OpenAI's opaque reasoning technique sparks AI safety alarm over monitoring

OpenAI has integrated a reasoning approach known as "opaque recurrence" into its upcoming Astra model, marking a significant shift in how advanced AI systems process information. Unlike traditional reasoning models that follow a transparent, step-by-step chain of thought, this technique processes queries iteratively through loops, leaving fewer interpretable traces of the model's decision-making process—and that's worrying safety-focused researchers.

The technique and its risks#

In conventional AI reasoning, models show their work through a chain of thought: a sequential record of how they arrived at answers. This transparency serves as a critical monitoring tool for detecting misalignment or unexpected behavior. "Opaque recurrence" operates differently. Rather than displaying a linear path of reasoning, the model cycles through the same problem multiple times, with intermediate computations happening in ways that resist human interpretation.

The concern isn't merely theoretical. As The Information first reported, even limited adoption of this method could eventually expand. Redwood Research CEO Buck Shlegeris warned that scaling up opaque recurrence could "totally destroy[] CoT monitorability"—the ability to monitor a model's chain of thought at all.

Safety community response#

Leading AI safety advocates have raised alarms. Longtime researcher Zvi Mowshowitz cautioned that the technique breaches an implicit "taboo" labs like OpenAI and Anthropic have maintained: preserving readable reasoning. He suggested regulation might be necessary to prevent a "race to the bottom" where labs compete by making their models harder to monitor.

Redwood Research's Ryan Greenblatt articulated the steepest concern: that opaque reasoning could naturally scale to the point where "the model reasons entirely or almost entirely in latent space," removing reasoning from any human-visible channel.

OpenAI's position#

OpenAI has pushed back against the alarm, with chief scientist Jakub Pachocki emphasizing the lab's ongoing commitment to legible chains of thought. The company notes that Astra's use of opaque recurrence remains limited and that its chain of thought remains largely interpretable. Plans for "extensive chain-of-thought monitoring systems" form part of the company's stated safety roadmap.

Reports indicate that Anthropic and Google DeepMind are also exploring the technique—suggesting this may become a broader industry trend rather than an isolated experiment.