OpenAI Grants Board Veto Power Over High-Risk AI Models
OpenAI introduces a new Preparedness Framework that gives its board veto power over risky AI models, establishing a dedicated advisory group to evaluate catastrophic threats.
OpenAI expands its internal safety processes by introducing a new safety advisory group that sits above technical teams and makes recommendations to leadership. The company also grants its board veto power over risky AI models, a significant move that warrants attention following the recent leadership shake-up that removed two prominent "decelerationist" members.
The updated Preparedness Framework provides a clear path for identifying, analyzing, and addressing "catastrophic" risks, which OpenAI defines as threats that could cause hundreds of billions of dollars in economic damage or lead to severe harm and death. The framework divides oversight into three distinct teams: a "safety systems" team for in-production models, a "preparedness" team for frontier models in development, and a "superalignment" team for theoretical superintelligent models.
For real-world models, OpenAI evaluates risks across four categories: cybersecurity, persuasion like disinformation, model autonomy, and CBRN threats such as the creation of novel pathogens. Teams rate each model in these areas and apply various mitigations, assuming reasonable restrictions on dangerous content like instructions for creating weapons or harmful substances.