Delivering Inflated Explanations

Inflated explanations reveal feature flexibility: Unlike traditional explanations that only show which features matter, inflated explanations demonstrate the range of values each feature could take while still producing the same decision—helping users understand whether a precise value is critical or if nearby values would work equally well.

Current AI explanations may be too narrow: The paper identifies a limitation in formal explainability approaches: standard abductive explanations only confirm that selected features with their given values lead to the same decision, but don't show what other values those features could have taken, potentially giving users an incomplete picture of how the AI system actually works.

Better understanding with modest computational cost: The authors demonstrate that computing these more informative inflated explanations requires only slightly more computational effort than traditional explanations, suggesting this approach could be practically implemented without significant efficiency trade-offs.

Implications for AI transparency and trust: By showing the full range of acceptable feature values, inflated explanations could help stakeholders better understand AI decision-making boundaries and identify whether decisions are robust or dependent on precise, potentially fragile conditions—important for assessing whether an AI system's reasoning is trustworthy.