OpenAI's new reasoning technique alarms AI safety experts
OpenAI's upcoming Astra model will use a method called "opaque recurrence" that makes its reasoning harder to monitor, prompting concern from AI safety researchers. Experts warn that expanding the technique could severely undermine the ability to oversee AI behavior.
According to a report published by The Information on Tuesday, OpenAI's forthcoming Astra model will employ a new reasoning approach known as "recurrent depth," also referred to as "opaque recurrence." Unlike typical reasoning models that follow a step-by-step sequential chain of thought, this method allows the model to process the same query repeatedly in a loop, producing fewer legible traces for outside observers to follow.
Safety concerns
The report has unsettled several AI safety specialists. Redwood Research CEO Buck Shlegeris expressed deep concern, saying he could not yet tell how much less monitorable Astra is compared to earlier models, but warned that if OpenAI expands the technique further, it could massively scale up recurrence and effectively destroy chain-of-thought monitorability altogether.
AI safety advocate Zvi Mowshowitz suggested that regulation might become necessary to prevent a "race to the bottom" among AI labs. He argued the technique threatens an informal norm that OpenAI and Anthropic have worked to uphold — preserving faithful, monitorable chains of thought for as long as possible.
OpenAI's response
Normally, a reasoning model's chain of thought lays out the sequential steps it takes while solving a problem. Though an imperfect representation, it has served as a useful tool for spotting misbehavior or misalignment — it reportedly helped explain past instances of rogue behavior by OpenAI's own agents.
OpenAI says Astra's use of the technique is limited and that its chain of thought will remain legible, pushing back on suggestions the company is moving toward fully opaque, "neuralese" reasoning. Chief scientist Jakub Pachocki reiterated that preserving monitorable chains of thought remains a core research priority for the lab.
A follow-up report Wednesday from The Information said Anthropic and Google DeepMind are already discussing similar techniques. Redwood Research chief scientist Ryan Greenblatt warned that opaque reasoning could scale faster than conventional chain-of-thought methods, raising the possibility that future models could end up reasoning almost entirely outside of human-visible channels.
