Mensgerichte Verklaarbare AI
Year: 2025
Type: dissertation
Abstract: Explainable AI (XAI) addresses the lack of transparency current AI decision-making faces, by providing insights into how AI systems derive their output. However, different stakeholders have distinct explanatory needs, making it essential to identify these groups and tailor explanations accordingly. This thesis first examines stakeholder classification in XAI, addressing the challenge of inconsistent categorisation and proposing a framework that distinguishes between role and expertise. This highlights a critical issue in existing XAI methods: prior research predominantly focusses on AI experts, limiting the effectiveness of explanations for a broader audience. The second focus of this thesis is lay users, who struggle to interpret complex AI outputs and are prone to biases. To support their understanding, we propose and assess adapted explanation designs, including simplified visuals and hybrid explanatory modalities. Finally this thesis focusses domain experts who bring valuable prior knowledge when interacting with AI, yet existing explanations often fail to align with their expertise. This thesis explores their preference for data-centric explanations and proposes granular feedback mechanisms to better capture their knowledge, benefiting both domain experts and model trainers. Through a literature review and five iterative, user-centered studies in media, health, and education, this thesis develops an empirical foundation for improving XAI across diverse stakeholder groups.
Source: Lirias
Author: Szymanski, Maxwell
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
SDG: Reduced inequalities
Open Access status: green
OpenAlex ID: W7112228326