Trust, distrust, and appropriate reliance in (X)AI: a survey of empirical evaluation of user trust

Explainability alone doesn't guarantee trust: The paper finds that empirical investigations show "mixed or inconclusive" results regarding whether explainability methods actually increase user trust in AI systems, challenging the common assumption that more explanation automatically leads to greater confidence in AI.

Trust and distrust are distinct concepts requiring different approaches: The survey distinguishes between user trust and distrust as separate psychological states (not simply opposites), each with different relationships to how people rely on AI systems—a distinction that matters for designing responsible AI systems.

"Appropriate reliance" is the real goal, not maximum trust: Rather than aiming to maximize user trust in AI, the research emphasizes that the actual objective should be "appropriate reliance"—users trusting AI systems to the degree that is actually warranted by the system's actual capabilities and limitations.

Interdisciplinary understanding is essential: The paper draws from machine learning, human-computer interaction, and social sciences to clarify trust concepts, suggesting that responsible AI development requires insights beyond technical fields alone to properly understand how people actually interact with and rely on AI systems.

Current empirical evidence is insufficient: The survey reveals that existing studies investigating how AI systems and explainability methods affect user trust remain inconclusive, indicating researchers need better frameworks and methodologies to properly measure and evaluate user attitudes toward AI systems.