OpenAI is building a new model called Astra. Reports say it will use a reasoning method known as recurrent depth.
The method is also called opaque recurrence. It changes how the model works through a problem.
Most reasoning models work in a straight line. They write out each step in order until they reach an answer.
Astra's method works differently. It runs the same query through a loop several times before landing on a response.
This looping process leaves fewer visible clues about how the model reached its answer.
Why Researchers Are Worried
AI safety researchers rely on a model's written steps, known as its chain of thought, to check its work.
Redwood Research CEO Buck Shlegeris said he is worried about the new method. He said scaling it up could remove much of the ability to monitor a model's thinking.
Writer Zvi Mowshowitz said rules might be needed to stop AI companies from racing toward less visible reasoning.
He said OpenAI and Anthropic have worked to keep their models' thinking visible. He said the new method could put that effort at risk.
Redwood chief scientist Ryan Greenblatt raised similar concerns. He said opaque reasoning could scale faster than normal step by step reasoning.
Greenblatt said his biggest worry is that future models could end up reasoning almost entirely outside of any visible text.
OpenAI's Response
OpenAI says Astra's use of the new method is limited for now. The company says the model's chain of thought is still readable.
OpenAI chief scientist Jakub Pachocki addressed the concerns in a public post. He said keeping reasoning visible has been a goal since the company's first reasoning models.
He said this remains a core part of OpenAI's current research work.
OpenAI has also announced plans to build systems that watch a model's chain of thought for safety issues.
Researchers point out that all AI models already do some hidden reasoning. Few experts treat a model's written steps as a full picture of its thinking.
Still, the concern is that this new method could make hidden reasoning more common across the AI industry.
A follow up report from The Information said Anthropic and Google DeepMind are already discussing similar reasoning techniques.
