OpenAI’s new reasoning technique alarms AI safety experts

OpenAI’s new Astra model will reportedly use recurrent depth, a technique that lets AI reason outside the step-by-step sequence common in current reasoning models. The approach has alarmed some AI safety experts because it could make model behavior harder to interpret or control.

The technique points to a broader push toward more flexible and powerful AI reasoning systems. But as models move beyond transparent sequential reasoning, AI safety researchers may face greater challenges in auditing how decisions are made.