Beneficent Intelligence: A Capability Approach to Modeling Benefit, Assistance, and Associated Moral Failures through AI Systems

AI ethics needs better language to address real harms: The paper argues that current AI ethics discussions lack the formal framework to properly identify and address the diverse ethical problems that emerge when AI systems affect people's lives, particularly in high-stakes domains like healthcare, criminal justice, and finance.

Meaningful AI benefit requires protecting both well-being and rights: Drawing on established ethical philosophy, the framework identifies that truly beneficial AI systems must do more than just produce good outcomes—they must also preserve people's fundamental rights and enable individuals to pursue their own life plans rather than having outcomes imposed on them.

AI systems can fail morally in specific, identifiable ways: The paper characterizes five distinct failure modes—unjustified paternalism, coercion, deception, exploitation, and domination—that represent how AI systems can harm people even when they appear to be helping, providing concrete language for recognizing these problems in practice.

Ethics must be built in from the start, not added later: The authors emphasize that an "ethics-led approach" is necessary from the inception of AI system design, rather than treating ethical concerns as an afterthought, suggesting that responsible AI requires fundamental changes to how systems are developed.

Individual capabilities and autonomy matter as much as outcomes: The framework prioritizes whether AI systems expand or restrict people's ability to make meaningful choices about their own lives, not just whether the systems produce beneficial results.