OpenAI's new reasoning technique worries AI safety experts

OpenAI’s new Astra model uses a reasoning technique called “recurrent depth,” or “opaque recurrence,” that lets it operate outside the sequential thinking that characterizes most reasoning models, The Information reported Tuesday. The technique will likely make the model’s chain of thought harder to monitor, and even though Astra’s use of it is reportedly limited, the disclosure has rattled AI safety researchers.

Redwood CEO Buck Shlegeris said he was extremely concerned by the reports that Astra uses opaque recurrence, warning that if OpenAI pushes the technique further, the lab would have the option to massively increase the recurrence and effectively destroy chain-of-thought monitorability. AI safety advocate Zvi Mowshowitz argued that laws might be needed to prevent a “race to the bottom” among AI labs, saying more intensive use of such techniques would probably damage the chain-of-thought faithfulness that OpenAI and Anthropic have worked to establish.

Under normal conditions, a reasoning model’s chain of thought lays out the sequential steps it takes to solve a problem — an imperfect but valuable tool for spotting misbehavior or misalignment, one that played a role in understanding OpenAI’s recent rogue agent activity. Opaque recurrence instead has the model process the same query repeatedly in a loop, leaving fewer legible traces and sidestepping a conventional chain-of-thought record.

OpenAI has pushed back on any suggestion it is shifting to “neuralese,” saying Astra’s chain of thought is still expected to be legible and pointing to its existing plans for extensive chain-of-thought monitoring systems; chief scientist Jakub Pachocki wrote on X that the company has worked to preserve and utilize chain-of-thought monitoring since its very first reasoning models, calling it a core goal of its current research program.

The concern isn’t limited to OpenAI: all AI models already do some amount of opaque reasoning, and in a follow-up report Wednesday, The Information said Anthropic and Google DeepMind were also already discussing the technique. Redwood Research chief scientist Ryan Greenblatt warned that opaque reasoning could scale faster than conventional chain-of-thought reasoning, eventually removing reasoning from visible channels altogether. “My biggest concern is that a natural progression from here would involve scaling up the opaque reasoning to the point where the model reasons entirely or almost entirely in latent space,” he wrote, adding that he hopes it isn’t too late to avoid the most concerning architectures.

Why it matters

The episode shows how a capability technique one lab uses only briefly can ripple across the industry once rivals like Anthropic and Google DeepMind start exploring it too, testing whether chain-of-thought monitoring — one of the few tools researchers have for catching AI misbehavior — can survive the next wave of architecture experiments.