New Framework Defines Epistemic Trustworthiness for Generative AI

As generative AI systems become integral to high-stakes professional workflows, the question of when users can genuinely trust and rely on their outputs becomes paramount. Existing responsible AI frameworks often focus on accuracy, fairness, or explainability, but these do not directly address the conditions for "epistemically warranted reliance"—that is, when users are justified in incorporating AI outputs into their own reasoning processes. This research proposes a new normative framework centered on "epistemic trustworthiness," defining what makes an AI system worthy of intellectual reliance. Drawing from philosophical concepts of trustworthiness, the framework outlines three jointly necessary and non-fungible conditions. First, "epistemic humility" requires systems to clearly communicate their competence limits. Second, "epistemic access" demands that systems allow users to inspect, question, and contest outputs within their context. Third, "resistance to epistemic injustice" mandates that systems recognize users as legitimate knowledge agents and avoid marginalizing their expertise. Through case studies in legal, medical, and hiring domains, the paper illustrates how failures in these three conditions can lead to significant harms that go unaddressed by traditional metrics like accuracy or fairness alone. The framework provides a crucial lens for designing and evaluating GenAI systems to foster justified reliance rather than merely correct outputs.