OpenAI’s new reasoning technique alarms AI safety experts

OpenAI’s upcoming Astra model is expected to use a reasoning technique that could make its thinking harder to monitor. The approach, known as “recurrent depth” or “opaque recurrence,” processes queries through repeated internal loops rather than relying entirely on the sequential reasoning traces used by most reasoning models.
The Information first reported Astra’s use of the technique. Although its reported application is limited, AI safety researchers are concerned that broader use could make models’ reasoning increasingly invisible to conventional monitoring systems.

Why chain-of-thought monitoring matters

A reasoning model’s chain of thought records the sequential steps it takes while working through a problem. Researchers do not consider these records a perfect transcript of a model’s internal reasoning, but they can still provide useful clues when investigating misbehavior or possible misalignment.
The source material notes that chain-of-thought records played an important role in understanding recent rogue agent activity. If models increasingly reason in ways that leave fewer legible traces, safety researchers could have less visibility into why an AI system produced a particular result or took a particular action.
Redwood CEO Buck Shlegeris said he was “extremely concerned” by reports that Astra uses opaque recurrence. He warned that increasing the amount of recurrence could “totally destroy” chain-of-thought monitorability.

How opaque recurrence works

Under conventional reasoning, a model’s work appears as a sequence of steps. Opaque recurrence instead allows the model to process the same query multiple times in a loop, using a less linear approach.
That process produces fewer readable traces and can effectively bypass a conventional chain-of-thought record. Redwood Research chief scientist Ryan Greenblatt said the technique could scale faster than ordinary chain-of-thought reasoning, potentially moving more of a model’s reasoning into latent space rather than visible channels.
Greenblatt said his biggest concern was a progression in which models reason “entirely or almost entirely in latent space.” Longtime AI safety advocate Zvi Mowshowitz similarly argued that laws might eventually be needed to prevent an AI lab “race to the bottom” on reasoning monitorability.

OpenAI says Astra will remain legible

The concerns are tempered by reports that Astra’s use of opaque recurrence will be limited. Its chain of thought is still expected to remain legible, and OpenAI pushed back against suggestions that the model would shift to “neuralese.”
OpenAI chief scientist Jakub Pachocki emphasized the company’s stated commitment to preserving chain-of-thought monitoring. “OpenAI has worked to preserve and utilize chain-of-thought monitoring since our very first reasoning models,” he wrote on X, calling it a core goal of the company’s research program.
OpenAI has also announced plans for extensive chain-of-thought monitoring systems as part of its forward-looking safety work. Still, the emergence of opaque recurrence has prompted concern beyond OpenAI: The Information reported that Anthropic and Google DeepMind were discussing the technique as well.
The central question for safety researchers is not whether all opaque reasoning is inherently unsafe—AI models already perform some reasoning opaquely—but whether expanded use could outpace the tools designed to monitor it. As recurrent techniques become more prominent, preserving meaningful visibility into model behavior may become increasingly difficult.