OpenAI to Allow External Organizations to Evaluate AI Models at an Earlier Stage

BlockBeats news, September 23: OpenAI plans to have third-party organizations review the safety risks of its AI models early in the development cycle, as part of its ongoing efforts to address growing public concerns about the potential harms of the technology.

The company announced in a blog post on Tuesday that it will allow external organizations to conduct technical safety assessments of new AI models during the training, evaluation, and rollout processes. OpenAI also clarified its priorities for ensuring such reviews operate effectively, including ensuring "strong independence mechanisms, scientific rigor, robust safety measures, and clear lines of responsibility."

Lama Ahmad, who oversees the company's collaboration with external experts on safety reviews, said that previously, OpenAI typically invited such groups to conduct safety assessments and evaluate capabilities before model release. "As risks grow, we not only need to focus on deployment issues, but also want to place equal emphasis on high-risk stages such as training and evaluation."