What was reported, and what OpenAI confirmed
The Information first reported that OpenAI uses a technique called recurrent depth — sometimes described as looped transformers — in GPT-6 Astra. Fortune reported on 3 September that the technique alarmed a set of AI safety researchers, and that OpenAI’s chief scientist Jakub Pachocki acknowledged its use while promising more architectural detail later.
The mechanism is simple to state. A conventional reasoning model produces its intermediate steps as text: it writes tokens, and those tokens are readable. Recurrent depth instead runs the same transformer layers repeatedly over a hidden internal state before any token is emitted. The extra computation happens in activations rather than in words, so it leaves no trail in the output stream.
The pitch for it is efficiency: more effective depth without a proportionally larger model.
Why safety researchers object
Chain-of-thought monitoring is the main practical tool labs and outside evaluators use to catch a model misbehaving. It works for an unglamorous reason: if a model has to write its reasoning down to use it, an auditor can read what it wrote. That property is a side effect of the architecture, not a designed safeguard, and it disappears when the reasoning stops being written.
Steven Adler, a former OpenAI safety researcher, told Fortune the approach potentially crosses one of the few redlines the industry has. Peter Wildeford of the AI Policy Network called it potentially very concerning and potentially reckless. Daniel Kokotajlo of the AI Futures Project made the argument that matters most for the field rather than for one model: that others may follow, and opacity spreads by precedent.
OpenAI’s answer
Pachocki said the company cares deeply about chain-of-thought monitoring and has worked to preserve and use it since its first reasoning models. He and other OpenAI researchers pushed back on the coverage, saying it created alarm out of proportion to what the company has done — that OpenAI has limited how far the looped architecture is used precisely so the model’s reasoning stays legible.
Both things can be true. A bounded use of recurrent depth may leave enough written reasoning to monitor. What the disagreement is really about is whether that boundary is a commitment or a current engineering choice, and there is no published specification to check it against.
What would settle it
The architectural details Pachocki promised. Specifically: how much of Astra’s effective reasoning depth occurs in latent space, whether the proportion is fixed or scales with the task, and whether external evaluators — the organisations that ran OpenAI’s own capability evaluations this year — get access sufficient to test monitorability rather than only performance.
Until that exists, the question is not whether Astra is monitorable. It is that nobody outside OpenAI can currently check.
