A Theoretical Framework for AI Models Explainability with Application in Biomedicine

Explanations require both technical accuracy and human understanding: The paper defines explanations as a combination of evidence from the AI model itself and how humans interpret that evidence, meaning a good explanation must satisfy both the computer scientists (being technically faithful) and the end users (being understandable).

Two critical properties must be balanced: The framework identifies "faithfulness" (whether an explanation truly describes how the model actually works) and "plausibility" (whether the explanation seems convincing to users) as essential, suggesting that explanations can sometimes seem reasonable to people while not accurately reflecting what the model is actually doing.

Shared terminology is missing in explainable AI: The authors note that despite growing research in XAI, the field "still lacks shared terminology and a framework," which could lead to confusion about what different explanation methods actually accomplish—a significant concern for responsible deployment in high-stakes domains like biomedicine.

One-size-fits-all explanations are insufficient: By recognizing that explanations combine model evidence with human interpretation, the framework suggests that effective explanations may need to be tailored to different audiences (doctors, patients, regulators) rather than providing a single standardized explanation.

The framework applies practical analysis to existing methods: The paper uses biomedical AI as a case study to demonstrate how their theoretical framework can evaluate common explanation techniques, providing a tool for practitioners to assess whether their explanations are both faithful and plausible.