(Solved): This section presents a risk taxonomy drawn from 16 policies of 9 foundation model developers (Fig ...

This section presents a risk taxonomy drawn from 16 policies of 9 foundation model developers (Figure 2). We focus on two types of company policies that seek to govern generative AI in order to address specific risks: platform-wide acceptable use policies and model-specific acceptable use policies [49]. An overview of the company policies we consider in this study organized into 13 sets is listed in Table 1. ?abie 1 : Uverview of the company pol?c?es ( 10 documents organized into 15 sets) we consider in th1s study. \table[[L1-Name,(3),(5),Al,\oo ,OS,G,G,8,4,S.,Es,%,%],[\table[[System & Operational],[Risks (6)]],6,4.5,5,3,5,5,5,3,1,5,4,1,3],[\table[[Content Safety Risks],[(17)]],13,11,14,9,12,11,11,11.5,11.5,14.5,14.5,6,11],[\table[[Societal Risks],[(14)]],6,9,8,4,4,2,3,8,0,5,9,2,7],[\table[[Legal & Rights-Related],[Risks (8)]],6.5,7,5,7,8,6,7,6,5,7,8,3,6],[Iotall(%),70%,70%,71%,51%,64%,53%,57%,63%,39%,70%,79%,27%,60%]] Table 2: Risk categories covered by each company's policies at level-3 risks in our AIR Taxonomy. Categories that are referenced without further elaboration are counted as 0.5 .4.1.A European Union Figure 3: EU regulations specified AI risks mapped as 23 level-3 categories in the AIR 2024.AI Risk Categorization Decoded (AIR 2024): Summary This paper introduces the AI Risk Taxonomy (AIR 2024), a comprehensive framework for categorizing risks associated with generative AI models and systems. It addresses the need for a unified language for evaluating AI safety by drawing on regulations and policies from governments and companies around the world. Here's a breakdown of the key points: Motivation: Different sectors (government, industry) use diverse terminology for AI risks, hindering communication and collaboration. AIR 2024 aims to bridge this gap. Methodology: Researchers analyzed policies from eight government bodies (EU, US, China) and 16 companies to identify a total of 314 unique risk categories. Structure: The taxonomy is hierarchical, with four levels: Level 1: Broad categories (System & Operational, Content Safety, Societal, Legal & Rights) Level 2: More specific categories (e.g., Operational Misuses under System & Operational) Level 3: Even more granular breakdowns (e.g., Automated Decision-Making under Operational Misuses) Level 4: Most detailed risk categories explicitly mentioned in policies (e.g., Financing eligibility/Creditworthiness under Automated Decision-Making) Benefits of AIR 2024: Standardized Framework: Enables consistent AI safety evaluations across regions and sectors. Insights into Corporate Risk Management: Reveals how companies prioritize and perceive AI risks. Comparative Analysis of Regulations: Helps understand how different governments approach AI governance. Improved Communication: Provides a common language for policymakers, researchers, and industry leaders. Safer AI Development and Deployment: Guides efforts to mitigate risks and promote responsible AI practices. The paper also discusses challenges encountered during taxonomy development, such as the use of varying terminology in different policies. It emphasizes the importance of ongoing updates as regulations and company policies evolve. Overall, AIR 2024 is a valuable resource for anyone involved in the development, deployment, or governance of generative AI. It promotes a shared understanding of AI risks and facilitates collaboration towards safer and more beneficial AI applications. can analyze scholastic the images of the paper?Figure 1: Overview of the AI risk taxonomy derived from 24 policy and regulatory documents, encompassing 314 unique risk categories. Charts on the right-hand side map to major AI regulations.