【英文】OpenAI+风险预防框架（28页）

Our rationale for grouping and naming these specific risk categories is informed by three  considerations. First, fine-tuning or other domain-specific enhancements (e.g., tailored  prompts or language model programs) may better elicit model capabilities along a particular  risk category. Our evaluations will thus include tests against these enhanced models to  ensure we are testing against the “worst case” scenario we know of. Our procedural  commitments are triggered when any of the tracked risk categories increase in severity,  rather than only when they all increase together. Because capability improvements across  different domains do not necessarily occur at the same rate, this approach ensures we err on  the side of safety. Second, this approach enables us to leverage domain-specific talent to  develop tailored suites of evaluations and monitoring solutions for each risk category. Third,  this approach increases options for tailored, domain-specific mitigations, to help minimize  the need for broader, more disruptive actions.