Counterfactuals and Causability in Explainable Artificial Intelligence: Theory, Algorithms, and Applications

Current explainability methods lack causal grounding: The paper finds that most popular model-agnostic counterfactual algorithms used to explain AI systems are not built on solid causal theory, meaning they cannot truly provide the "causability" (human-understandable causal explanations) that researchers claim they offer.

Explanations may be misleading rather than truthful: The algorithms reviewed often produce explanations based on spurious correlations rather than actual cause-and-effect relationships, which could lead decision-makers to draw incorrect conclusions about why an AI system made a particular decision.

Risk of biased or erroneous decisions: Because these explanations may not reflect true causal relationships, they have the potential to result in sub-optimal, erroneous, or biased decisions by humans who rely on them—a significant concern for high-stakes applications like healthcare, criminal justice, or lending.

Gap between theory and practice in explainable AI: The systematic review reveals a disconnect between what researchers claim their explainability methods can do and what they actually deliver, suggesting the field needs stronger theoretical foundations before these tools can be trusted for responsible AI deployment.

Need for causal theory in future AI transparency work: The paper calls for new directions in explainable AI that properly ground counterfactual explanations in formal causal theory to genuinely achieve human-understandable explanations rather than just appearing to do so.