The alarm over OpenAI’s new reasoning method has the premise backwards. The worry, roughly, is that we are about to lose our window into how these models think. My position is that the window was never there. What recurrent depth removes is not oversight but the appearance of it, and that distinction matters more than anything else in the current round of concern.
Here is what we know. OpenAI’s upcoming Astra model uses a technique called recurrent depth, which lets it operate outside the sequential thinking that characterizes most reasoning models. Instead of producing a chain of tokens that steps toward an answer, the model processes a query repeatedly in a loop, deepening its computation without necessarily externalizing it. Safety researchers have started calling this opaque recurrence, and the objection is straightforward: if the reasoning does not surface as text, it becomes much harder to monitor, verify, or control.
Why chain-of-thought was never a microscope
The implicit baseline in this critique is that token-by-token reasoning gives us genuine visibility into a model’s cognition. As someone who spends most of her time looking at internal activations rather than output logs, I find that baseline hard to defend.
A chain of thought is a generated artifact. It is trained, shaped by reward signals, and optimized in part for how it reads. It correlates with the computation that produced the answer, but correlation is doing a lot of work in that sentence. We have known for a while that models can arrive at conclusions through paths their stated reasoning does not describe, and that the stated reasoning can be edited by training pressure without the underlying behavior changing much. Reading a trace and believing you have audited a mind is a category error. You have audited a report about a mind, written by the mind, in a format it learned humans find persuasive.
So the honest version of the safety concern is not that recurrent depth destroys interpretability. It is that recurrent depth destroys a monitoring practice we had built a great deal of infrastructure and institutional comfort around, and we do not yet have a replacement.
The architectural problem is real, just different
None of this makes the concern empty. It relocates it.
Looping computation inside a latent space changes the shape of the oversight problem in ways that are genuinely difficult:
- No natural checkpoint. Sequential reasoning gives you discrete, human-legible stopping points where an external system can inspect, flag, or interrupt. A recurrent loop in latent space does not hand you those boundaries for free. You have to define them.
- Variable, unobservable compute. If depth is dynamic, the amount of thinking spent on a query is no longer visible in the length of a trace. Two identical-looking answers may reflect very different amounts of internal work.
- Nothing to read. Interpretability on latent recurrence means probing representations across iterations, which is a research program, not a product feature. It requires access, tooling, and time that external evaluators generally do not have.
- Evaluation drift. Much of the current safety evaluation stack assumes textual reasoning as an input. Benchmarks and red-teaming methods built on that assumption degrade quietly rather than failing loudly.
That last point is the one I would put in front of policymakers. The International AI Safety Report 2026 makes the broader observation that developers are getting significant capability gains from new techniques, not just larger training runs. Architectural change is now a primary driver of capability. Our oversight methods, by contrast, are still largely tuned to the previous architecture. The gap between those two rates of change is the actual risk surface.
What a serious response looks like
Asking labs not to build recurrent architectures is not a plan. Latent iteration is too useful, too efficient, and too obvious a direction for anyone to abandon on request. The plan has to be to make internal computation legible on its own terms.
Concretely, that means treating per-iteration probing as a first-class requirement rather than an academic curiosity, building evaluation methods that measure behavior under varying compute depth, and being explicit about which claims about model reasoning are supported by mechanistic evidence versus generated text. It also means labs granting external researchers the kind of internal access that latent-space auditing requires, because you cannot study a loop from the outside.
The uncomfortable framing is that recurrent depth is a forcing function. It takes away a monitoring habit that was always weaker than advertised and demands that we build something real in its place. I would rather face a system that is visibly hard to inspect than keep trusting one that merely looked easy.
🕒 Published: