Architecting Trust in Artificial Epistemic Agents

LLMs are reshaping how we access information: Large language models now function as "epistemic agents" that curate and generate the information we rely on for decisions, often replacing traditional search methods. This shift means their reliability directly impacts the quality of personal and specialized advice people receive.

Calibration to human norms is critical: The paper emphasizes that AI systems must be properly calibrated to match human epistemic standards and societal knowledge practices. Without this alignment, there's a risk of "cognitive deskilling" where people become less capable of independent judgment, and "epistemic drift" where shared understanding becomes distorted.

Current evaluation methods are insufficient: The authors argue that existing approaches to assessing AI systems don't adequately address their role as knowledge curators in complex multi-agent environments. New evaluation and governance frameworks are needed that specifically account for how these systems shape our collective knowledge ecosystem.

Trustworthiness requires multiple dimensions: Beyond accuracy, trustworthy epistemic AI agents must demonstrate epistemic competence, robust falsifiability (ability to be proven wrong), and alignment with human epistemic goals—suggesting that transparency and the capacity to be corrected are as important as getting answers right.

The stakes extend beyond individual users: Because LLMs influence shared knowledge environments that many people depend on, their calibration affects not just personal decision-making but collective societal understanding, making responsible development a high-stakes necessity for democratic and social functioning.