OpenAI's Astra model reportedly uses hidden reasoning technique
OpenAI's unreleased Astra model reportedly uses a hidden reasoning technique called 'recurrent depth' that makes it harder for safety auditors to monitor the model's thinking process through standard chain-of-thought logs. AI safety researchers have warned that broader adoption of such techniques could undermine transparency and trigger a race to the bottom in safety practices across the industry.
What happened
OpenAI's unreleased Astra model reportedly uses a reasoning technique called 'recurrent depth' or 'opaque recurrence' that processes queries in hidden loops within the model's internal state before producing visible output.
Chain-of-thought monitoring, which auditors rely on to detect misbehavior, depends on visible reasoning steps, but recurrent depth erodes that visibility by leaving fewer legible traces of the model's reasoning process.
Reported impact
Redwood Research's CEO warned that if OpenAI pushes this technique further, safety auditors could lose the ability to monitor chain-of-thought output effectively.
AI safety advocate Zvi Mowshowitz expressed concern that widespread adoption could trigger a 'race to the bottom' among AI labs, damaging the monitorability of reasoning that companies have worked to maintain.
Index analysis
Reports indicate that Anthropic and Google DeepMind reportedly explore similar hidden reasoning techniques, raising concerns about an industry-wide transparency race.
OpenAI stated that Astra's use of the technique is limited and that its chain of thought is still expected to be legible, and the company emphasized its commitment to preserving chain-of-thought monitoring systems.
Governance controls
OpenAI chief scientist Jakub Pachocki posted that the company has worked to preserve and utilize chain-of-thought, though full details of monitoring systems were not disclosed in the reporting.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.