How AI Explanations Affect Trust and Decision-Making: An Experimental Analysis of Explainability Formats

AbstractWe combine subjective measures (trust in the system, perceivedtransparency, perceived usefulness, and self-rated understanding on7-point Likert scales) with behavioral outcomes (decision accuracyagainst ground truth, response time, acceptance/reliance rate, andpatterns of overrides). Calibration is quantified using Brier scores,mean calibration error (the mean absolute discrepancy between statedconfidence and actual accuracy), and reliability diagrams. Primary inferential tests use repeated-measures ANOVA to examine main effectsand interactions of explanation format and AI reliability; mixed-effectsmodels (participant as a random effect, trial as nested observations)probe trial-level variability and order effects. Secondary analyses explore mediation (does perceived understanding mediate the effect ofexplanation type on acceptance?) and moderation by prior AI experience and domain familiarity.