From Transparency to Agency: The MICA Framework for Addressing the Understanding-Trust Gap in AI-Assisted Camouflage Analysis

Abstract

Effective integration of Artificial Intelligence (AI) into high-stakes visual detection tasks requires not only technical accuracy but calibrated human trust. Explainable AI (XAI) approaches assume that increasing transparency improves user understanding and, in turn, increases trust and reliance. However, this assumption remains insufficiently validated in operationally relevant domains. This dissertation investigates the relationships among explainability, user understanding, agency, and trust in Camouflaged Object Detection and Segmentation (CODS). A controlled between-subjects factorial study (N = 150) compared an interactive explainable CODS system (MURDOC) to a baseline system (FACE). Participants completed structured detection tasks followed by standardized assessments of understanding, trust, usability, and perceived interactivity. Results revealed a persistent dissociation between understanding and trust. While interactive explainability significantly increased perceived interactivity and engagement, it did not significantly improve user understanding and produced only modest increases in trust. Participants demonstrated high comprehension of system capabilities but only moderate willingness to rely on system outputs. This finding—termed the understanding–trust gap—challenges the assumption that transparency alone produces calibrated trust. In response, this dissertation proposes the Mixed-Initiative Camouflage Analysis (MICA) framework, which reconceptualizes explainability as an active human–AI partnership rather than passive explanation delivery. MICA introduces structured interactive controls, including region-of-interest selection and sensitivity adjustment, to operationalize user agency within the detection process. A functional prototype demonstrates the feasibility of mixed-initiative CODS and establishes infrastructure for future empirical validation. Together, these contributions provide empirical clarification of the limits of transparency-based approaches and introduce a framework that positions structured user agency as a mechanism supporting trust calibration in AI-assisted visual detection.