Open AI plans to allow external agency safety verification from the early stages of AI model development

(Seoul = Yonhap News) Reporter Joo Jong-guk = Open AI plans to allow external organizations to verify the safety risks of the model from the early stages of artificial intelligence (AI) model development.

Rama Ahmad, who is in charge of safety review cooperation with external experts at OpenAI, said in an interview with Bloomberg News on the 22nd (local time) that the company plans to allow third-party organizations to perform technical safety assessments during the training, evaluation, and deployment stages of AI models in the future.

Until now, external organizations had been involved for safety evaluation or performance verification at the stage just before the launch of a new model.

He said, “As the importance and impact of AI technology grows, we want to closely review safety not only at the actual service deployment stage, but also at the equally important learning and evaluation stage.”

OpenAI also presented “strong independence guarantees, scientific rigor, thorough security practices, and clear accountability” as key priorities for this external review to function effectively.

OpenAI plans to soon announce a policy to receive verification from external experts at every cycle.

Ahmad added that for the most sensitive work, external evaluators can be brought directly to the OpenAI office to conduct the work, a method that has been attempted in the past.

OpenAI said in a post that it is currently in discussions with organizations that can perform such evaluation work. External organizations also include AI research group METR and Redwood Research, which investigated the Hugging Face hacking incident of the Open AI model.

OpenAI said, “AI research labs have a responsibility to protect sensitive information while ensuring that meaningful verification can take place.”

Dario Amodei, CEO of competitor Antropic, recently raised the need to control the speed of AI development and announced that he would implement new safety measures, such as introducing a third-party evaluation agency, and OpenAI CEO Sam Altman also said he agreed to this.

On the 18th, Antropic announced that it had designated Accenture as a resident independent safety evaluation agency and that this evaluation team would have a similar level of system access as Antropic executives and employees and would be used to verify the safety of the company's cutting-edge AI model.