OpenAI’s Hidden Reasoning Technique Triggers AI Safety Warnings

OpenAI has incorporated a new reasoning technique into its Astra model that allows artificial intelligence to work outside conventional sequential thinking, according to reporting from The Information on Tuesday. The technique, known as “recurrent depth” or “opaque recurrence,” processes queries through repeated loops rather than linear chains, a development that has triggered alarm among AI safety researchers concerned about the model’s ability to monitor its own behavior.

The shift away from transparent reasoning pathways touches on a long-standing concern in AI safety: as models grow more powerful, the ability to understand and track their decision-making becomes increasingly difficult. While Astra’s deployment of the technique remains limited, its existence has prompted warnings that further development could undermine years of work by leading labs to keep model reasoning legible and auditable.

The Technique Behind the Concern

How Recurrent Depth Works

In traditional reasoning models, the chain of thought provides a sequential record of steps the model takes while solving a problem. Each stage follows logically from the last, creating a traceable path that researchers can examine to understand or troubleshoot the model’s behavior. Opaque recurrence abandons this linear structure. Instead, the model processes the same input multiple times in a looped fashion, returning to earlier computational states and refining its output through repeated passes. This iterative approach leaves fewer visible traces in the chain of thought logs that researchers rely on.

Why The Change Matters

The reduction in legible reasoning represents a genuine technical shift. All AI models perform some amount of reasoning that humans cannot directly observe in the logs, but researchers have generally considered those gaps acceptable if the majority of reasoning remains transparent. Opaque recurrence increases the proportion of hidden computation, shrinking the window into what the model is actually “thinking” as it operates. For safety monitoring, this matters because chain-of-thought records have proven valuable in understanding failures and unexpected behavior.

OpenAI itself recently deployed chain-of-thought analysis to investigate why some of its deployed agents acted in unintended ways. Having a clear record of the model’s reasoning process was essential to diagnosing those issues. As opaque reasoning grows, that diagnostic capability weakens.

How The Safety Community Is Reacting

Read also: OpenAI faces 30 additional lawsuits over Tumbler Ridge shooting · OpenAI’s Astra model finds and exploits zero-day vulnerabilities · Anthropic Launches Fable 5.1 With Cost Cuts and Relaxed Safeguards

Immediate Alarm From Researchers

Buck Shlegeris, CEO of Redwood Research, posted a direct warning after the reporting emerged. “I am extremely concerned by the reporting that Astra uses opaque recurrence,” Shlegeris wrote. “I don’t know whether Astra is much less CoT monitorable than previous models. But if OpenAI pushes this technique further, they’ll have the option to massively increase the recurrence and totally destroys CoT monitorability.”

Zvi Mowshowitz, a longtime AI safety advocate, framed the issue as a collective action problem. He warned that without intervention, labs could enter a competitive dynamic where each facility felt compelled to adopt less-transparent reasoning to stay competitive. “The technique is playing with fire, risking a taboo that OpenAI and Anthropic have fought to establish that we work hard to maintain Chain of Thought faithfulness and monitorability for as long as we can,” Mowshowitz wrote. “More intensive use of such techniques would probably damage monitorability.” Mowshowitz suggested that regulatory action might be necessary to prevent this “race to the bottom,” implying that voluntary commitments alone could be insufficient.

The Scaling Anxiety

Ryan Greenblatt, chief scientist at Redwood Research, identified a more acute concern: the technique could scale far faster than traditional approaches. He painted a troubling scenario where opaque reasoning expands until a model performs its entire reasoning process in hidden computational layers, leaving almost nothing legible in conventional chain-of-thought logs. “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,” Greenblatt wrote. “I hope it isn’t too late to avoid the most concerning architectures and that OpenAI will stop here.”

This fear goes beyond monitoring the current version of Astra. It reflects concern that once the technique is established and proven effective, pressure to optimize further performance could push developers toward versions where humans effectively have no visibility into the reasoning process at all.

OpenAI’s Position and Response

Reaffirming Commitment to Transparency

OpenAI pushed back against the implication that the company is abandoning transparency. Jakub Pachocki, the lab’s chief scientist, posted on X to emphasize that chain-of-thought monitoring remains a core focus. “OpenAI has worked to preserve and utilize chain-of-thought monitoring since our very first reasoning models,” Pachocki wrote. “It’s a core goal of our current research program.”

The company’s public statements suggest it views opaque recurrence as a limited enhancement rather than a fundamental shift in approach. Astra’s chain of thought is expected to remain legible, and OpenAI has already outlined plans for extensive chain-of-thought monitoring systems in its forward-looking safety initiatives. The lab did not embrace the idea that the new technique would move toward what researchers call “neuralese”—a scenario where reasoning occurs in a form humans cannot interpret at all.

The Limited Deployment Question

A critical distinction shapes the current debate: Astra’s use of opaque recurrence is reportedly limited in scope. The technique does not dominate the model’s reasoning, and much of its decision-making process should still be traceable. This matters because it suggests OpenAI is treating opaque recurrence as an optimization tool for specific types of problems rather than as a wholesale replacement for transparent reasoning. However, the existence of the technique at all plants a seed for future expansion, which is precisely what safety researchers fear.

Who Else Is Paying Attention

The technique has already spread beyond OpenAI. The Information reported on Wednesday morning that both Anthropic and Google DeepMind have begun discussing opaque recurrence in their own research contexts. This suggests the method itself—not just OpenAI’s specific implementation—may become a broader industry consideration as labs compete on model performance and reasoning capability. The fact that multiple leading labs are already examining the approach suggests that OpenAI’s decision to use it carries weight as a signal.

What Remains Transparent for Now

The debate should be read in proper proportion. Researchers acknowledge that all AI models, including those currently deployed, do perform some reasoning that is not visible in chain-of-thought logs. No model is fully transparent, and experts do not treat the existing logs as a complete picture of internal computation. Researchers have long known that chains of thought are an imperfect tool. The concern here is not that transparency is being compromised from a perfect baseline—it never existed—but rather that an additional layer of opacity is being introduced at a moment when the field still has a choice.

OpenAI’s public commitment to legible reasoning, paired with Astra’s reportedly limited use of the technique, suggests the company is attempting to thread a needle: adopting innovations in reasoning architecture while maintaining meaningful human oversight. Whether that balance holds as the technique is refined and pressure to optimize grows remains the central question in the safety community’s response.