OpenAI's 'Opaque Recurrence' Reasoning Technique Alarms AI Safety Experts

A new reasoning method could make AI's chain of thought harder to monitor — and safety researchers are worried.

Published: 2026-09-03 Category: Quick Take Sources: TechCrunch

The Story

OpenAI's upcoming Astra model will reportedly use a reasoning technique called "recurrent depth" — also referred to as "opaque recurrence" — that lets it operate outside the sequential thinking that characterizes most reasoning models. The Information first reported the news on Tuesday. The technique processes the same query several times in a loop rather than in a linear chain, leaving fewer legible traces and effectively side-stepping a conventional chain-of-thought record.

That has rattled AI safety experts. Redwood CEO Buck Shlegeris said he is "extremely concerned" by the reporting, warning that if OpenAI pushes the technique further, it could "massively increase the recurrence and totally destroy CoT monitorability." Longtime safety advocate Zvi Mowshowitz went further, suggesting laws might be needed to prevent a "race to the bottom" among AI labs, calling the technique "playing with fire."

Why It Matters

Chain-of-thought records have become a cornerstone of AI safety practice. They provide the sequential steps a model takes while solving a problem, and they proved valuable in teasing out why OpenAI's recent rogue agent activity behaved the way it did. If models increasingly reason in "latent space" — with no legible trace — that monitoring tool degrades or disappears entirely.

Redwood Research chief scientist Ryan Greenblatt warned the natural progression could be scaling opaque reasoning "to the point where the model reasons entirely or almost entirely in latent space." He hopes it isn't "too late to avoid the most concerning architectures."

The Counterpoint

OpenAI pushed back. Chief scientist Jakub Pachocki emphasized the lab's commitment to legible chains of thought, calling CoT monitoring "a core goal of our current research program." Astra's use of the technique appears limited, and the company denies any shift to "neuralese." All AI models do some opaque reasoning, and few researchers treat CoT logs as a direct representation of reasoning.

Still, the concern is that the technique scales. The Information reported Wednesday that both Anthropic and Google DeepMind are already discussing it — meaning this isn't just an OpenAI question, but an industry-wide trajectory.

Source: TechCrunch, "OpenAI's new reasoning technique alarms AI safety experts" (Sept 2, 2026).