xxAI – Beyond Explainable AI represents a significant milestone in the ongoing mission to bridge the gap between complex machine learning models and human interpretability. Edited by Andreas Holzinger, Randy Goebel, Ruth Fong, Taesup Moon, Klaus-Robert Müller, and Wojciech Samek, this volume encapsulates the forefront of research in explainable artificial intelligence (xAI), extending its boundaries to address new challenges and propose innovative solutions.
In an era where statistical machine learning has revitalized artificial intelligence, leading to sophisticated models like deep neural networks (DNNs) that excel in predictive capabilities, there emerges an inherent trade-off: the increased complexity of these models often comes at the expense of human interpretability. The dichotomy between correlation and causality becomes pronounced, necessitating a deeper exploration into not just how these models make predictions, but why they make specific decisions. This is where the field of explainable AI steps in, striving to create tools and models that are both predictive and interpretable, fostering trust and transparency between AI systems and human users.
The book goes into the limitations of current xAI methods, acknowledging that while they have made significant strides—such as developing robust heatmap-based explanations for DNN classifiers—there is still a pressing need to move beyond these approaches. The editors emphasize the importance of addressing new scenarios, such as explaining unsupervised and reinforcement learning models, and creating explanations optimally structured for human decision-makers with varying levels of prior knowledge. The goal is not merely to interpret the outputs of AI systems but to enhance the transparency, efficiency, and generalization ability of the models themselves.
One of the central themes explored is the concept of causability, which extends explainable AI by incorporating causal reasoning into the explanation process. This is particularly crucial in domains like medicine, where understanding the causal factors behind a diagnosis or treatment recommendation can significantly impact patient outcomes. By formalizing human knowledge to build structural causal models of decision-making, features can be traced back to train AI systems, contributing to more effective and trustworthy AI solutions.
The volume recognizes the critical role of human-in-the-loop approaches, where human domain experts can augment AI with implicit knowledge, experience, conceptual understanding, and context awareness. This synergy between artificial intelligence and human intelligence is posited as a pathway to develop AI systems that not only perform at high levels but also align with human values, ethical principles, and legal requirements. The integration of human expertise is seen as a means to imbue AI with the robustness and generalization capabilities that humans naturally possess, such as the ability to understand context from minimal data and to integrate new insights into their conceptual knowledge base.
The editors and contributors also address the increasing need for trustworthy AI solutions in sensitive and safety-critical application domains. Trusted AI requires both robustness and explainability, balanced with considerations of privacy, security, and safety for individuals. The European General Data Protection Regulation (GDPR) and other legal frameworks underscore the importance of transparency and accountability in AI systems, especially as they become more deeply integrated into aspects of human life that have profound ethical implications.
Andreas Holzinger’s background and contributions are particularly noteworthy. Starting his career as an IT apprentice in 1978 and progressing to an academic with extensive experience in both industry and research, Holzinger brings a unique perspective to the field. His work focuses on the synergistic combination of human-computer interaction (HCI) and knowledge discovery and data mining (KDD), aiming to support human learning with machine learning. By extending advanced methods to include dimensions such as time (information entropy) and space (computational topology), he and his team strive to create interactive software for mobile applications and content analytics, pushing the boundaries of how we interact with and understand complex data.
The book doesn’t shy away from acknowledging the challenges and pitfalls of current xAI methods. It critiques the reliance on black-box statistical ML methods, emphasizing the urgent need to go beyond explainable AI to develop methods that can measure the quality of explanations and build efficient human-AI interfaces. The limitations of current methods, such as their focus on supervised learning models and classification tasks, are discussed, highlighting the necessity to expand the scope of xAI to include unsupervised and reinforcement learning models.
Furthermore, the volume explores the historical context of explainable AI, recognizing that the quest for understanding the “why” behind phenomena is a longstanding pursuit in science. The current resurgence of interest in xAI is seen as a continuation of this endeavour, now applied to the realm of artificial intelligence where the complexity of models often obscures the reasoning behind their outputs.
The contributions within the book are drawn from leading researchers in the field, both from academia and industry, reflecting a clear interdisciplinary approach to problem-solving. The discussions encompass concepts such as explainability, causability, and the interfaces between AI and humans, with applications spanning image processing, natural language processing, law, fairness, and climate science. This breadth illustrates the pervasive impact of AI across various domains and the universal need for explanations that are accessible and meaningful to human users.
xxAI – Beyond Explainable AI also emphasizes the importance of moving beyond mere explanations to directly improving the models themselves. By focusing on transparency and efficiency, the book advocates for the development of AI systems that are not only interpretable but also inherently aligned with human cognition and decision-making processes. This approach requires a fundamental rethinking of how AI models are designed, trained, and deployed, with an emphasis on human-centric principles.
The editors highlight the significance of formalizing human knowledge to create structural causal models, which can be used to enhance AI training and performance. This integration of causality is seen as a critical step in advancing beyond the current state-of-the-art in AI, enabling systems to not just correlate data but to understand the underlying causal relationships that drive outcomes.
Moreover, the volume addresses the challenges of evaluating and comparing different xAI methods. It acknowledges the need for standardized benchmarks and evaluation criteria to assess the effectiveness of explanations, recognizing that the quality of an explanation can significantly impact its utility in real-world applications. The book references tools like Quantus and datasets like CLEVR-XAI, which offer methods to objectively evaluate explanations and foster systematic comparisons between different approaches.
In the realm of practical applications, the book underscores the importance of explainable AI in domains where trust and transparency are paramount. For instance, in the medical field, the ability to explain AI-driven diagnoses or treatment recommendations is crucial for both clinicians and patients. Explanations can facilitate informed decision-making, enhance patient trust, and ultimately improve health outcomes. Similarly, in areas like autonomous driving or financial services, explainability is essential to ensure safety, compliance, and ethical considerations are adequately addressed.
The philosophical underpinnings of the book are rooted in the recognition that artificial intelligence does not exist in a vacuum but is deeply intertwined with human values, ethics, and societal norms. The development of AI systems that can explain their decisions is not merely a technical challenge but also a philosophical one, raising questions about the nature of understanding, trust, and the relationship between humans and machines.
xxAI – Beyond Explainable AI serves as both a comprehensive overview of the current state of explainable AI and a visionary guide for future research. It invites readers to consider not just how we can explain AI systems but how we can fundamentally redesign them to be more transparent, interpretable, and aligned with human cognition. By doing so, the book contributes to the ongoing dialogue about the role of AI in society and how we can harness its potential while mitigating risks associated with opacity and lack of accountability.
This volume is an invaluable resource for researchers, practitioners, and policymakers interested in the forefront of explainable AI. It provides a look into the challenges, methodologies, and philosophical considerations that underpin the quest to make AI systems more interpretable and trustworthy. By bringing together diverse perspectives and cutting-edge research, xxAI – Beyond Explainable AI pushes the boundaries of what is possible in the field, charting a course towards AI systems that are not only powerful but also comprehensible and aligned with human values.
Simon Gros 
Leave a comment