Bridging Explainability and Interactivity to foster Trust and increase Interpretability in AI-systems
Marco Fries, Louisa Maria Sauter, Julia Nießner, Markus Keyser, Thomas G. Ludwig · 2025
AI-based text recognition in technical drawings faces challenges due to their complexity and variability. While Optical Character Recognition (OCR) improves data extraction, AI systems often operate as black boxes, limiting trust and interpretability. Explainable AI (XAI) techniques, such as SHapley Additive exPlanations (SHAP) and Gradient-weighted Class Activation Mapping (Grad-CAM), enhance transparency but are typically designed from experts for experts, making them inaccessible to broader user groups. Interactive approaches like dialogic structures, including feedback mechanisms, offer a potential solution by guiding and assisting users throughout the sensemaking process, clarifying questions. This study examines how XAI and user feedback impact transparency, interpretability, and trust in AI-based text recognition. We developed four prototypes with different levels of explainability and interactivity and conducted a qualitative study. We show that combining XAI with feedback increases transparency and trust compared to standalone methods. While explainability in the form of Grad-CAM heatmaps improved traceability, manual corrections fostered user engagement and led to deeper insights. However, XAI often failed to explain why errors occurred, leaving users uncertain about AI decision-making. Our findings highlight the need for holistic AI explainability that combines visual explanations with interactive elements, like context-aware chatbots, and other concepts. This study contributes to human-centered AI, emphasizing the importance of interactive XAI systems.