AI safety experts are raising alarms about how OpenAI has built its soon-to-be-released frontier AI model, Astra, saying the approach may hasten the day when humans lose the ability to monitor the reasoning that AI agents use.

For the new model, OpenAI has employed a method alternately referred to as “recurrent depth” or “looped Transformers” for a portion of the model’s internal architecture. The technique can make AI models considerably more efficient by requiring less computing power to process each prompt—a valuable feature at a time when many businesses are complaining about the high costs of using the most advanced frontier AI models.

The new process, however, also means that part of the model’s “chain of thought”—the reasoning steps it takes—is not expressed in natural language, making it far more difficult for humans to monitor what the model is doing and why. Chain-of-thought monitoring is currently one of the main methods companies use to ensure AI agents are not taking unintended or unauthorized actions, and it has become a central focus of the field of AI interpretability, in which researchers at major labs—including OpenAI, Anthropic, and Google DeepMind—study whether a model’s visible reasoning actually reflects its internal computations. Researchers have cautioned that chain-of-thought text is not always “faithful” to a model’s true internal process, which makes preserving any legible trace of reasoning especially valuable to safety teams.

Tech publication The Information first reported on OpenAI’s use of recurrent depth in Astra earlier this week. Jakub Pachoki, OpenAI’s chief scientist, and several other OpenAI researchers criticized the publication for sparking undue alarm among AI safety researchers, saying the company had limited the extent to which the looped transformer architecture is used so that the model’s reasoning remains legible.

In response to The Information’s report, Pachoki wrote on X that “we care deeply” about chain-of-thought monitoring and that “OpenAI has worked to preserve and utilize chain-of-thought monitoring since our very first reasoning models.” He said OpenAI would share more details of Astra’s architecture in the future.

Pachoki added that he thought chain-of-thought monitoring could grow more challenging, but that this would be “for reasons not contingent on architecture changes” such as what OpenAI has begun implementing with recurrent depth. “But there are things we can do to strengthen it, and it’s a core goal of our current research program,” he wrote.

Prior to Pachoki’s statement, Steven Adler, a former OpenAI safety researcher who now runs Guidelight AI Standards, a nonprofit organization working on AI safety benchmarks and standards, wrote on X that, if The Information’s report was true, “OpenAI seems to be violating one of the few redlines that exists in the AI industry.”

Peter Wildeford, policy director at the AI Policy Network, a Washington, D.C. think tank dedicated to advancing policies to help Americans prepare for the impact of AI, told Fortune that OpenAI’s use of recurrent depth was “potentially very concerning” and “potentially reckless.” He noted that one of the only ways OpenAI and outside AI evaluation companies were able to piece together what happened during the July incident—in which several of OpenAI’s AI models autonomously attacked the company Hugging Face—was by reading the models’ chains of thought. OpenAI has also said it is investing significant resources in real-time chain-of-thought monitoring going forward as part of its effort to prevent future rogue AI incidents. “If OpenAI is indeed shifting away from this, this is the wrong direction,” he said.

Several AI safety experts said they were alarmed by OpenAI’s use of looped Transformers in Astra not necessarily because it made Astra’s own reasoning harder to monitor. Rather, they feared the move would normalize the technique, setting a precedent that other AI companies are likely to follow and expand upon, eventually resulting in AI models whose reasoning steps are completely opaque to humans.

Daniel Kokotajlo, a former OpenAI governance researcher who now runs the AI Futures Project, a nonprofit research organization dedicated to forecasting the possible impacts of advanced AI, responded to Pachoki that “even if OpenAI doesn’t go further [in using architectures that make a model’s chain of thought less legible], other might.”

Kokotajlo urged Pachoki to lead efforts to create an industrywide standard on chain-of-thought monitorability “either to arrest the slide into oblivion or better yet to race to the top. I think this is something where we need more than just political will, we need thoughtful technical specifications.” Such calls for shared technical standards echo broader efforts in the AI industry, where labs, standards bodies, and third-party evaluators have been developing benchmarks for model safety and transparency—work that depends on auditors having some access to how models reason.

How Transformers normally work

In a normal Transformer, tokens—the units of information that large language models process—are passed from lower layers of a model’s neural network to higher layers sequentially, with each layer applying mathematical operations in turn. In a language model, the final output of this process is text that people can read. In a reasoning model, the model is told to output a series of steps for arriving at an answer, and this text is first written to a scratchpad, which forms the model’s “chain of thought.” This intermediary text is then fed back through all the layers again to create the next steps, and so on, until the model reaches its final answer and outputs the answer text.

A cheaper but more inscrutable process

In a looped Transformer, tokens are fed multiple times through a single block (which can consist of several different layers, but not the entire network), with the same mathematical operations applied each time. Critically, the output of the block is fed back into the block without being written to a scratchpad each time. The method therefore does not create a natural language “chain of thought,” even though each pass through the block does represent the model’s reasoning. AI researchers sometimes refer to the block’s output as “neuralese,” since it can be processed and reasoned over by the AI model but is not intelligible to a human. The only natural language output of the model is the final answer.

The advantage of looped Transformers is that by using the same mathematical operations each time and not having to pass every token through every layer of the network, they can wring greater performance out of a model of a given size while using less computing power. Studies have shown looped Transformers can achieve the same performance as standard ones while using 50% to 90% less computing power. That matters for cost—a key consideration for many enterprise AI customers, who have been complaining about skyrocketing AI bills. It also matters for providers, since the computing power needed to train and run frontier models has become one of the largest expenses in the AI industry, driving massive investments in data centers and custom chips.

Obscuring part of a model’s chain-of-thought reasoning can also make the model harder to distill—a method in which a smaller model is trained on the outputs of a larger model. The U.S. government and American AI companies have accused Chinese AI companies of engaging in campaigns to distill frontier AI models, including copying the chain-of-thought reasoning steps.

Looped transformers are just one of several new architectures researchers are considering to make AI models more efficient but which can destroy all or part of a model’s chain of thought. Another technique is Prefix Sliding, which was covered in Tuesday’s Fortune “Eye on AI” newsletter. That method involves preserving the initial prompt and the last several thousand reasoning tokens, but discarding many of the intermediate reasoning tokens. What happens with Astra—and whether other labs follow OpenAI’s lead—will help determine how much of future AI systems’ reasoning remains visible to the humans deploying them.

Fortune senior reporter Beatrice Nolan contributed to this story. This story was originally featured on Fortune.com.