SIDEs: Separating Idealization from Deceptive Explanations in xAI

Not all inaccuracies in AI explanations are problematic: The paper argues that some distortions in explainable AI (xAI) methods can be legitimate "idealizations"—similar to how scientists intentionally simplify complex phenomena to make them understandable—rather than inherently deceptive, challenging the view that xAI explanations "must be wrong."

We need better frameworks to distinguish helpful simplifications from misleading ones: The paper introduces the SIDEs framework (Separating Idealizations from Deceptive Explanations) to evaluate whether xAI methods' limitations represent useful simplifications or genuinely deceptive distortions that undermine trust.

Current criticism of xAI may be too absolute: By drawing parallels to how idealizations function successfully in natural sciences, the paper suggests that rejecting all black-box models in high-stakes cases (as some critics propose) may be overly restrictive if we can better understand which explanations are responsibly simplified versus genuinely misleading.

Feature importance and counterfactual methods warrant closer scrutiny: The paper's qualitative analysis examines leading xAI approaches to determine whether their limitations represent acceptable idealizations or problematic deceptions, suggesting these popular methods need more rigorous evaluation.

Transparency about trade-offs is essential: The framework implies that responsible AI deployment requires explicitly acknowledging what xAI methods simplify or distort, rather than presenting explanations as complete truth or dismissing them as entirely false.