When an AI shows its work, auditors can catch it doing something dangerous. When it reasons in secret, they’re reading a blank page. That tension sits at the center of what’s rattling AI safety researchers right now — and it centers on a technique reportedly buried inside OpenAI‘s unreleased Astra model. Chain-of-thought monitoring, the primary tool labs use to catch misbehavior, depends on visible reasoning. Recurrent depth, as described in recent reporting, quietly erodes that visibility.
According to reporting by The Information, summarized by TechCrunch, Astra uses something called “recurrent depth” — also described as “opaque recurrence.” The details remain anonymously sourced and unconfirmed by OpenAI officially. But the architecture is real, and the implications are serious enough that prominent safety researchers are already going public with their concerns.
The Technique, Explained
Here is what recurrent depth actually does — and why it differs from the reasoning transparency you’re used to.
- Standard AI reasoning models output chain-of-thought (CoT) — visible, sequential thinking steps that labs and auditors use to diagnose misbehavior, flag deception, and analyze rogue behavior. CoT logs were reportedly crucial in diagnosing a recent OpenAI rogue agent incident.
- Recurrent depth loops the same transformer layers over a hidden internal state multiple times before producing any output token — no intermediate text, no readable trail left in the token stream.
- All extra reasoning happens in activations (latent space), making it inaccessible to ordinary text-based safety monitors without specialized probing tools.
- Astra’s reported use is described as limited and constrained — computational depth reportedly within roughly 2x of GPT-4, according to OpenAI chief scientist Jakub Pachocki.
- Key uncertainty: no official OpenAI architecture documentation or system card has been published; all technical claims about Astra rest on The Information’s anonymously sourced reporting.
Safety researchers are not alarmed by what Astra does today. They’re alarmed by where this trajectory leads.
Buck Shlegeris, CEO of Redwood Research, put it plainly: “I am extremely concerned… if OpenAI pushes this technique further, they’ll have the option to massively increase the recurrence and totally destroy CoT monitorability.”
Ryan Greenblatt, Redwood’s chief scientist, sees a near-inevitable progression: scale opaque reasoning far enough, and the model reasons almost entirely in latent space, with only its final answers visible. It is the difference between a chef cooking in an open kitchen and one working entirely behind closed doors — you receive the dish, but no one can examine what went into it or whether the process was safe.
OpenAI’s Defense – and the Race Nobody Wants to Win
OpenAI insists CoT oversight remains a core commitment, but competitive pressure across the industry is already pushing in the opposite direction.
Pachocki says OpenAI has “worked to preserve and utilize chain-of-thought monitoring since our very first reasoning models,” calling CoT oversight a core research goal. OpenAI has also announced plans for extensive CoT monitoring systems and collaboration with governments and independent safety organizations before Astra ships. That is the responsible answer.
The uncomfortable part: Anthropic and Google DeepMind are reportedly already discussing the same technique, according to The Information. Safety advocate Zvi Mowshowitz called this “playing with fire,” warned it could trigger a competitive race to the bottom on transparency, and suggested laws may be the only guardrail that prevents it. Once one major lab normalizes opaque recurrence, staying transparent starts to feel like showing up to a knife fight with a handshake.
The safety norm around legible reasoning is not being abandoned with a press release. It is being optimized away, one hidden loop at a time. For organizations deploying AI in high-stakes decisions, the erosion of legible reasoning is a material governance risk — not a hypothetical one.
