OpenAI is shifting its safety strategy by inviting third-party organizations to conduct technical assessments earlier in the AI model development cycle. Previously, the company primarily engaged external experts only during the final stages before a model's public release.
The company aims to integrate independent oversight into the training and evaluation phases. According to Lama Ahmad, who leads these external collaborations, the goal is to ensure critical development stages are scrutinized as potential risks associated with AI models continue to grow.
OpenAI has outlined key priorities for these evaluations, emphasizing scientific rigor, robust safety practices, and clear accountability. For highly sensitive projects, the company may grant external evaluators physical access to its offices to perform on-site testing.
The developer is currently in discussions with various organizations to facilitate these assessments. Potential partners include research groups like METR and Redwood Research, both of which have previously collaborated with OpenAI on security investigations.
This move aligns with broader industry trends toward increased transparency. Competitor Anthropic has also committed to third-party safety testing, recently announcing a partnership with Accenture to evaluate its own frontier AI models.
