Computational Virtue
under Double Uncertainty

Humans, animals, and artificial systems alike act in environments that are noisy, ambiguous, and only partially observable. Classical decision theory and much of cognitive science model this as uncertainty about the state of the world: an agent must infer what is going on and choose actions accordingly. Nascent literature in moral philosophy has attended to moral uncertainty: not knowing which moral theory is true or how to weigh competing values. Yet these two dimensions of uncertainty — epistemic and evaluative — are usually treated separately.

Time is orthogonal to space, but we work in spacetime. The descriptive and evaluative worlds we inhabit may be orthogonal, but orthogonality does not justify separate treatment.

Of course, this is intellectually tempting; few concepts are truly thick (being both descriptive and evaluative)² and, famously, is does not imply ought — normatively or prospectively. Indeed, “moral decision-making under uncertainty”, as a recognised research programme, is new and fails to include virtue at all (Tarsney, 2024) and considerations of double uncertainty are, as of yet, either inadvertent and tangential (Friederich, 2025) or non-existent.

Most moral theories take a single aspect of agency as primary. Deontological views foreground rules and constraints. Consequentialist views foreground outcomes. Even decision-theoretic treatments of moral uncertainty largely follow this structure, adding a distribution over moral theories and then maximising expected “choiceworthiness” (MacAskill, Bykvist, & Ord, 2020).³

Virtue ethics, by contrast, shifts the focus from isolated actions to the kind of agent one is becoming. Virtues⁴ are stable, cultivated traits that shape perception, direct attention, temper emotion, and guide action. I agree with Vallor that virtue ethics is ideally suited for managing complex, novel, and unpredictable moral landscapes (2016).

The central claim of the thesis is that we can make philosophical and practical progress by treating virtue as a matter of moral concept learning (Tenenbaum, 1999 & 2000) under double uncertainty, and practical wisdom as a kind of good inductive and attentional bias in that learning process.

The first part of the thesis will articulate and defend a core framework in which virtue is modelled as moral concept learning within a Bayesian moral-world model. At the representational level, the agent maintains and refines a repertoire of thick ethical concepts. These are not thin evaluative predicates attached to a fixed outcome space⁵, but ways of carving up situations that are at once descriptive and evaluative.

At a higher hypothesis level, the agent maintains a distribution over evaluative “mini-theories”: candidate ways of organising and weighting these thick concepts across contexts, linking them to roles, practices, and ready-to-hand consideration sets (Nelle & Cushman, 2022), and projecting them forward in novel cases.

Virtues as Inductive Bias

On this picture, an agent’s virtues are not just behavioural dispositions but the inductive and attentional biases that:

• Structure the space of representable thick concepts and evaluative hypotheses
• Set priors and likelihoods over those concepts and hypotheses
• Guide how the agent seeks and interprets new evidence, and which actions refine their character and reshape their environment

Practical wisdom, in this framework, is a form of good inductive bias under double uncertainty: a configuration of concept space and meta-level learning dynamics that makes agents more likely, over time, to track morally significant features of a changing technosocial world, without prematurely converging on simplistic or self-serving framings.

Distinction from Deontology and Consequentialism

Deontological rules can be recovered as derived regularities within a learned concept space, but they are not taken as fundamental — their normative authority is downstream of the quality of the underlying concept learning. Rather than assuming a fixed outcome space or value function, this approach treats the evaluative concept-space itself as the object of learning. Both the hypothesis space and the possibility space are endogenous to agents’ practices and environment.

Moral life is irreducibly social and historical: agents learn their concepts and virtues in communities (Maier, Cheung, & Lieder, 2025), and those communities themselves change over time. The second part of the thesis will extend the core framework to multi-agent and multi-generational settings.

In a multi-agent environment, each agent maintains a moral world model. They interact in a shared material environment, but their concepts and hypotheses may conflict. The very space of hypotheses they can entertain is itself shaped by cultural evolution and cumulative knowledge.

Technomoral Virtues as Social Epistemic Virtues

Contested morals arise when agents disagree not only about what is happening, but about which concepts apply and how to respond to new possibilities. On my account, technomoral virtues in such environments become social epistemic virtues:

• Normative theory of mind: the capacity to model others as having their own moral world models, and to treat disagreement as evidence both about the world and about one’s own concepts.
• Epistemic humility and charity: caution about applying familiar concepts in novel contexts, and openness to revising them in light of others’ testimony.

Cultural Evolution and Technomoral Wisdom

A culture’s technomoral wisdom can be understood as its capacity to keep its moral concept-learning process corrigible under conditions where both its hypothesis space and its possibility space are in flux, rather than allowing either to be fixed by opaque processes.