OpenAI plans to let third-party groups conduct technical safety evaluations of its AI models during the training, evaluation, and deployment phases, extending outside scrutiny to earlier stages of the development cycle1.
The move, reported by Rachel Metz at Bloomberg, means independent assessors will vet models before they reach public release, rather than only after deployment3. OpenAI characterized the change as introducing third-party independent safety assessments in the early stages of AI model development2.
Previously, external red-teaming and safety evaluation at OpenAI occurred primarily around or after a model's launch window. Under the new plan, outside groups will gain access during training itself, giving them visibility into a model's risk profile while it is still being shaped. The policy covers three distinct phases: training, evaluation, and deployment.
The announcement lands as OpenAI faces legal and reputational pressure on safety. British Columbia filed a lawsuit against OpenAI and CEO Sam Altman in San Francisco federal court on Monday, alleging the company's failure to flag a shooter's ChatGPT activity enabled a mass shooting that killed eight people in the province[1]. Separately, OpenAI is in the middle of an aggressive pricing battle with Anthropic after both labs cut frontier-model prices on September 22[2].
ANALYSIS Granting evaluators access at the training stage, rather than only post-deployment, shifts the point of external accountability upstream, where design choices can still be altered before a model ships.