Even-if Explanations: Formal Foundations, Priorities and Complexity

Different AI models have different levels of explainability: The paper finds that linear models and tree-based models are "strictly more interpretable than neural networks," meaning simpler models can provide clearer explanations for their decisions. This has practical implications for high-stakes decisions where explainability is critical.

"Even-if" explanations offer a different perspective than counterfactuals: While most explainability research focuses on counterfactual explanations (what would need to change for a different outcome), this paper emphasizes semifactual "even-if" explanations that show what factors don't matter for a decision. For example, an even-if explanation might reveal that a loan was denied regardless of your zip code, highlighting what didn't influence the outcome.

One-size-fits-all explanations may not serve users well: The paper introduces a preference-based framework allowing users to customize explanations based on what matters to them. This recognizes that different stakeholders—such as regulators, affected individuals, or model developers—may need different types of information to understand AI decisions fairly.

Computing explanations has real computational limits: The paper analyzes the computational complexity of generating these explanations across different model types, finding that some explanation queries are computationally expensive. This matters because it affects whether explanations can be provided quickly in real-world applications where timely decisions are needed.