Bounded Morality: Defining the Space of Moral Computation

What Happened

A new preprint on arXiv (2607.00002v1) introduces "Bounded Morality," a formal framework that reframes moral cognition not as adherence to fixed ethical theories but as a computational process operating under constraints. The authors argue that traditional approaches—deontology, consequentialism, virtue ethics—have been implemented in AI as static rules or value functions, which fail to capture the dynamic, resource-limited nature of real-world moral reasoning.

The framework proposes that moral decision-making is fundamentally bounded by computational resources, information availability, and time constraints. Rather than selecting a single ethical theory, Bounded Morality models moral cognition as a search over a space of possible moral computations, where the agent must navigate trade-offs between accuracy, speed, and cognitive cost. This aligns with the broader bounded rationality tradition in cognitive science but applies it specifically to moral domains.

Why It Matters

This paper addresses a persistent blind spot in AI ethics research. Most current approaches to machine ethics treat morality as a set of static principles to be encoded or learned—whether through rule-based systems, reward engineering, or constitutional AI. Yet human moral reasoning is deeply contextual and resource-aware: we make different ethical judgments when pressed for time, when information is incomplete, or when cognitive load is high.

By formalizing morality as a bounded computation, the framework offers a more realistic model for AI systems that must operate in messy, real-world environments. It suggests that the goal is not to build perfectly moral agents that always follow a single ethical theory, but rather agents that can dynamically allocate moral reasoning resources based on context. This has direct implications for safety-critical applications—autonomous vehicles, medical diagnosis, military systems—where perfect moral reasoning is impossible but better-than-random decisions are required.

Implications for AI Practitioners

For engineers and researchers building ethical AI systems, Bounded Morality points to several practical shifts:

First, it challenges the assumption that more ethical reasoning is always better. If moral computation has costs, then systems should be designed with explicit resource budgets for ethical deliberation. This could mean tiered reasoning: fast heuristics for low-stakes decisions, deeper analysis for high-stakes ones.

Second, it provides a vocabulary for discussing moral trade-offs that are currently handled implicitly. When an LLM refuses a harmless request due to over-cautious safety filters, that is a bounded morality failure—the system allocated too many resources to harm avoidance at the expense of helpfulness. The framework gives practitioners a way to diagnose and tune these trade-offs.

Third, it opens the door to empirical evaluation of moral reasoning efficiency. Rather than asking only "did the AI make the right choice?", practitioners can ask "did the AI allocate moral reasoning resources appropriately given the context?" This shifts evaluation from outcome-only to process-aware.

Key Takeaways

• Bounded Morality reframes ethical AI from static rule-following to resource-constrained moral computation, acknowledging that perfect moral reasoning is computationally infeasible.
• The framework provides a formal basis for designing AI systems that dynamically adjust moral reasoning depth based on context, stakes, and available resources.
• Practitioners should consider explicit resource budgets for ethical deliberation, moving beyond the assumption that more moral reasoning is always better.
• Evaluation metrics should expand from outcome correctness to include reasoning efficiency and resource allocation appropriateness.