Problem

The rapid advancements in Artificial Intelligence, particularly in Machine Learning and Deep Learning, have led to highly performant systems, but often at the cost of transparency and human understanding. This opacity erodes trust, especially in critical applications requiring traceability, reliability, or accountability for safety and regulatory compliance. Despite the growing popularity of Explainable AI (XAI), there is a lack of unified semantics to define and categorize the diverse explanatory requirements across different applications. Existing definitions tend to be too narrow or focus on presentation rather than the underlying source and detail of explanations, making it difficult to properly evaluate and develop XAI capabilities that truly meet user needs.

Method

The authors introduce a novel, unified categorization for explanations along three dimensions: Source, Depth, and Scope. The Source dimension distinguishes between "Post-Hoc Rationalisation," where explanations come from an external system observing a black-box process, and "Introspective," where explanations derive from the AI system's actual decision-making process, retaining symbolic meaning. The Depth dimension comprises "Attribute" explanations, which describe how features were used in a decision, and "Model" explanations, which also include how the model itself was generated. A sub-dimension for Attribute explanations differentiates between "Attribute Identity" (what attributes were used) and "Attribute Use" (how they were used meaningfully). Finally, the Scope dimension differentiates between "Justification" explanations, which focus on specific decisions, and "Teaching" explanations, which aim to provide a broader understanding of the system's general behavior or decision boundaries.

Results

The paper applies its three-dimensional framework to various AI techniques and application requirements. For instance, in applications like service robots, Post-Hoc Rationalisation and Attribute Identity/Use explanations may suffice, prioritizing user understanding over strict introspectiveness. However, forensic and compliance applications, such as those governed by GDPR or critical safety systems like self-driving cars, demand Introspective, Model, and Teaching explanations to ensure full transparency and accountability. The authors note that much of the contemporary ML literature, especially concerning deep learning, focuses primarily on Post-Hoc Rationalisation, Attribute Identity, Justification explanations. While deep learning can offer partially Introspective and Attribute Use explanations through gradient-based methods, true introspection remains challenging. In contrast, inherently transparent models like Decision Trees, Inductive Logic Programming, and Ripple-Down Rules are capable of providing explanations across all eight octants of the classification scheme, including Introspective, Model, and Teaching explanations, provided their underlying concepts are meaningfully expressed.

Implications

The proposed categorization framework provides a structured approach for practitioners, developers, and users to specify and evaluate the explainability requirements of AI systems. By clearly defining explanation dimensions, it enables a more informed trade-off between explainability and other factors like efficiency, predictive accuracy, and representational flexibility. This framework is crucial for building trust in AI by ensuring that systems can provide the appropriate level and type of explanation for a given application. Future work includes extending these semantics to define "repairability" based on underlying AI techniques and developing specific metrics for different categories of explanations, addressing the growing societal demand for reliability, transparency, and accountability in AI.