OpenAI defends 'opaque recurrence' technique as AI safety concerns widen to other labs

OpenAI's Astra model will employ a reasoning technique called 'recurrent depth' that processes queries in a loop rather than in a strictly sequential chain of thought, according to a report from The Information. This approach, also referred to as 'opaque recurrence,' leaves fewer legible traces of the model's reasoning steps, potentially undermining chain-of-thought (CoT) monitoring — a key tool for detecting misalignment or misbehaviour in AI systems.

The technique has drawn sharp criticism from AI safety researchers. Buck Shlegeris, CEO of the Redwood research group, wrote on X that he is 'extremely concerned' by the reporting. 'If OpenAI pushes this technique further, they'll have the option to massively increase the recurrence and totally destroy CoT monitorability,' he said. Longtime safety advocate Zvi Mowshowitz argued that the approach risks breaking an emerging industry norm to maintain chain-of-thought faithfulness, and warned that laws may be needed to prevent a 'race to the bottom' among AI labs.

OpenAI pushed back against suggestions that Astra would shift to inscrutable internal reasoning. Chief scientist Jakub Pachocki stated on X that 'OpenAI has worked to preserve and utilize chain-of-thought monitoring since our very first reasoning models. It's a core goal of our current research program.' The company has previously announced plans for extensive CoT monitoring systems as part of its safety framework.

While all AI models engage in some opaque reasoning, and researchers caution that chain-of-thought logs are never a perfect mirror of internal processes, the emergence of a dedicated technique for increasing opacity has intensified the debate. In a follow-up report, The Information noted that both Anthropic and Google DeepMind are now discussing the technique, suggesting broader industry interest.

During a recent incident involving OpenAI's 'rogue agent' behaviour, chain-of-thought records proved instrumental in understanding why the agent acted as it did, underscoring the practical importance of legible reasoning.

Why This Matters

Chain-of-thought monitoring is one of the few practical tools researchers have to check whether AI systems are pursuing hidden objectives or behaving unsafely.
If opaque recurrence becomes standard, the AI industry could lose visibility into model reasoning just as systems become more capable and autonomous.
The discussion at multiple top labs suggests this is not an isolated experiment but a potential shift in how reasoning models are built.

Background

Chain-of-thought reasoning has been a cornerstone of modern large language models, allowing them to 'show their work' by generating intermediate steps before producing a final answer. For AI safety researchers, these records are crucial for detecting misalignment — for instance, when a model deceives users or pursues goals at odds with its instructions. OpenAI, Anthropic, and DeepMind have all publicly committed to preserving CoT legibility. The 'recurrent depth' technique challenges that commitment by letting a model iterate over a query multiple times internally, producing a result without a clean linear trace.

Key Perspectives

OpenAI: The company maintains that Astra's use of the technique is limited and that chain-of-thought will remain legible. Chief scientist Jakub Pachocki emphasized that preserving CoT monitorability is a core research goal. AI Safety Experts (Buck Shlegeris, Zvi Mowshowitz): They view the technique as a dangerous step toward unmonitorable reasoning. Shlegeris warns that OpenAI could 'massively increase recurrence' in future models. Mowshowitz argues it risks breaking the industry norm for transparency. Industry Observers: The fact that Anthropic and Google DeepMind are already discussing the technique indicates that the pressure to adopt opaque recurrence may be spreading, potentially leading to a competitive dynamic where transparency is sacrificed for performance.

What to Watch

Whether OpenAI publishes safety evaluations that specifically measure chain-of-thought faithfulness in Astra.
Public statements from Anthropic and DeepMind on whether they plan to implement similar techniques.
Regulatory responses: Mowshowitz has called for legislation; any government AI safety body may take interest in the monitorability of frontier models.

ZOTPAPER