Model explainability or interpretability is often demanded, posed as a requirement. But not always. Under what circumstances is it pragmatically necessary (or even legally required) — and, when it is called for, what exactly does it mean? Would it suffice to explain each model prediction by showing what differences in inputs would have changed the prediction? Or must the model be “understood” globally? Is understanding how the model derives its predictions sufficient, even without the why — that is, without causal explanations for the correlations it encodes?

Join this expert panel to hear seasoned experts weigh in on these tough questions.