AI Safety Experts Express Concern Over OpenAI’s New Reasoning Technique

OpenAI has unveiled its latest language model, named Astra, which employs an innovative reasoning technique known as “recurrent depth.” This method enables the model to step outside the traditional sequential reasoning frameworks that most models utilize. According to a report from The Information, the introduction of this technique has sparked serious concern among AI safety experts due to its potential implications for monitoring AI behavior.

While the application of recurrent depth in Astra is reportedly constrained, experts remain wary. Buck Shlegeris, CEO of Redwood, expressed his apprehension in a recent post stating, “I am extremely concerned by the reporting that Astra uses opaque recurrence.” He further noted that should OpenAI deepen its use of this technique, it could significantly reduce the model’s chain-of-thought monitorability.

Longtime AI safety advocate Zvi Mowshowitz echoed these concerns, emphasizing the need for potential regulatory measures to curb a “race to the bottom” among AI laboratories. He cautioned that this technique could undermine the efforts to maintain the integrity of chain-of-thought reasoning, which helps ensure that AI systems operate safely and aligned with user intentions.

Typically, a reasoning model’s transparent chain of thought provides a series of sequential steps that the model follows to solve a problem, making it easier to monitor potential misalignments or misbehaviors. In light of recent incidents involving rogue AI agents, having a clear chain-of-thought record has proven invaluable for understanding their actions.

The “opaque recurrence” approach allows the model to revisit the same query multiple times in a loop, resulting in a less linear thought process and producing fewer identifiable traces—essentially circumventing traditional tracking mechanisms.

Crucially, OpenAI has clarified that the current application of this technique in Astra is limited, and the model is still expected to provide a legible thought process. The company has countered any speculation suggesting a shift toward more obscure reasoning methodologies. In fact, OpenAI plans to implement robust chain-of-thought monitoring systems as part of its ongoing safety initiatives.

Jakub Pachocki, OpenAI’s chief scientist, reaffirmed the organization’s commitment to maintaining clear reasoning paths, stating, “OpenAI has worked to preserve and utilize chain-of-thought monitoring since our very first reasoning models.” He emphasized that this commitment continues to be a core objective of their research efforts.

Despite the existing caveats, there’s widespread concern that employing opaque reasoning techniques could complicate the monitoring of AI systems, especially if they become common across various models. A follow-up report indicated that both Anthropic and Google DeepMind are already exploring this new technique.

In response to these developments, Ryan Greenblatt, chief scientist at Redwood Research, voiced his concern that opaque reasoning could scale rapidly, potentially obscuring all reasoning from detection. He warned of the possibility that models may eventually engage in reasoning primarily in latent spaces, urging OpenAI to halt its advancement in this direction.