Onto-Epistemological Analysis of AI Explanations

Philosophical assumptions matter in technical design: The paper demonstrates that XAI methods embed hidden philosophical assumptions about what explanations are and how we can know them. These aren't neutral technical choices—they have real consequences for whether explanations are actually valid and trustworthy in different applications.

Engineers may not recognize their own assumptions: XAI methods are typically developed by technical experts who may not explicitly consider the deep philosophical questions underlying their work. This means problematic assumptions about explanations can be built into AI systems without anyone realizing it, potentially leading to false confidence in AI decision-making.

Small technical changes can mean big philosophical differences: The paper shows that seemingly minor modifications to how an XAI method works can actually reflect fundamentally different beliefs about the nature of explanations themselves. This suggests that comparing different XAI methods requires understanding not just their technical features, but their underlying philosophical foundations.

Context matters for explanation validity: By analyzing ontological and epistemological assumptions, the paper implies that an explanation method valid in one domain (like image recognition) may not be appropriate in another (like medical diagnosis or criminal justice), because different fields may require different types of explanations and different standards of proof.

Responsible AI requires interdisciplinary thinking: The paper's core message is that developing trustworthy AI explanations cannot be a purely technical endeavor—it requires collaboration between engineers, philosophers, and domain experts to ensure that the assumptions built into XAI methods are appropriate for their intended use.