Academic Research - United States

The United States leads global XAI research, with major contributions from both academic institutions and industry labs. This means that U.S. institutions set the research agenda for the global XAI community. Historically, the modern XAI field originated from early work at Carnegie Mellon in the 1990s on rule extraction, but gained momentum when DARPA launched the Explainable AI program in 2017, investing $75 million over four years to develop interpretable AI systems. According to Scopus data, U.S. institutions authored approximately 35% of XAI publications in 2024—more than any other country—with particularly strong representation in foundational methods (LIME, SHAP, TCAV) and healthcare applications. Compared to European research that emphasizes formal verification, U.S. work tends to prioritize practical deployment and scalability. The research landscape is characterized by close collaboration between universities and technology companies, with many influential researchers holding joint appointments or moving between academia and industry.

Two distinct schools of thought have emerged in U.S. XAI research, representing fundamentally different philosophies about how to achieve interpretable AI. The post-hoc explanation school, centered at University of Washington and Microsoft Research, focuses on explaining existing black-box models through techniques like LIME and SHAP. In contrast, the inherently interpretable school, led by Duke University's Cynthia Rudin, argues that black-box models should be replaced with transparent alternatives in high-stakes settings. According to Rudin (2019), this distinction matters because post-hoc explanations can be unfaithful to actual model behavior, whereas inherently interpretable models guarantee explanation accuracy. This debate has shaped the field's development and is discussed in the Evaluation & Frameworks page.

Academic Research - Europe

European XAI research is distinguished by its emphasis on formal verification, theoretical foundations, and regulatory compliance. This orientation reflects the European Union's proactive approach to AI governance, particularly through the GDPR's "right to explanation" and the 2024 AI Act. Consequently, European researchers have developed methods that provide mathematical guarantees about explanation properties rather than relying solely on empirical evaluation. Unlike U.S. research that often prioritizes practical deployment, European work tends to emphasize theoretical rigor. However, both traditions have strengths: U.S. methods are more widely adopted in industry, while European methods offer stronger formal guarantees. According to bibliometric analysis, European institutions produced 28% of XAI publications in 2024—less than the U.S. but more than Asia-Pacific—with Germany and the United Kingdom contributing approximately 40% of the European total.

A distinctive feature of European XAI research is the integration of causality and uncertainty quantification into explanation frameworks. Because the Berlin-Tübingen axis (Fraunhofer HHI, MPI, TU Berlin) has established deep expertise in both neural network theory and causal inference, European methods like Layer-wise Relevance Propagation (LRP) offer formal decomposition of neural network outputs that can be verified mathematically. Compared to LIME and SHAP which rely on perturbation-based approximations, LRP provides exact attribution through backpropagation, making it particularly suitable for applications requiring mathematical guarantees. According to Miller (2019), European research has also contributed significantly to understanding explanation from a social science perspective, emphasizing that effective explanations must align with human cognitive processes. This theoretical rigor complements the more empirically-driven approach dominant in U.S. research, as discussed in the Evaluation & Frameworks page.

Academic Research - Asia

Asian XAI research has grown rapidly, contributing approximately 25% of global publications in 2024, with China, India, and Japan as primary contributors. This means that the Asia-Pacific region has emerged as a significant force in XAI development. Studies show that Asian research groups have particularly focused on XAI applications in healthcare, manufacturing, and smart city infrastructure, reflecting regional priorities in AI deployment. For instance, BITS Pilani has pioneered XAI for IoT edge computing, whereas Tsinghua has focused on large-scale NLP explanations. Research demonstrates that Asian contributions have increased by over 300% since 2019, outpacing growth in both North America and Europe. Because many Asian countries are simultaneously developing AI capabilities and regulatory frameworks, XAI research in the region effectively bridges fundamental methods with immediate application needs. In other words, Asian XAI research is characterized by its practical orientation toward deployment rather than theoretical exploration.

A notable characteristic of Asian XAI research is the integration of explainability with edge computing and IoT systems. Groups at BITS Pilani, Tsinghua, and NUS have developed lightweight explanation methods suitable for resource-constrained environments, addressing the challenge of deploying interpretable AI in mobile health applications and industrial IoT. This practical orientation has resulted in techniques optimized for real-time explanation generation. Alternatively, some Asian groups focus on domain-specific applications rather than general methods. On the other hand, the advantage of this approach is faster deployment cycles versus the longer research timelines typical of Western fundamental research. For application-specific requirements, see the Applications & Domains page.

Industry Research Labs

Industry research labs have become central to XAI development because major technology companies face both internal needs (debugging production models, improving user trust) and external pressures (regulatory compliance, liability concerns). Research demonstrates that Microsoft Research and Google DeepMind have produced some of the most widely-adopted XAI tools, including SHAP and TCAV, which are now integrated into production systems serving billions of users. Therefore, industry XAI research tends to prioritize scalability and practical deployment over theoretical elegance. In practice, this means that industry tools often outperform academic prototypes in computational efficiency. Together, these industry labs have collectively advanced the field from academic concepts to production-ready tools used by millions of developers worldwide.

The relationship between industry and academia in XAI is particularly close, with frequent researcher movement and collaborative projects. For example, Scott Lundberg developed SHAP while at the University of Washington before moving to Microsoft Research, while Been Kim's TCAV work bridges her DeepMind position with ongoing academic collaborations. This industry-academia synergy has accelerated the translation of XAI methods from research papers to production tools, as evidenced by the rapid adoption of techniques covered in the Techniques & Methods page.

Industry labs differ from academic groups in their emphasis on production deployment. Whereas academic research often prioritizes theoretical properties (as in Lundberg & Lee's SHAP paper), industry implementations focus on computational efficiency and scalability. For instance, Microsoft's TreeSHAP achieves polynomial-time exact Shapley values specifically for tree ensembles, enabling SHAP deployment on models with millions of predictions per day. Similarly, Google's Integrated Gradients method was designed with gradient computation efficiency in mind, making it practical for explaining large neural networks in production. This practical focus has made industry toolkits like InterpretML, AI Explainability 360, and Captum the standard infrastructure for XAI deployment.

Government Programs

Government initiatives have provided crucial funding and strategic direction for XAI research, particularly in defense and healthcare applications where accountability is paramount. For example, the DARPA XAI program (2017-2021), with over $75 million in funding, catalyzed the modern XAI field by establishing systematic evaluation methodologies. This means that explanation quality could be rigorously measured for the first time. As a result, many current XAI evaluation frameworks trace their origins to DARPA-funded research. Specifically, the program's emphasis on user studies and psychological evaluation influenced subsequent academic research. In other words, DARPA shifted the field from purely technical metrics toward human-centered assessment. Compared to earlier ad-hoc approaches, this systematic methodology represented a significant advancement. The literature shows that DARPA-funded teams produced over 200 publications that collectively shaped modern XAI practices.

Government interest in XAI reflects the growing deployment of AI in consequential domains including national security, healthcare policy, and social services. Because algorithmic decisions in these domains can affect fundamental rights and resource allocation, agencies like NIST have developed risk-based frameworks that specify explainability requirements proportional to decision impact. The EU AI Act, which entered into force in 2024, codifies similar risk-tiered explainability mandates as discussed in the Applications page.

Paper Authors

The authors of "Interpreting Black-Box Models: A Review on Explainable Artificial Intelligence" represent institutions across multiple countries:

Key Journals

• Cognitive Computation (Springer) - Brain-inspired computing and XAI (Editor-in-Chief: Amir Hussain)
• Nature Machine Intelligence - High-impact AI interpretability research
• Artificial Intelligence (Elsevier) - Foundational AI and explanation theory
• Journal of Machine Learning Research - Open-access XAI methods
• ACM TIIS - Human-AI interaction and explanation interfaces
• ACM FAccT - Fairness, accountability, and transparency
• Frontiers in Artificial Intelligence - Open-access XAI research