Effects of Error & Expertise on Example-Based Explainable Artificial Intelligence (XAI)

In an era where Artificial Intelligence (AI) systems are increasingly employed for critical decision-making across various sectors, understanding and trusting these "black box" systems has become paramount. Addressing this issue, the thesis presents a comprehensive investigation in the field of eXplainable AI (XAI), targeting two crucial challenges: (i) identifying effective strategies to elucidate the workings of black-box AI systems, especially when they are in error and (ii) evaluating the impact of these strategies on users with different levels of domain expertise. The research examines the role of post-hoc, example-based explanations in enhancing user comprehension, especially when an AI system has made errorful predictions and users have limited domain knowledge. Employing a twin-system approach, using datasets like MNIST and Kannada-MNIST, the experiments evaluate the effectiveness of these example-based explanations in making AI systems more comprehensible and relevant to a broad range of users, especially those lacking specialised knowledge. The thesis also advances a new explanation algorithm, the Altered Neighbours Method (ANM), designed to adapt explanations to improve on the "traditional" nearest-neighbour-based approach. This method aims to alleviate the challenges posed by system errors and the varied expertise of users, thereby enhancing the overall user performance with these AI systems. Comprehensive user studies are conducted as part of the research to test example-based explanations, deepening user understanding and confidence in AI systems. The findings contribute to the field of XAI, suggesting potential pathways to more transparent, user-friendly, and reliable AI applications, aligning with the increasing integration of AI in various sectors.