The Feature Understandability Scale for Human-Centred Explainable AI: Assessing Tabular Feature Importance

Explainability is not one-size-fits-all: The paper emphasizes that different audiences (e.g., regulators, domain experts, general users) have varying needs for AI explanations, meaning a single explanation approach cannot serve everyone equally well.

Feature understandability matters more than we realize: The research identifies a critical gap—most AI explanation methods focus on how features influence predictions, but ignore whether users actually understand what those features mean. This disconnect undermines the effectiveness of explanations.

New measurement tools have been developed: The authors created two validated scales (one for numerical features, one for categorical features) that can quantitatively measure how well users understand input features, enabling AI designers to prioritize more understandable features when building explanations.

Feature selection should precede explanation: Rather than explaining with whatever data is available, system designers should strategically pre-select understandable features for explanations—a practice the paper suggests is currently overlooked in most explainability approaches.

Auditing AI requires human-centered design: The work underscores that responsible AI auditing depends on designing explanations around actual human comprehension, not just technical accuracy, making user understanding a key metric for responsible AI systems.