Abstraction, Validation, and Generalization for Explainable Artificial Intelligence

Explanation is fundamentally a communication act: The paper proposes that AI explanations should be understood as deliberate attempts to shift what stakeholders believe or understand, rather than simply displaying information. This reframes how we think about whether an explanation is actually "good"—it's not just about showing your work, but about whether it successfully changes someone's understanding.

XAI methods can be broken down into four testable components: By decomposing any explanation method into (1) what's being explained, (2) how it's explained, (3) who is receiving it, and (4) who is explaining it, the paper suggests we can validate each piece separately rather than treating entire systems as black boxes. This modular approach could make it easier to identify which parts of an explanation system are actually working.

Validation can happen before deployment through component testing: Instead of evaluating every new explanation method from scratch, the paper argues that if individual components have been validated, new combinations of those components can be trusted without extensive re-testing. This could significantly speed up responsible AI deployment by reducing redundant validation work.

Current XAI methods lack unified theory: The paper identifies that without a common framework, different explanation approaches develop in isolation, making it harder to learn systematically what works and what doesn't. The Bayesian Teaching framework attempts to provide this common language so the field can build cumulative knowledge rather than reinventing solutions.