人工智能伤害的机制：从人工智能事件中吸取的教训

Executive Summary With recent advancements in artificial intelligence—particularly,powerful generativemodels—private and public sector actors have heralded the benefits of incorporating AImore prominently into our daily lives. Frequently cited benefits include increasedproductivity, efficiency, and personalization. However, the harm caused by AI remainsto be more fully understood. As a result of wider AI deployment and use, the number ofAI harm incidents has surged in recent years, suggesting that current approaches toharm preventionmay befalling short. This report argues that this is due to a limitedunderstanding of how AI risks materialize in practice. Leveraging AI incident reportsfrom the AI Incident Database, it analyzes how AI deployment results in harm andidentifies six key mechanisms that describe this process (Table 1). A review of AI incidents associated with these mechanisms leads to several keytakeaways that should inform AI governance approaches in the future. 1.A one-size-fits-all approach to harm prevention will fall short.This reportillustrates the diverse pathways to AI harm and the wide range of actorsinvolved. Effective mitigation requires an equally diverse response strategy thatincludes sociotechnical approaches.Adopting model-based approaches alonecouldespeciallyneglect integration harms and failures of human oversight. 2.To date, risk of harm correlates only weakly with model capabilities.Thisreport illustrates many instances of harm that implicate single-purpose AIsystems. Yet many policy approaches use broad model capabilities, often proxiedby computing power, as a predictor for the propensity to do harm. This fails tomitigate the significant risk associated with the irresponsible design,development, and deployment ofless powerfulAIsystems. 3.Tracking AI incidents offers invaluable insights into real AI risks and helpsbuild response capacity.Technical innovation, experimentation with new usecases, and novel attack strategies will result in new AI harm incidents in the future. Keeping pace with these developments requires rapid adaptation andagile responses. Comprehensive AI incident reporting allowsforlearningandadaptationat an accelerated pace, enabling improved mitigation strategies andidentification of novel AI risks as they emerge. Incident reporting must berecognized as a critical policy tool to address AI risks. Table of Contents Executive Summary................................................................................................................................1Introduction...............................................................................................................................................4Methodology............................................................................................................................................6Limitations............................................................................................................................................6AI Harm Mechanisms.............................................................................................................................9Intentional Harm.................................................................................................................................9Harm by Design..............................................................................................................................9AI Misuse........................................................................................................................................10Attacks on AI Systems...............................................................................................................12Unintentional Harm........................................................................................................................14AI Failures......................................................................................................................................14Failures of Human Oversight...................................................................................................16Integration Harm.........................................................................................................................19Discussion..............................................................................................................................................22Conclusion..............................................................................................................................................23Appendix................................................................................................................................................25Authors....................................................................................................................................................27Acknowledgments...................................................................................