Bounded Morality (Kanwal et al.) defines moral computation

Bounded Morality, a paper by Max Kanwal, Caryn Tran and Patrick Mineault, was submitted to arXiv on 1 Apr 2026 and lays out a formal framework for the computational demands of moral problems. The authors present a 24-page treatment with two figures, published as part of the Proceedings of the AAAI-26 Workshop on Machine Ethics, and assigned arXiv:2607.00002 (doi:10.48550/arXiv.2607.00002).

What is Bounded Morality?

Bounded Morality is a formal framework that frames moral cognition as constrained computation, defining moral problems along two orthogonal dimensions: moral breadth and moral depth. Moral breadth measures the scope of entities treated as morally relevant, while moral depth measures the inferential integration required to evaluate interactions among those entities. The paper positions ethical theories such as deontology, consequentialism and virtue ethics not as competing truths but as locally efficient strategies that occupya particular region of the feasible space defined by these constraints.

The authors extend Herbert Simon's notion of bounded rationality to moral reasoning, arguing that limited resources force an unavoidable tradeoff between breadth and depth and thereby limit the set of moral computations a finite agent can perform.

How does the framework work?

The framework maps moral situations into a two-dimensional demand space and defines feasible strategies as points within that space constrained by computational resources. Moral breadth and moral depth are the primary axes; as capacity shifts, the optimal allocation across those axes changes, producing different locally efficient ethical strategies. Within this space the paper formalizes notions of moral regret and moral progress as metrics for performance "under constraint," tying those evaluations explicitly to resource allocation.

Kanwal, Tran and Mineault treat ethical theories as strategies adapted to different demand regimes rather than universal prescriptions. The framework therefore predicts that increasing or reallocating moral reasoning capacity will change which strategies are feasible or efficient, and that moral alignment for artificial systems depends on the scaling and allocation of moral reasoning capacity rather than direct imitation of human judgments.

Why it matters

The paper reframes debates about machine ethics from matching human judgments to engineering questions about computational capacity and allocation. If ethical theories are strategies optimized for different positions in a constrained demand space, then aligning systems requires specifying how much and what kind of moral computation a system must perform, not just demonstrating fit to human labels. This shifts the engineering target from imitation to capacity design and prioritization.

The authors also introduce formal metrics for moral regret and moral progress under constraint, giving designers a way to quantify tradeoffs when full moral reasoning is computationally infeasible. That quantification could change how researchers evaluate moral competence in agents: not by a single ideal benchmark but by performance relative to resource-bounded optima.

What to watch

Look for follow-up work that applies the framework to concrete agent architectures or benchmarks, and for efforts that operationalize the paper's definitions of moral regret and moral progress into measurable tasks. The paper appears in the AAAI-26 Workshop on Machine Ethics proceedings and may prompt empirical tests that trace how reallocating reasoning capacity alters which ethical strategies an agent uses.