Explainable Artificial Intelligence and Multicollinearity : A Mini Review of Current Approaches

Feature explanations can be misleading when variables are correlated: When AI systems use features that are highly correlated with each other (multicollinearity), XAI methods may incorrectly identify which features are actually important for the model's decisions, potentially giving false credit to some variables while overlooking others.

Current XAI methods have significant blind spots: The review found that existing explainability approaches have notable limitations when dealing with multicollinearity, meaning many widely-used XAI tools may not be reliably identifying the true drivers of AI decisions in real-world scenarios where correlated data is common.

This is an understudied but critical problem: Despite multicollinearity being a well-known statistical issue, the paper notes that "no review has been dedicated to investigate the current approaches to handle such significant issue," suggesting that responsible AI practitioners may not have adequate guidance on this important challenge.

Better methods are needed for trustworthy AI: The authors call for "future directions" in developing improved approaches, indicating that organizations relying on XAI to justify AI decisions should be cautious, as current methods may not provide fully reliable explanations in many practical applications.