Chain-of-thought monitoring is not a perfect safety net. But right now, it’s one of the only ones we have. So when a new AI reasoning technique threatens to make that monitoring significantly harder, “we’re still committed to transparency” is not the answer safety researchers want to hear.
According to TechCrunch, OpenAI’s upcoming Astra model uses a reasoning approach called “recurrent depth,” also known as “opaque recurrence.” Instead of processing a query in a linear, step-by-step sequence, the model loops through the same query multiple times. The result is fewer readable traces of how it reached its conclusions. That’s a problem for anyone trying to audit what the model is actually doing inside its own head.
Under normal conditions, a reasoning model’s chain of thought gives researchers and developers a sequential record of the steps the model took to solve a problem. It’s imperfect, and no serious researcher treats it as a direct window into model cognition. But it still matters. When OpenAI’s agents recently started behaving unexpectedly, chain-of-thought logs were a key tool in figuring out why. Opaque recurrence quietly sidesteps that record.
The safety community’s reaction has been sharp. Redwood Research CEO Buck Shlegeris said he is “extremely concerned” by the reports. His worry isn’t just about Astra’s current, apparently limited use of the technique. It’s about where this goes next. “If OpenAI pushes this technique further,” he wrote, “they’ll have the option to massively increase the recurrence and totally destroy CoT monitorability.” Redwood Research chief scientist Ryan Greenblatt went further, warning that opaque reasoning could scale faster than conventional chain-of-thought reasoning, eventually pushing all model reasoning into what he called “latent space,” invisible and unreadable to outside observers.
AI safety writer Zvi Mowshowitz argued that legislation might be the only way to prevent a race among labs to adopt techniques that erode safety norms. He described opaque recurrence as “playing with fire,” specifically because it threatens a norm that OpenAI and Anthropic have both worked to establish around chain-of-thought faithfulness.
OpenAI’s response has been a mix of reassurance and deflection. Chief scientist Jakub Pachocki posted on X that the company has prioritized chain-of-thought monitoring since its first reasoning models and called it a “core goal” of current research. The company also pushed back against the idea that it would move toward fully opaque “neuralese” reasoning. And yes, OpenAI has announced plans for extensive chain-of-thought monitoring as part of its safety roadmap.
But the broader context makes those reassurances harder to take at face value. A follow-up report noted that both Anthropic and Google DeepMind are already discussing the technique internally. That’s how norms erode. One lab adopts something at the margins, frames it as limited and controlled, and then the rest of the industry starts treating it as standard practice. The competitive pressure to ship faster, reason better, and outperform rivals does not naturally favor slowing down to preserve safety tooling that inconveniences no one until something goes wrong.
The core tension here is not technical. It’s about who gets to watch the machines and how. Opaque recurrence, even in limited form, shifts more of the model’s activity out of readable channels. For users and regulators trying to understand AI behavior, that’s a step in the wrong direction, regardless of what any lab’s safety roadmap says.