought monitoring as one oversight signal among several rather than the only one. OpenAI calls controllability a noisy, imperfect proxy, and reports an iso-capability comparison in which the more controllable model did not show more monitor evasion, so a rise in controllability is a warning sign and not proof of evasion.

## Quick facts

- **Proxies measured:** No-CoT capability and CoT controllability
- **Lifecycle stages:** Pretraining, architecture, RL, model-generated data
- **GPT-6 Astra:** Not off-trend for no-CoT capability against perplexity
- **Time horizon:** 30.9 minute median, reported by UK AISI

## Questions

### Does reinforcement learning make a chain of thought harder to monitor?

Reinforcement learning raised CoT controllability substantially in OpenAI's measurements, even with no process supervision beyond length constraints and final-answer grading. Higher controllability means a model can more readily add, remove or reshape its reasoning tokens on request. OpenAI treats that as a warning sign for monitoring rather than evidence of evasion, and calls the proxy noisy and imperfect.

### Can chain-of-thought monitorability be inherited from another model?

Controllability can be distilled, according to OpenAI's results: training on rollouts from a more controllable teacher produces a more controllable student. Teams that distil from a frontier model therefore inherit some of that model's control over its own reasoning traces. A monitorability check on the teacher is not a substitute for checking the student that was trained on its output.

### Was GPT-6 Astra unusual on OpenAI's monitorability measures?

GPT-6 Astra is not off-trend with respect to perplexity on no-CoT capability, OpenAI states. On a separate measure, UK AISI reported a time horizon median of 30.9 minutes for the model, against a previously projected median of around 7 minutes by the end of 2028. UK AISI notes the estimate could be inflated by contamination.

## Sources

1. [OpenAI Research Post](https://alignment.openai.com/cot-monitorability-model-lifecycle/)
2. [UK AISI Evaluation](https://deploymentsafety.openai.com/gpt-6-astra/external-evaluation-for-monitorability---uk-aisi)