OpenAI said on Tuesday it will allow external organizations to conduct technical
safety assessments earlier in the development cycle of new AI models, including
during training, evaluation and rollout. The company said reviews will be
governed by strong independence mechanisms, scientific rigor, robust safety
measures and clear accountability; Lama Ahmad, who oversees OpenAI’s external
safety reviews, said the firm previously typically invited such groups before
model release but will now extend scrutiny to higher-risk stages such as
training and evaluation as risks rise.
