A Synthetic Ground-Truth Framework for the Evaluation of Explainable AI Methods

Current XAI evaluation methods are fundamentally flawed: The paper demonstrates that existing approaches measure only whether explanations match a model's predictions (fidelity), not whether they actually reflect how the model makes decisions—meaning different explanations can score similarly while providing contradictory interpretations of the same model.

A new evaluation framework using synthetic data can reveal hidden problems: By creating controlled datasets where the importance of different input features is known by design, researchers can establish "ground truth" explanations and test whether popular XAI methods accurately identify them across images, tables, and time series data.

Nine widely-used XAI methods show significant limitations: The experimental evaluation of common explanation techniques revealed substantial shortcomings when tested against this more rigorous synthetic ground-truth standard, suggesting that many explanations currently trusted in practice may be unreliable.

The absence of ground truth has enabled misleading explanations to proliferate: Without reliable evaluation procedures, organizations deploying XAI methods have had no way to verify whether explanations are actually trustworthy, creating a gap between perceived and actual transparency in AI systems.

Responsible AI deployment requires better evaluation standards: This framework provides a practical path forward for developing more rigorous testing procedures across different data types, which is essential before relying on explanations for high-stakes decisions.