OpenAI’s reported ‘opaque recurrence’ technique raises fresh AI safety concerns

A reasoning technique reportedly being used in OpenAI’s forthcoming Astra model has prompted concern among AI safety researchers who fear increasingly complex systems could become harder to monitor.

TechCrunch, citing reporting by The Information, said Astra will make limited use of a technique known as “recurrent depth”, also described as “opaque recurrence”. Instead of following a largely sequential reasoning process, the technique allows a model to process a problem repeatedly through loops.

Tap here to add The Daily Star as a trusted source

Researchers are concerned that greater reliance on such processing could reduce the usefulness of chain-of-thought monitoring, which is used to examine intermediate reasoning signals for potentially problematic model behaviour.

AI safety researchers, including Redwood Research figures Buck Shlegeris and Ryan Greenblatt, have warned about the implications if opaque reasoning becomes substantially more prominent in future systems. Safety advocate Zvi Mowshowitz has also argued that competition among AI companies could create pressure to adopt techniques that make models less transparent.

OpenAI has pushed back against suggestions that Astra represents a wholesale move towards unreadable reasoning. The reported use of recurrence is limited, and the model’s chain of thought is still expected to remain accessible for monitoring.

OpenAI chief scientist Jakub Pachocki has also said preserving and using chain-of-thought monitoring remains a central objective of the company’s research programme.