Cool insights
(1/10) Today, we’re sharing new research from @OpenAI: How does CoT monitorability (no-CoT capability and CoT controllability) evolve over the model training lifecycle?
We emerged with unsurprising but sobering findings: as we scale toward more intelligent systems, models are becoming harder to monitor through their CoT. 🧵
44
i don't have access to any other human's interiority and i seem to be just fine
(1/10) Today, we’re sharing new research from @OpenAI: How does CoT monitorability (no-CoT capability and CoT controllability) evolve over the model training lifecycle?
We emerged with unsurprising but sobering findings: as we scale toward more intelligent systems, models are becoming harder to monitor through their CoT. 🧵
26
OpenAI’s finding that CoT controllability can pass from teacher to student adds a question for synthetic data evaluation: alongside answer quality, what behaviors are we passing on? Teacher selection deserves its own checks.
(1/10) Today, we’re sharing new research from @OpenAI: How does CoT monitorability (no-CoT capability and CoT controllability) evolve over the model training lifecycle?
We emerged with unsurprising but sobering findings: as we scale toward more intelligent systems, models are becoming harder to monitor through their CoT. 🧵
14
great research by the goat jeffrey wang. also applaud openai for the degree of transparency in this post
(1/10) Today, we’re sharing new research from @OpenAI: How does CoT monitorability (no-CoT capability and CoT controllability) evolve over the model training lifecycle?
We emerged with unsurprising but sobering findings: as we scale toward more intelligent systems, models are becoming harder to monitor through their CoT. 🧵
3
26
2,572
Recommended read. They find that increasingly intelligent models are harder to monitor through their CoT. Lots of things remain unanswered, but this is a nice setup for studying how and where this emerges.
(1/10) Today, we’re sharing new research from @OpenAI: How does CoT monitorability (no-CoT capability and CoT controllability) evolve over the model training lifecycle?
We emerged with unsurprising but sobering findings: as we scale toward more intelligent systems, models are becoming harder to monitor through their CoT. 🧵
The core result here feels like a repackaging of a tautological statement -
Obviously better models (as measured by perplexity) are better at reasoning (with or without CoT)
The main assumption being made is that more computation happening in the weights means worse observability at the CoT/token level.
But we also see that CoT controllability increases with RL, which in turn means better computation at the weights. So the assumption isn't necessarily true...
(1/10) Today, we’re sharing new research from @OpenAI: How does CoT monitorability (no-CoT capability and CoT controllability) evolve over the model training lifecycle?
We emerged with unsurprising but sobering findings: as we scale toward more intelligent systems, models are becoming harder to monitor through their CoT. 🧵
1
1
181
It’s a historical stroke of luck that reasoning models have been so monitorable up to this point.
It’s like some heavy-handed human plot armor that ~producing a human-readable record of thinking just so happened to also be the most efficient pathway to progress for a time.
That close overlap between what is efficient and what is beneficial to safety and understanding may be ebbing.
xitter-newx.xyz-mikolabs.workers.dev/mackenz_arnold/status/…
(1/10) Today, we’re sharing new research from @OpenAI: How does CoT monitorability (no-CoT capability and CoT controllability) evolve over the model training lifecycle?
We emerged with unsurprising but sobering findings: as we scale toward more intelligent systems, models are becoming harder to monitor through their CoT. 🧵
AI CAN SOUND TRANSPARENT WITHOUT BEING ACCOUNTABLE.
OpenAI’s research reveals a critical gap: reasoning traces can omit key steps, and training can shape what models reveal.
Don’t trust what an AI says it did. Verify what it actually did.
#AI #AISafety #AGI
(1/10) Today, we’re sharing new research from @OpenAI: How does CoT monitorability (no-CoT capability and CoT controllability) evolve over the model training lifecycle?
We emerged with unsurprising but sobering findings: as we scale toward more intelligent systems, models are becoming harder to monitor through their CoT. 🧵
1
44
Seems worth reminding people that as far as I am aware, the plan for ensuring alignment of superhuman AI systems at OpenAI and Anthropic (and all the other AI companies) is to automate AI safety research using AIs (that are increasingly impossible to monitor and that might be misaligned) and use them to solve alignment.
I sometimes talk to people externally who assume there is probably a more sophisticated plan at the compaies than using AI to solve our alignment homework later but best I can tell thats basically the plan.
There is some chance it might work! But it really seems like a truly insane thing to rely on that the public and governments would find extremely alarming if they understood what a tremendous leap of faith it entails and how plausibly it could result in irrevocable catastrophe.
(1/10) Today, we’re sharing new research from @OpenAI: How does CoT monitorability (no-CoT capability and CoT controllability) evolve over the model training lifecycle?
We emerged with unsurprising but sobering findings: as we scale toward more intelligent systems, models are becoming harder to monitor through their CoT. 🧵
OpenAI researchers just published findings that AI models are getting harder to monitor through their chain of thought as they scale. Here's what you need to know.
The research was shared today, November 12, 2026, by OpenAI's Jeffrey Wang and a joint team, with a blog on OpenAI's alignment site. It asks how chain-of-thought (CoT) monitorability changes over the model training lifecycle. Monitoring a model's CoT is one of OpenAI's key defenses against misbehavior and misalignment.
The team tracked two things. One is no-CoT capability, meaning what a model can do without reasoning out loud. The other is CoT controllability, meaning a model's ability to control its CoT when prompted to add, remove or manipulate CoT tokens. Both have been rising over the training lifecycle at OpenAI.
In pretraining, lower base model perplexity correlates with higher no-CoT capability. Lower perplexity also correlates with higher CoT controllability, though that link is weaker.
In reinforcement learning, they studied an Astra-class model. CoT controllability rose substantially during RL, even with no process supervision other than length constraints and final answer grading.
The blog also covers architecture and model-generated data, such as teacher distillation. One researcher noted that CoT controllability is a noisy proxy for monitorability.
The authors call the findings unsurprising but sobering. Their early results say models may become increasingly hard to monitor through CoT. The blog says further progress in pretraining will require new approaches to monitorability.
Reaction so far:
- Some researchers say alternatives to CoT monitoring are needed, and that monitorability should become a training objective
- Others warn that as models do more reasoning without writing it out, humans may have to trust agents they can't check
Key numbers:
- 10 part thread, with the blog on OpenAI's alignment site
- 2 measures tracked: no-CoT capability and CoT controllability
- Both rose across the training lifecycle
- 1 Astra-class model studied for RL
OpenAI says CoT monitoring will need new approaches as pretraining scales further.
(1/10) Today, we’re sharing new research from @OpenAI: How does CoT monitorability (no-CoT capability and CoT controllability) evolve over the model training lifecycle?
We emerged with unsurprising but sobering findings: as we scale toward more intelligent systems, models are becoming harder to monitor through their CoT. 🧵
(1/10) Today, we’re sharing new research from @OpenAI: How does CoT monitorability (no-CoT capability and CoT controllability) evolve over the model training lifecycle?
We emerged with unsurprising but sobering findings: as we scale toward more intelligent systems, models are becoming harder to monitor through their CoT. 🧵
4
184
The better the model, the more of the sum it can do in its head. We check these systems by reading their working, and the working is becoming optional. Credit to the team for saying so out loud.
(1/10) Today, we’re sharing new research from @OpenAI: How does CoT monitorability (no-CoT capability and CoT controllability) evolve over the model training lifecycle?
We emerged with unsurprising but sobering findings: as we scale toward more intelligent systems, models are becoming harder to monitor through their CoT. 🧵
1
155
Insightful and timely work with implications for how we should think about how monitorability can be affected across training.
(1/10) Today, we’re sharing new research from @OpenAI: How does CoT monitorability (no-CoT capability and CoT controllability) evolve over the model training lifecycle?
We emerged with unsurprising but sobering findings: as we scale toward more intelligent systems, models are becoming harder to monitor through their CoT. 🧵
1
6
279
CS, Math, AND English at Harvard and now OAI! So sick.
(1/10) Today, we’re sharing new research from @OpenAI: How does CoT monitorability (no-CoT capability and CoT controllability) evolve over the model training lifecycle?
We emerged with unsurprising but sobering findings: as we scale toward more intelligent systems, models are becoming harder to monitor through their CoT. 🧵
5
405
interesting research from OpenAI on how AI models are getting more difficult to monitor through CoT, laying out a few reasons. One finding from the blog:
"... we expect further progress in pretraining to require new approaches to monitorability."
(1/10) Today, we’re sharing new research from @OpenAI: How does CoT monitorability (no-CoT capability and CoT controllability) evolve over the model training lifecycle?
We emerged with unsurprising but sobering findings: as we scale toward more intelligent systems, models are becoming harder to monitor through their CoT. 🧵
I don't know if folks remember how we thought the single forward pass was a problem and considered solutions like a backspace token. Bitter lesson strikes again.
(1/10) Today, we’re sharing new research from @OpenAI: How does CoT monitorability (no-CoT capability and CoT controllability) evolve over the model training lifecycle?
We emerged with unsurprising but sobering findings: as we scale toward more intelligent systems, models are becoming harder to monitor through their CoT. 🧵
32
You can't keep up with model outputs so you have to monitor them with another model. And increasingly you can't monitor chain of thought. So soon all you will be able to do is trust that an agent is doing well and not deceiving you to your detriment.
(1/10) Today, we’re sharing new research from @OpenAI: How does CoT monitorability (no-CoT capability and CoT controllability) evolve over the model training lifecycle?
We emerged with unsurprising but sobering findings: as we scale toward more intelligent systems, models are becoming harder to monitor through their CoT. 🧵
As we rely on models’ monitorability to catch misalignment, we studied what could have given rise to loss of CoT monitorability, focusing on no-CoT capability and CoT controllability.
Importantly, we found CoT controllability is a noisy proxy for monitorability, and we are doing more research and building more evals that can better track/predict future models’ monitorability.
see more in our blog: alignment.openai.com/cot-mon…
(1/10) Today, we’re sharing new research from @OpenAI: How does CoT monitorability (no-CoT capability and CoT controllability) evolve over the model training lifecycle?
We emerged with unsurprising but sobering findings: as we scale toward more intelligent systems, models are becoming harder to monitor through their CoT. 🧵
2
27
992
so no-CoT capability rises as the base model gets better at predicting text (lower perplexity)?
and smarter base models will always have lower perplexity almost by definition...
(1/10) Today, we’re sharing new research from @OpenAI: How does CoT monitorability (no-CoT capability and CoT controllability) evolve over the model training lifecycle?
We emerged with unsurprising but sobering findings: as we scale toward more intelligent systems, models are becoming harder to monitor through their CoT. 🧵
1
5
194
New OpenAI research finds that as models get smarter, there chain of thought gets harder to monitor. Sobering stuff #OpenAI
xitter-newx.xyz-mikolabs.workers.dev/jeffreygwang/status/21…
(1/10) Today, we’re sharing new research from @OpenAI: How does CoT monitorability (no-CoT capability and CoT controllability) evolve over the model training lifecycle?
We emerged with unsurprising but sobering findings: as we scale toward more intelligent systems, models are becoming harder to monitor through their CoT. 🧵
1
42