Updating the Frontier Safety Framework

Title: Responsible and Safety Introducing the Frontier Safety Framework Introduction: We have created this Frontier Safety Framework to ensure that new AI models, which are likely to have significant impacts on society in high-risk domains such as transportation, healthcare, and energy, play a role in causing severe harm. We recognize that there is significant uncertainty surrounding the capabilities of these models, so our approach involves analyzing and mitigating future risk posed by advanced AI models. This framework outlines the criteria for identifying critical capabilities, the minimal levels of capabilities necessary to cause severe harm, and the strategies to mitigate such harms. Body: 1. Critical Capability Levels (CCLs): We define a CCL as “a minimum level of capabilities that a model must possess before it can play a role in causing severe harm.” These CCLs are determined based on a risk analysis undertaken by the AI Principles project, which is supported by major technology companies. 2. Mitigation Strategies: We develop mitigation strategies to reduce the severity of harms caused by these models if they do cause harm. These strategies include regulatory oversight, privacy safeguards, and transparency measures. 3. Research Program: We initiate a research program to identify critical capabilities of AI models in high-risk domains. This program involves conducting technical assessments, expert interviews, and user studies to identify potential harms of these models. 4. Evaluation Criteria: We define the evaluation criteria for identifying critical capabilities of AI models based on various risk factors such as societal implications, technological feasibility, and public perception. These criteria are then used in our technical assessments to identify potential harms of these models. 5. Mitigation Measures: Based on the identified critical capabilities, we develop mitigation measures for these models. This includes regulatory oversight, privacy safeguards, and transparency measures. These measures aim to reduce the severity of harms caused by AI models if they do cause harm. Conclusion: Our approach provides a framework for ensuring that new AI models play a role in causing severe harm in high-risk domains. We recognize that there is significant uncertainty surrounding the capabilities of these models, so our approach involves analyzing and mitigating such harms using various strategies. Through this framework, we hope to promote responsible and safe AI development.