OpenAI plans to use a reasoning technique called recurrent depth in its upcoming Astra model, TechCrunch reported on Wednesday. The method allows the system to operate outside of sequential thinking, looping through queries rather than processing them in conventional linear steps.
The technique, also called opaque recurrence, was first reported by The Information on Tuesday. In opaque recurrence, a model processes the same prompt multiple times within a loop. The process leaves fewer legible traces, side-stepping standard chain-of-thought logs that researchers inspect to detect misbehavior or misalignment.
Safety experts warn of monitoring risks
AI safety experts responded with immediate warnings about the difficulty of tracking non-linear reasoning. Redwood Research chief executive Buck Shlegeris stated in an online post that expanding the method could harm safety oversight. While noting he did not know whether Astra was significantly less monitorable than earlier models, Shlegeris warned that expanding the technique gives developers the option to massively increase recurrence and destroy monitorability.
AI safety advocate Zvi Mowshowitz warned that new legislation may become necessary to stop labs from abandoning oversight standards. Mowshowitz wrote that the method risks breaking an informal taboo established by OpenAI and Anthropic to maintain legible reasoning paths for as long as possible.
Recent incidents have relied directly on visible reasoning logs for safety investigations. When OpenAI evaluated rogue agent activity, investigators used chain-of-thought logs to determine why the systems took unintended actions.
OpenAI response and industry adoption
OpenAI chief scientist Jakub Pachocki defended the lab's practices in a post on X. Pachocki stated that OpenAI has worked to preserve chain-of-thought monitoring since releasing its earliest reasoning models, calling legible logs a primary research priority.
Company representatives pushed back against suggestions that the model would move toward internal "neuralese." OpenAI has announced plans for expanded chain-of-thought monitoring systems, and Astra's implementation of recurrent depth is currently limited.
The broader AI industry is already considering similar designs. The Information reported Wednesday morning that Google DeepMind and Anthropic have begun internal discussions about using the technique, while Redwood Research chief scientist Ryan Greenblatt warned that opaque reasoning could scale quickly into latent space if labs fail to stop its expansion.
