OpenAI has published its initial beta version of its Preparedness Framework, outlining safety precautions for its AI models. The company commits to running consistent evaluations on its frontier models and showing the findings in risk “scorecards” that will be continuously updated. The risk thresholds will be classified into four safety levels: low, medium, high, and critical. OpenAI is also restructuring its decision-making process, with a dedicated Preparedness team and a cross-functional Safety Advisory Group. Leadership will remain the decision-maker, but the Board of Directors will have the right to reverse decisions. Other framework elements include developing a protocol for added safety, collaborating with external and internal teams to track real-world misuse, and pioneering new research in measuring how risk evolves as models scale.
OpenAI’s Groundbreaking Safety Plan for Frontier AI Models and Its Implications for Future Development
