For decades, moral cognition has been modeled as the application of static ethical theories—deontology, consequentialism, virtue ethics—translated into fixed rules or value functions. A new formal framework proposed by a research team breaks that tradition: Bounded Morality shifts the focus to finite agents, where computational capacity becomes the critical variable.

Inspired by Herbert Simon’s bounded rationality, the work identifies two orthogonal dimensions characterizing any moral situation: moral breadth, the scope of entities deemed morally relevant, and moral depth, the inferential integration needed to evaluate their interactions. With finite computational resources, every agent faces a trade-off: pushing on breadth means sacrificing depth, and vice versa. This space of possibility defines what is actually computable by a finite agent, introducing a formal notion of moral regret and moral progress under constraint.

Traditional ethical theories, from this perspective, are no longer seen as competing accounts of moral truth but as locally efficient strategies adapted to different computational demand regimes. A rigid deontological system, for example, may be an optimal response when the agent cannot afford deep evaluations across the entire set of involved entities.

For those working on AI alignment and Large Language Models, the message is clear: teaching a model to imitate human judgments is not enough. Moral alignment depends on the ability to allocate moral reasoning capacity in a scalable way. In on-premise deployment scenarios—where memory, throughput, and latency are bounded by specific hardware choices, think of VRAM available on local inference GPUs or the energy cost of deep evaluations—the breadth-versus-depth trade-off becomes concrete. A self-hosted system operating air-gapped might need to reduce moral breadth to preserve analytical depth, or sacrifice depth to monitor a wider set of stakeholders.

The Bounded Morality framework does not offer recipes, but a formal language to make these dilemmas explicit. Shifting the debate from behavioral imitation to the management of a moral computational budget could become a key step for those designing LLMs aimed at ethically sensitive decisions, from content moderation to resource allocation in regulated environments.