Modeling Typicality and Novelty Perception with Adaptive Explainable AI in Aesthetic Cognition

Advances in computational power and large-scale perceptual data enable new integrations of machine learning and psychological research on aesthetic judgment. Yet predicting how non-experts evaluate product form remains challenging. This study presents an AI-assisted design support system operationalizing the Unified Model of Aesthetics (UMA) and the MAYA principle (Most Advanced Yet Acceptable) within a human–AI collaboration framework. A two-phase experiment combined Bayesian Adaptive Design Optimization (ADO) with an interactive Explainable AI (XAI) interface. In Phase I, an improved deep learning model classified PC designs into typical, novel, and MAYA-balanced categories, and predictions were compared with ratings from 234 non-design participants. In Phase II, 18 professional designers were randomly assigned to AI-assisted or control groups to assess whether Grad-CAM visualizations and textual explanations improved alignment with lay preferences. Model predictions correlated strongly with human ratings for novelty and typicality. The AI-assisted group achieved higher Top-2 cognitive alignment and greater decision confidence. Thematic analysis indicated that XAI outputs enhanced empathy for non-expert perspectives. By embedding psychological constructs into computational modeling, this approach demonstrates the potential of combining ADO and XAI to address methodological challenges in applying machine learning to perception and preference research.