Explanation from Specification

Explanations should be customized to user needs rather than one-size-fits-all: The paper argues that instead of defaulting to familiar explanation types like linear models or decision trees, AI systems should ask users what kind of explanation they actually need. This is particularly important in high-stakes domains like medicine and forensics where different stakeholders may require different types of information.

Domain experts should help design explanation types: The authors propose that specialists in fields like medicine, chemistry, and forensics should collaborate with AI researchers to specify what kinds of explanations make sense in their context. For example, a chemical query language could let users ask about molecular properties when an AI classifies molecules, rather than receiving generic feature importance scores.

Different AI models need different explanation approaches: The paper demonstrates that explanation methods should be tailored to the underlying model—showing how argumentation theory can explain Bayesian networks and how graph-based approaches can explain graph neural networks. This suggests there's no universal explanation solution across all AI systems.

Explanation design is fundamentally a knowledge representation challenge: The authors frame building good explanations as an exploratory task requiring algorithm designers to understand what types of explanations are actually computable and practical for each domain, not just theoretically possible.

Philosophy of science should inform AI explainability: The work is grounded in philosophical theories of explanation, suggesting that responsible AI development benefits from engaging with how scientists and philosophers think about what makes something truly "explained."