Towards Human-Compatible XAI: Explaining Data Differentials with Concept Induction over Background Knowledge

Logical reasoning can make AI explanations more human-understandable: The paper demonstrates that formal logical methods (specifically concept induction over description logics) can generate explanations of AI decisions that are more meaningful to people than traditional black-box approaches, moving toward more "human-compatible" explainability.

Background knowledge improves explanation quality: By leveraging structured knowledge from Wikipedia's category hierarchy as background context, the approach can explain "data differentials" (differences in how AI systems treat different inputs) in ways that connect to concepts people already understand, rather than purely technical feature descriptions.

Ontology engineering techniques have practical XAI applications: Methods originally developed for building knowledge bases and taxonomies can be repurposed to address a key responsible AI challenge—helping stakeholders understand why AI systems make different decisions for different groups or cases.

Structured knowledge enables fairer AI auditing: This approach provides a systematic way to identify and explain disparities in AI system behavior, which is essential for detecting potential bias and ensuring accountability in high-stakes applications where different groups may receive different treatment.