AI Virtue: What is "Good" Knowledge in the Age of Artificial Intelligence?

What Happened

A forthcoming paper in Modern Fiction Studies (2027) applies digital humanities methods to map "epistemic virtues"—criteria like truth, objectivity, and reliability—in the context of AI-generated knowledge. By analyzing how literary and cultural texts have historically defined "good" knowledge, the research asks a deceptively simple question: as AI systems increasingly produce and curate information, what standards should we use to judge whether that knowledge is valuable or trustworthy?

The paper bridges two fields that rarely converse: literary criticism's nuanced understanding of how knowledge claims are constructed, and AI's engineering focus on accuracy and performance. It treats epistemic virtues not as fixed philosophical ideals but as historically contingent values that shift across contexts—and that AI systems may be reshaping, often invisibly.

Why It Matters

This research arrives at a moment when AI-generated content is flooding every domain, from academic publishing to customer support. The dominant metrics for evaluating AI outputs—perplexity, BLEU scores, human preference ratings—are technical proxies that say little about whether the knowledge produced is actually good in any meaningful sense. A model can generate factually correct, coherent text that is nonetheless misleading, biased, or epistemically hollow.

The paper's intervention is to force a conversation about epistemic virtue that goes beyond "is it true?" to ask: Is the knowledge transparent about its provenance? Does it acknowledge uncertainty? Does it serve the user's genuine need for understanding, or merely produce plausible-sounding answers? These are questions that literary scholars have explored for decades through the study of narrative authority, unreliable narrators, and the social construction of knowledge. Applying those methods to AI reveals that many current systems are optimized for persuasive knowledge rather than virtuous knowledge—a distinction with profound consequences for trust.

Implications for AI Practitioners

For engineers and product teams, this research suggests several actionable shifts:

• Rethink evaluation metrics. Accuracy alone is insufficient. Practitioners should consider incorporating measures of epistemic quality: source transparency, acknowledgment of competing viewpoints, calibrated confidence expressions, and resistance to sycophancy (telling users what they want to hear).

• Design for epistemic humility. Current LLMs project unwarranted certainty. Systems that can express "I don't know," cite sources, and distinguish between established facts and plausible inferences would better embody the epistemic virtue of honesty—and likely earn greater user trust over time.

• Consider domain-specific virtue profiles. The "good knowledge" required for medical diagnosis differs from that needed for creative writing or historical analysis. AI systems should be configurable to prioritize different epistemic virtues depending on context, rather than applying a one-size-fits-all optimization.

• Prepare for regulatory scrutiny. As governments move to regulate AI, concepts like "truthfulness" and "reliability" will be operationalized into legal standards. Understanding the philosophical and cultural history of these concepts—as this paper does—gives practitioners a head start in building compliant, defensible systems.

Key Takeaways

• A forthcoming humanities-AI crossover paper uses digital methods to map how "good knowledge" has been defined historically, challenging AI's narrow focus on factual accuracy.
• Current AI evaluation metrics (perplexity, preference ratings) fail to capture epistemic virtues like transparency, humility, and source accountability.
• AI practitioners should develop metrics and design patterns that reward epistemic virtue, not just persuasive output.
• Domain-specific and culturally aware definitions of "good knowledge" will be essential as AI regulation matures and user trust becomes a competitive differentiator.