Voorbij verklaringen: het ontwerpen van interactieve verklaarbare AI-systemen voor eindgebruikers

With the rapid adoption of Artificial Intelligence (AI) across critical domains, the need for transparency, user trust, and effective human-AI collaboration has never been more urgent. While Explainable AI (XAI) and Interactive Machine Learning (IML) have made notable progress in improving model interpretability and user involvement, most existing approaches treat explainability and interactivity in isolation. Such approaches involve either providing static, one-off explanations or limiting user input to model training phases. However, recent literature in XAI underscores that static explanations alone are often insufficient for meaningful understanding or trustworthy decision-making. This thesis addresses this gap by investigating the design and utility of interactive Explainable AI (iXAI) systems that integrate explanation and interactivity to help end users with minimal AI knowledge (such as domain experts and non-experts) to better understand, evaluate, and guide AI-driven decisions. By adopting a user-centric design and development methodology, the research focuses on three core questions: RQ1. How do iXAI systems influence end users' understanding of AI and calibrate trust in its outputs? RQ2. In what ways can domain experts contribute to refining AI workflows through iXAI interfaces? RQ3. How can iXAI support end users in taking informed action for achieving their end goals? To explore these questions, the thesis presents a series of application-grounded studies in healthcare and recruitment in Chapters 4 to 6. Across these research studies, several iXAI systems were developed that combine diverse explanation types (model-centric, data-centric, and outcome-centric) with interactive mechanisms (such as what-if explorations, conversational interactions, steering and debiasing of prediction models). These systems were evaluated through a combination of quantitative, qualitative, and mixed-method user studies involving both expert and non-expert participants. Through the individual research studies included in this thesis, this thesis offers the following key contributions: Empirical contributions that help us understand how different explanatory and interactive strategies in iXAI system can support user understanding, trust calibration, and actionable insights. Artifact contributions in the form of system designs, source code of the iXAI systems, study procedure and the data collected through the application-grounded user studies. Theoretical contributions through several design guidelines proposed in the research studies for the design of multifacted explanatory systems to involve non-experts and facilitate domain experts for steering and debiasing prediction models without impacting the model performance. Methodological contributions outlining practical approaches for evaluating iXAI systems through application-grounded user studies. The thesis concludes with a critical reflection on the practical challenges encountered and proposes future research directions to address open questions at the intersection of explainability, interactivity, and human-centred AI.