Disentangling AI Alignment: A Structured Taxonomy Beyond Safety and Ethics

AI alignment is broader than just "safety": The paper identifies multiple distinct alignment aims including safety, ethicality, and legality—meaning responsible AI requires attention to different types of concerns beyond preventing harmful outcomes, each requiring different approaches and expertise.

Alignment can target different levels: The framework distinguishes between "outcome alignment" (what an AI system does) and "execution alignment" (how it does it), suggesting that an AI could produce good results through problematic methods, or vice versa—both matter for responsible deployment.

Different stakeholders may have conflicting alignment needs: By introducing "constituency" as a key dimension, the paper highlights that alignment for individuals may differ from alignment for collectives or society broadly, creating genuine tensions that need explicit acknowledgment rather than being treated as a single unified problem.

Current AI research lacks clear conceptual boundaries: The paper reveals that AI Safety, AI Alignment, and Machine Ethics fields operate with vague and overlapping definitions, leaving researchers and policymakers without clear guidance on how different responsible AI approaches relate to each other or which tools apply to which problems.

"All-things-considered" alignment requires integration across domains: Rather than one field solving AI responsibility, the paper suggests that meaningful alignment requires coordinating insights from multiple disciplines with different aims, scopes, and constituencies—a practical challenge for interdisciplinary AI governance.