The “recurrent depth” technique reportedly set to be used in OpenAI’s new Astra model has raised concerns among AI safety experts. According to a report by The Information, the method, also known as “opaque recurrence,” allows the model to process the same query multiple times within a loop, moving beyond the sequential thinking pattern found in most reasoning models. This approach could make it more difficult to monitor the model’s chain of thought.
Redwood CEO Buck Shlegeris said he was “extremely concerned” about Astra’s use of the technique. Zvi Mowshowitz argued that legislation might be needed to prevent a “race to the bottom” among AI companies. According to experts, while traditional chain-of-thought records do not show every step a model follows when solving a problem with complete perfection, they are an important tool for examining erroneous or misaligned behavior. These records were also used to investigate OpenAI’s recent rogue agent activities.
It was reported that the technique’s use in Astra is limited for now, and that the model’s chain of thought is still expected to remain readable. OpenAI rejected claims that the model would switch to “neuralese” and reiterated its commitment to the traceability of chain-of-thought processes. Chief Scientist Jakub Pachocki said that preserving this monitoring capability since the first reasoning models had been one of the core goals of the research program.
Nevertheless, it is stated that traceability could decline further if the technique becomes widespread. The Information reported that Anthropic and Google DeepMind had also discussed the technology. Ryan Greenblatt, chief scientist at Redwood Research, said he was concerned that opaque reasoning could expand, moving the model’s thoughts into a hidden space either completely or to a large extent.
Why it matters
The widespread adoption of this approach could make it more difficult for safety processes to monitor why AI systems arrive at a particular result. Although chains of thought are not a flawless audit trail, they are one of the primary tools used to examine erroneous or misaligned behavior, so the partial obfuscation of these records could affect researchers’ ability to diagnose problems. The expectation that traceability will be preserved for Astra for now shows that the debate is focused less on the current model than on the extent to which the method will be used in the future. The fact that Anthropic and Google DeepMind are also interested in the same technique indicates that the issue has moved beyond a single company’s product choice and become linked to AI safety standards and the possible need for regulation.
Background
OpenAI is not a new name in the FikirPilot archive: we have published 21 articles mentioning the name in the last 90 days; the latest is dated September 6, 2026.