OpenAI will allow independent third-party groups to conduct technical safety assessments of its artificial intelligence models while they are still being trained and evaluated, expanding external oversight beyond checks conducted before launch.

Bloomberg reported that the company is in discussions with external organisations, including METR and Redwood Research, about carrying out the assessments.

The move would allow outside evaluators to examine models during the development process rather than waiting until they are ready for release.

However, OpenAI has not yet named a formal partner or disclosed the terms under which external groups would gain access to its models.

The announcement comes about 10 days after OpenAI chief executive Sam Altman said the company would give independent evaluators access to its facilities and tools.

On September 12, Altman said evaluators would receive desks, badges and laptops and would have the right to publish their findings.

The latest plan appears to broaden that commitment by allowing external groups to assess models at multiple stages of training and evaluation, potentially providing earlier insight into safety risks before products are launched.