OpenAI’s new reasoning technique alarms AI safety experts

OpenAI's upcoming Astra model will utilize a reasoning technique called 'recurrent depth' or 'opaque recurrence,' operating outside sequential thinking. This approach has sparked serious concerns among AI safety experts regarding the transparency and monitorability of the model's decision-making process. Industry leaders and safety advocates warned that intensive use of opaque recurrence could destroy the ability to audit a model's chain of thought, which serves as a critical diagnostic tool for detecting misalignment or rogue behaviors. Unlike traditional reasoning models that output clear sequential steps to solve problems, opaque recurrence processes queries in iterative loops that leave fewer legible traces, effectively bypassing conventional accountability records.