Is It Possible to Make Chatbots Virtuous? Investigating a Virtue-Based Design Methodology Applied to LLMs

TL;DR

The paper investigates whether virtue ethics can guide LLM design by applying Catholic Social Teaching (CST) within the Virtue-Guided Technology Design framework. It extends Conwill et al.'s method to derive five CST-inspired design patterns for LLMs and evaluates them through an IRB-approved, semi-structured interview study with 13 technologists, analyzed via reflexive thematic analysis. Results show that, while patterns are largely favorable for improving robustness, safety, and access, participants note risks such as jailbreaking, cultural generalization, and implementation challenges, highlighting trade-offs. The work argues for smaller, specialized LLMs with ethical patterns and calls for algorithmic development to operationalize virtue-based guidelines, with Rights Reinforcement as a particularly promising next step.

Abstract

With the rapid growth of Large Language Models (LLMs), criticism of their societal impact has also grown. Work in Responsible AI (RAI) has focused on the development of AI systems aimed at reducing harm. Responding to RAI's criticisms and the need to bring the wisdom traditions into HCI, we apply Conwill et al.'s Virtue-Guided Technology Design method to LLMs. We cataloged new ethical design patterns for LLMs and evaluated them through interviews with technologists. Participants valued that the patterns provided more accuracy and robustness, better safety, new research opportunities, increased access and control, and reduced waste. Their concerns were that the patterns could be vulnerable to jailbreaking, were generalizing models too widely, and had potential implementation issues. Overall, participants reacted positively while also acknowledging the tradeoffs involved in ethical LLM design.

Figures (5)

• Figure 1: A summary of the conceptual inquiry, detailing how the six principles of Catholic Social Teaching apply to LLMs. For brevity we only show the points from the conceptual inquiry that ended up inspiring the five design patterns in this paper.
• Figure 2: The five design patterns for virtuous LLMs that we documented through the Virtue-Guided Technology Design process.
• Figure 3: Total numbers of participants who described each pattern as a net positive, net negative, or ambivalent, and the score describing if the pattern was overall seen as strongly favorable, leaning favorable, leaning unfavorable, or strongly unfavorable by participants. All design patterns were favorable, and all patterns were strongly favorable except for Careful Context.
• Figure 4: Overarching themes for what participants valued in corresponding proposed design patterns (green check).
• Figure 5: Overarching themes for participants' concerns with indicated corresponding proposed design patterns (red X).