Can Transparency Help Clinicians Trust AI? Reframing Trust as an Information Foraging and Sensemaking Loop

Abstract

Large language models (LLMs) are increasingly used in biomedical question-answering (QA) systems, yet clinician adoption remains constrained by challenges in trust. Prior work has largely operationalized trust as a static outcome. Such measures capture whether clinicians trusted a system, but obscure how they arrived at that judgment. We reframe clinician trust in LLM-based biomedical QA as an ongoing process of information foraging and sensemaking. In this view, trust calibration emerges through iterative verification shaped by interface-level transparency. Using an exploratory study design that holds AI-generated answers constant while varying only the structure of evidence access—unlike much prior work that manipulates model accuracy or explanation type—we examined how three participants with medical training navigated, verified, and judged AI-generated biomedical answers across three transparency conditions. Pilot observations indicate that transparency features structured verification behavior rather than directly increasing or decreasing trust: participants selectively attended to the specificity of evidence, revisited claims after inspecting supporting or conflicting snippets, and deferred or rejected answers when evidence appeared incomplete. These observations suggest that effective transparency in clinical AI systems should support controlled cognitive engagement—enabling clinicians to perform their own verification—rather than aim to reduce cognitive effort through persuasive justification.

Ähnliche Arbeiten