AI Risk Categorization Decoded (AIR 2024): From Government Regulations to Corporate Policies

Governments and companies define AI risks differently: The paper identifies significant overlaps and discrepancies between how public regulators (EU, US, China) and private companies conceptualize AI safety risks, suggesting that a unified language is needed for consistent risk management across sectors.

AI risks fall into four major categories: The taxonomy organizes 314 unique risk categories into System & Operational Risks, Content Safety Risks, Societal Risks, and Legal & Rights Risks, providing a structured framework that organizations can use to systematically evaluate generative AI safety.

Current approach lacks standardization: With eight different government policies and 16 company policies analyzed, the fragmented landscape of AI risk definitions creates challenges for information sharing and best practice adoption—the paper's unified framework aims to address this coordination problem.

The taxonomy enables cross-sector learning: By mapping connections between government regulations and corporate policies, the framework facilitates knowledge transfer between public and private sectors, potentially improving how both groups identify and mitigate AI risks.

Generative AI safety requires comprehensive evaluation: The identification of 314 distinct risk categories demonstrates that responsible AI governance cannot focus on a single risk type but must address technical, content-based, societal, and legal dimensions simultaneously.