OpenAI's New Reasoning Technique Raises Concerns Among AI Safety Experts

OpenAI is preparing to launch a new model, known internally by the code name Astra, which could fundamentally change the way artificial intelligence systems process information. According to TechCrunch, the new model will use a method called "recurrent depth" — a radical departure from the approaches companies in the field have relied on until now.

What is "recurrent depth"?

Most current reasoning models, including those developed by OpenAI itself, operate through a sequential process: the model works through logical steps one after another, in a linear order, to arrive at an answer. The "recurrent depth" technique proposes a different approach, in which the model can revisit its own calculations, processing information in repeated loops rather than strictly moving forward step by step.

In practice, instead of "thinking" linearly from one conclusion to the next, the system could return to earlier stages of processing to refine its answers — a process that resembles the workings of a recurrent neural network more than the step-by-step logic characteristic of "chain-of-thought" models.

An architecture based on recurrent processing, one that doesn't follow an easily interpretable, linear logical path, could make it far more difficult to understand how the model arrives at certain conclusions. Many researchers view this lack of visibility into the decision-making process as a significant risk — particularly as AI models become increasingly capable and are deployed in high-stakes contexts.

Context and outlook

OpenAI has not yet released full technical details about Astra or a timeline for its launch. The company has, however, found itself at the center of similar debates before, as it has rolled out increasingly advanced reasoning models that raise questions about interpretability and control.

It remains to be seen whether the performance gains offered by "recurrent depth" will, in the eyes of the research community, outweigh the challenges it poses for transparency and oversight of model behavior.