OpenAI's New 'Recurrent Depth' Technique Raises Chain-of-Thought Monitoring Concerns

News Summary

OpenAI's next-generation model, reportedly code-named Astra, is drawing sharp criticism from AI safety researchers after reports surfaced that it relies on a reasoning technique known as "recurrent depth," which reuses internal computation layers in a loop rather than producing the fully sequential, human-readable chain of thought that has become a standard tool for monitoring advanced AI systems. The reports, first detailed by The Information on the morning of September 2, 2026 (Eastern Time), and later corroborated by TechCrunch and other outlets, describe the approach as a meaningful departure from how leading reasoning models have operated to date, and safety researchers say it could make it substantially harder to detect when a model is behaving in unintended or deceptive ways.

What Is Recurrent Depth

Recurrent depth, sometimes called "opaque recurrence" or described as a "looped transformer" architecture, allows a model to reprocess the same query multiple times through a shared core set of internal layers instead of moving strictly forward through a linear sequence of reasoning steps. Proponents say the technique can improve a model's performance and computational efficiency without requiring additional parameters or memory, since the same layers are reused repeatedly rather than the model needing an ever-larger stack of unique layers to gain reasoning depth.

The tradeoff, according to researchers who reviewed early technical details, is that much of this repeated internal processing happens in the model's hidden states rather than being externalized as readable text. Traditional reasoning models, including OpenAI's earlier "o-series" systems, produce a chain of thought: a step-by-step written trace of how the model works through a problem before producing a final answer. That trace, while imperfect, has given researchers a way to peer into a model's decision-making process and flag signs of misalignment, deception, or unsafe reasoning before they show up in a user-facing response.

Why Safety Researchers Are Concerned

The central worry is not necessarily about Astra's current behavior, since multiple reports indicate its use of the technique so far appears limited and its chain of thought remains largely legible. Instead, researchers are alarmed by where the technique could lead if adopted more aggressively in future models.

Buck Shlegeris, CEO of AI safety research organization Redwood Research, said that if OpenAI expands the use of recurrent depth, the company would gain the option to "massively increase the recurrence," which he warned "totally destroys chain-of-thought monitorability." Ryan Greenblatt, Redwood's chief scientist, echoed that concern, saying opaque reasoning techniques could eventually expand to the point where models reason almost entirely within hidden internal channels rather than in any externally observable form.

AI commentator Zvi Mowshowitz described the shift as "playing with fire," arguing that maintaining faithful, monitorable chains of thought has been treated as something close to an industry norm, one that both OpenAI and Anthropic have previously said they work hard to uphold. He suggested that if labs began moving away from that norm without clear safeguards, regulation might ultimately be needed to preserve monitorability as models become more capable.

Independent AI researcher Gary Marcus went further, publishing a piece describing the development as crossing what he called an AI safety "redline." He characterized chain-of-thought monitoring as one of the few reliable tools available for observing the internal workings of otherwise opaque large language models, and argued that eroding it, even gradually, removes a safeguard the field has come to depend on.

How OpenAI Has Responded

OpenAI's chief scientist, Jakub Pachocki, pushed back on characterizations of the technique as a wholesale departure from the company's safety commitments. He noted that Astra's effective computation depth remains close to that of GPT-4-era models and said OpenAI continues to prioritize chain-of-thought monitoring as a source of safety insight into how its models reach conclusions. Pachocki's comments suggest that, from OpenAI's perspective, the current implementation of recurrent depth is a targeted efficiency improvement rather than an attempt to obscure model reasoning from oversight.

That framing has not fully satisfied outside researchers, who note that even a limited or experimental use of the technique establishes a precedent. Their concern is less about Astra as it exists today and more about the incentive structure it creates: if recurrent depth proves effective at improving performance and lowering computational cost, competitive pressure could push OpenAI or other labs to expand its use in future models, potentially at the expense of transparency.

Broader Context for AI Oversight

Chain-of-thought monitoring has emerged over the past few years as one of the more practical tools available to researchers trying to understand and audit the behavior of large reasoning models, alongside interpretability research and red-teaming. Because these models are trained on vast datasets and arrive at answers through processes that are not fully understood even by their creators, a legible reasoning trace has offered a rare, if imperfect, window into how a system reaches its conclusions.

Safety researchers across multiple organizations, including Redwood Research and independent commentators, have previously called on frontier AI labs to preserve and strengthen chain-of-thought faithfulness as models grow more capable, rather than trading it away for performance gains. The debate around Astra's recurrent depth technique reflects a broader tension in the field between the drive to build faster, cheaper, and more capable models and the desire to keep those models' reasoning processes open to inspection. How OpenAI and its peers navigate that tension in future model releases is likely to remain a closely watched issue among researchers focused on AI safety and governance.