Virtue in Ethics, Science & AI

• Virtue is defined as stable, context-sensitive dispositions that drive agents towards excellence, flourishing, and the common good.
• Formal models quantify virtue by assigning real-valued parameters to ethical traits, enabling trust scoring and utility assessments in epistemic and AI systems.
• Virtue-guided design in human–AI interactions uses value-sensitive patterns and empirical validation to foster ethical, transparent, and responsible technology.

Virtue encompasses a family of concepts unifying moral philosophy, epistemology, scientific methodology, human–AI interaction, and machine learning. It refers to stable, context-sensitive dispositions—character traits, intellectual attributes, or policy-level habits—that drive agents to act, perceive, or decide in ways consistently oriented toward excellence, flourishing, and the common good. Modern research rigorously formalizes virtue for applications ranging from epistemic trust models and value-sensitive design to the engineering of virtuous artificial agents and advanced representation learning systems.

1. Virtue in Moral and Epistemic Philosophy

Virtue ethics, originating in Aristotelian thought and extended through Confucian, Buddhist, and contemporary traditions, asserts that the ethical quality of a life or system hinges on the cultivation and activation of virtues—traits such as courage, justice, temperance, and practical wisdom (phronesis). Unlike deontological (rule-based) and consequentialist (outcome-centric) theories, virtue ethics centers on the agent’s persistent character, emphasizing habituated excellence in reasoning, action, and affect across changing contexts. Intellectual virtue epistemology (notably Zagzebski’s neo-Aristotelian framework) translates this to epistemic agency: a virtuous knower employs conscientiousness, humility, open-mindedness, courage, attentiveness, perseverance, and rigor to regulate evidence gathering, deference to authority, and belief revision (Schwabe, 2 Dec 2025, Schwabe, 20 Dec 2025).

Within virtue-epistemic models, cognitive biases (e.g., confirmation bias, anchoring, Dunning–Kruger) are diagnosed as failures of virtue. Epistemic vices are systematic failures of such dispositions, leading to distorted trust formation or reasoning (Schwabe, 2 Dec 2025, Schwabe, 20 Dec 2025). The MEVIR and MEVIR 2 frameworks instantiate formal models in which trust lattices—graph-structured justifications—are modulated by vectors of virtue levels, which amplify or suppress the propagation of evidential credibility through reasoning chains.

2. Formalization and Models of Virtue

In epistemic and scientific contexts, virtue is not merely qualitative. Models assign real-valued parameters to core virtues, allowing for quantitative modulation of processes such as trust propagation, rational deliberation, and scientific pursuit-worthiness (Schwabe, 2 Dec 2025, Duerr et al., 9 Jan 2025).

Epistemic Trust Example

Let $G = (V, E)$ be a trust lattice. Each node $v \in V$ is assigned a base credibility $\sigma(v) \in [0,1]$. A vector $\vec{v} = (C, H, O, K, T, R) \in [0,1]^6$ captures agent virtues:

$$\mathrm{TrustScore}(G;\,\vec{v}) = \frac{\sum_{p \in \mathcal{P}}\prod_{(u \to u') \in p}\sigma(u') \prod_{\tau \in \mathcal{V}(p)}\phi_\tau(v_\tau)}{\sum_{p \in \mathcal{P}}\prod_{(u \to u') \in p}\sigma(u')}$$

where $\mathcal{P}$ is the set of evidence paths, and $\phi_\tau(v_\tau)$ applies virtue-specific checks to path segments (Schwabe, 2 Dec 2025).

Scientific Pursuit Example

Rational pursuit-worthiness of an idea $I$ is formulated by summing weighted estimates of virtues:

$$\mathrm{Utility}(I) = \sum_{V \in \mathcal{V}} w_V p_V(I) - \sum_{W \in \mathcal{W}} w_W p_{\neg W}(I)$$

where $p_V(I)$ is the probability that $v \in V$0 manifests virtue $v \in V$1, and $v \in V$2 reflects its context-dependent importance (Duerr et al., 9 Jan 2025).

Virtue-economic accounts impose rationality constraints: historical circumspection (fit to scientific exemplars), peer-level coherence, and systematicity across judgments, ensuring virtue assessments are not ad hoc or idiosyncratic.

3. Virtue in Artificial Intelligence and Machine Learning

Virtue is operationalized in reinforcement learning and AI as policy-level dispositions or traits:

• In multi-objective RL, virtues become dimensions (e.g., honesty, justice) with cumulative scores $v \in V$3, where $v \in V$4 indicates virtue-relevance (Ghasemi et al., 3 Dec 2025).
• Affinity-based regularization enforces proximity to a "virtue-prior" policy through penalization of deviations $v \in V$5, guaranteeing stable trait expression across distributional shifts (Ghasemi et al., 3 Dec 2025, Vishwanath et al., 2022).
• Ethical trade-offs are handled explicitly through Pareto multi-objective optimization and constrained formulations, ensuring that virtues are not collapsed into a single scalar reward but preserved as auditable, durable, policy-level characteristics (Ghasemi et al., 3 Dec 2025).

Explainable AI techniques attribute action selection to virtue dimensions, facilitating interpretability and systematic auditing of virtue expression (Vishwanath et al., 2022, Tzachristas et al., 27 Jun 2026).

Table: Virtue Formulation Across Domains

Domain	Formalization	Primary Virtues Modeled
Epistemic Trust (MEVIR)	Vector-modulated trust score	Conscientiousness, humility, openness, courage, attentiveness, rigor
Scientific Methodology	Weighted cost–benefit utility	Empirical adequacy, coherence, simplicity, fertility, testability
AI/ML (Virtuous RL/Agents)	Policy-level trait scores; affinity regularization	Honesty, justice, temperance, courage, practical wisdom

4. Virtue Ethics in Human–AI System Design

Virtue ethics guides the design of socio-technical systems, especially in HCI and technology for the common good (Gorichanaz, 2022, Conwill et al., 17 Jan 2025). The focus is on cultivating user and system virtues via explicit design patterns:

• Design patterns for social media—such as "Chats Over Feeds," "Friends Over Followers," and "Notification Intentionality"—are explicitly mapped to virtues (e.g., dignity, subsidiarity, solidarity) and empirically validated with technology practitioners (Conwill et al., 17 Jan 2025).
• Value Sensitive Design (VSD) is extended by using virtue traditions to supply normative orientation, avoiding mere stakeholder preference capture (Conwill et al., 17 Jan 2025).
• The process involves identifying target virtues, conceptually clarifying their local meaning, surveying and abstracting virtuous features in current systems, empirically validating virtuous patterns, and iterating for diverse cultural contexts (Conwill et al., 17 Jan 2025).

Virtue cultivation is linked to seven practical dimensions: moral habituation, relational understanding, reflective self-examination, self-direction, moral attention, prudential judgment, and extension of moral concern (Gorichanaz, 2022).

5. Virtue in Representation Learning and Embedding Models

Virtue, as a term, is also appropriated as an acronym for advanced systems in embedding and retrieval. In "VIRTUE: Versatile Video Retrieval Through Unified Embeddings," it denotes a unified framework for video retrieval using a single multimodal LLM backbone with low-rank adaptation (Halbe et al., 17 Jan 2026). Here, "virtue" characterizes:

• Efficient unified embedding via EOS token summarization.
• Broad functional coverage: corpus-level retrieval, fine-grained moment localization, and composed multimodal queries within a single model (Halbe et al., 17 Jan 2026).
• Contrastive alignment loss for shared text–video embedding space, supporting zero-shot performance across heterogeneous retrieval tasks.
• Empirical outperformance of prior MLLM-based methods and matching of specialized dual-encoder models using orders of magnitude less data (Halbe et al., 17 Jan 2026).

Similarly, "VIRTUE: Visual-Interactive Text-Image Universal Embedder" extends vision–LLMs to support region-level prompting (points, boxes, masks), fusing segmentation and vision-language features for localized, entity-aware embedding and state-of-the-art performance on complex retrieval benchmarks (Wang et al., 1 Oct 2025).

6. Evaluation, Benchmarking, and Operationalization of Virtue

Virtue is rendered operational in benchmarking LLMs and AI systems:

• The "VirtueMap" framework profiles LLMs by eliciting rankings across ethical dilemmas, each calibrated to express five Aristotelian virtues: practical wisdom, justice, truthfulness, courage, temperance (Tzachristas et al., 27 Jun 2026). Ground truth orderings are rigorously validated in human samples, and Borda alignment quantifies how closely LLM outputs manifest each virtue dimension.
• In empirical studies, LLMs exhibit high mean rank consistency (90.3%) and varying strengths across virtues, with models differentiated most on Courage, Temperance, and Justice dimensions (Tzachristas et al., 27 Jun 2026).
• System behavior is thus not reduced to categorical correctness, but profiled in multi-dimensional virtue space, enabling granular auditing and targeted fine-tuning (Tzachristas et al., 27 Jun 2026).

7. Implications, Limitations, and Future Directions

Virtue-centric models support normative pluralism, intersubjective critique, and transparent control in system design, scientific evaluation, AI training, and epistemic trust formation (Duerr et al., 9 Jan 2025, Schwabe, 2 Dec 2025, Conwill et al., 17 Jan 2025). Methodological pluralism is explicit: evaluators may weight virtues differently, but such variation is subject to rationality constraints and empirical testing (Duerr et al., 9 Jan 2025).

Key open problems include: scaling virtue-informed frameworks to more diverse data and contexts, codifying virtue trade-offs at scale in machine learning, improving cross-cultural calibration in empirical virtue benchmarks, and extending design patterns to encompass broader solidarity and common good concerns (Schwabe, 20 Dec 2025, Conwill et al., 17 Jan 2025, Gorichanaz, 2022).

In summary, virtue now substantiates a rigorous bridge between ancient ethical theory and modern technical systems: as a formal, quantitative, and empirically evaluable construct central to epistemic trust, scientific rationality, the engineering of ethical AI agents, and interface design for collective flourishing.