Bounded Morality framework redefines AI ethics as resource tradeoffs

A new paper by Max Kanwal, Caryn Tran, and Patrick Mineault introduces Bounded Morality, a formal framework that applies Herbert Simon's concept of bounded rationality to moral cognition. Traditionally, morality has been modeled as adherence to fixed ethical theories—deontology, consequentialism, virtue ethics—implemented as static rules or value functions. The authors argue this is computationally unrealistic for finite agents. Instead, they formalize moral situations along two orthogonal dimensions: moral breadth (how many entities are considered morally relevant) and moral depth (how much inferential integration is needed to evaluate their interactions). Limited resources force an unavoidable tradeoff between these dimensions, defining a feasible space of moral computation.

Within this space, ethical theories are not competing accounts of moral truth but locally efficient strategies adapted to different demand regimes. The framework introduces formal concepts of moral regret and moral progress under constraints. Most importantly, it implies that moral alignment in artificial systems depends on the scaling and allocation of moral reasoning capacity—not on direct imitation of human judgments. The 24-page paper, accepted at the AAAI-26 Workshop on Machine Ethics, offers a computational lens for understanding why different moral systems arise and how AI systems can be aligned under resource constraints.