Explainable AI needs formalization

Current explainability tools may give false confidence: Popular XAI methods "systematically attribute importance to input features that are independent of the prediction target," meaning they can highlight factors that actually have nothing to do with why an AI made its decision. This creates a misleading sense that we understand the system when we may not.

XAI methods lack rigorous standards: The paper argues that XAI researchers have not formally defined what "explanation" means or established clear criteria for measuring whether an explanation is actually correct. Without these standards, there's no reliable way to validate whether an XAI tool is working as intended.

Real-world applications are at risk: Because current XAI methods cannot reliably answer questions about models and their training data, they have limited usefulness for critical tasks like "diagnosing and correcting data and models" or identifying where interventions are needed—exactly the applications where we most need trustworthy explanations.

One-size-fits-all explanations won't work: The paper suggests that different use cases (healthcare, hiring, finance, etc.) will require different definitions of what makes an explanation "correct," meaning the field needs tailored approaches rather than universal XAI solutions.

The field needs a reset: Researchers should start by formally defining the specific problem they're trying to solve before designing XAI methods, then evaluate those methods against objective performance metrics—a more rigorous foundation than currently exists.