OpenAI's upcoming AI model, Astra, is drawing scrutiny from artificial intelligence safety researchers over a reasoning technique that could make parts of the model's internal decision-making more difficult to monitor.
The technique, known as "recurrent depth" or "opaque recurrence", allows an AI model to repeatedly process information internally rather than relying entirely on the sequential reasoning commonly associated with current reasoning models. Astra is reportedly using the approach on a limited basis.
The development has raised questions among researchers about chain-of-thought monitoring, a method used to examine the intermediate reasoning generated by AI models. Although these records do not provide a complete representation of how a model reaches a conclusion, researchers have used them as one mechanism for identifying potentially undesirable or misaligned behaviour.
With recurrent depth, a model can process the same problem multiple times through an internal loop. This can reduce the amount of reasoning that appears in a conventional, human-readable chain of thought, potentially making some aspects of the model's behaviour harder to examine.
The concerns do not mean Astra's reasoning will become entirely opaque. Reporting indicates that its use of recurrent depth remains limited and that the model is still expected to produce legible reasoning traces.
OpenAI Chief Scientist Jakub Pachocki has also reiterated the company's commitment to chain-of-thought monitoring. He said preserving and using such monitoring has been part of OpenAI's work since its early reasoning models and remains a core objective of its research programme.
The debate has nevertheless attracted attention from AI safety researchers. Buck Shlegeris, CEO of Redwood Research, expressed concern that greater use of opaque recurrence in future models could reduce the effectiveness of chain-of-thought monitoring. Redwood Research Chief Scientist Ryan Greenblatt similarly raised concerns about a scenario in which increasingly large portions of AI reasoning move into less visible internal representations.
The issue extends beyond OpenAI. Google DeepMind and Anthropic are also reportedly discussing the technique, suggesting that recurrent approaches could become a broader area of research as AI developers look for new ways to improve model performance and efficiency.
The debate comes as AI companies develop increasingly capable reasoning models and autonomous agents. As these systems are given greater ability to perform multi-step tasks, researchers and developers are also examining how their decisions can be monitored and evaluated.
For OpenAI, Astra's reported architecture brings that question into sharper focus. The immediate issue is not whether chain-of-thought monitoring is disappearing, but whether increasingly sophisticated reasoning techniques could make existing monitoring methods less informative as AI systems become more capable.
Disclaimer: This article may include information derived from interviews, press releases, public statements, research, company communications and other publicly available or third-party sources. Such material may be summarised, paraphrased or contextualised for journalistic and editorial purposes. All rights in third-party content remain with their respective owners.