The Problem with Current Explainability Metrics

Existing explainability methods often rely on correlation-based metrics, which fail to capture the true causal relationship between input features and model outputs. This leads to "explanation drift," where models appear interpretable but do not reflect the underlying logic driving the decision. The Causal Attribution Score (CAS) addresses this by formalizing interpretability through the lens of causal inference, ensuring that the features identified as "important" are those that actually exert a causal influence on the model's prediction.

Unified Local and Global Attribution

CAS functions as a dual-purpose metric that works across different scales of analysis:

• Local Attribution: By calculating the causal effect of specific features on an individual prediction, CAS provides a robust way to verify if a model's local decision-making aligns with expected causal pathways. This is critical for high-stakes domains like healthcare or finance where individual decisions must be auditable.
• Global Attribution: By aggregating these causal scores across a dataset, CAS allows developers to understand the model's general behavior. This helps identify systemic biases or reliance on spurious correlations that might not be obvious when looking at individual instances alone.

Implementation and Impact

By moving away from purely statistical feature importance (like SHAP or LIME) toward a causal framework, CAS provides a more stable and reliable metric for model evaluation. The approach allows practitioners to quantify the "causal fidelity" of their models, providing a concrete score that can be used to compare different model architectures or training techniques. This shift is essential for moving AI systems from black-box models toward transparent, verifiable architectures that behave predictably in real-world environments.