OpenAI has introduced a new reasoning approach called “recurrent depth,” or “opaque recurrence,” in its latest Astra model. Unlike traditional reasoning models that follow a clear sequential chain of thought, this technique enables the model to repeatedly process a query in a loop, leading to less transparent and non-linear reasoning steps. While this method could improve the model's capabilities, it also poses challenges for monitoring and understanding the model’s internal decision-making processes.

The introduction of this reasoning technique has sparked concern among AI safety experts. Prominent figures like Buck Shlegeris, CEO of Redwood Research, and advocate Zvi Mowshowitz have expressed fears that expanding the use of opaque recurrence could significantly diminish the monitorability of AI reasoning. They warn that this might accelerate a risky “race to the bottom” among AI developers, potentially making it harder to detect misalignment or misconduct in advanced AI systems.

Despite these worries, OpenAI has clarified that Astra’s implementation of recurrent depth is currently limited, with efforts to maintain legible chains of thought. According to OpenAI’s chief scientist Jakub Pachocki, chain-of-thought monitoring remains a core priority, and the model does not employ so-called “neuralese,” or fully latent reasoning inaccessible to scrutiny. OpenAI also plans to enhance chain-of-thought monitoring as part of its safety strategy moving forward.

Other major AI labs, including Anthropic and Google DeepMind, are reportedly exploring similar techniques, raising further debate about the future of AI transparency. AI safety experts like Redwood Research’s Ryan Greenblatt caution that a natural progression toward greater use of opaque reasoning could eventually render AI reasoning invisible, making governance and safety interventions far more difficult. This development signals growing urgency in balancing AI innovation with robust safety practices.