A predictive model can be accurate without making its reasoning easy to inspect. Interpretability asks what aspects of the input drive predictions and how those influences should be communicated.

For simple linear models, coefficients provide a direct local description: holding other modeled features fixed, changing feature $x_j$ changes the score according to coefficient $\beta_j$. For nonlinear models, more general tools are needed.

Permutation importance measures how much held-out performance worsens when one feature is randomly shuffled. If shuffling a feature destroys useful information while leaving others unchanged, the performance drop suggests that the fitted predictor relies on it.

This measure has limitations. Correlated features can substitute for one another, making each appear less important than the pair really is. Importance also describes the model's predictive dependence, not causal influence in the world.

Interpretability should therefore answer a precise question: global reliance on a feature, explanation of one prediction, sensitivity to a perturbation, or a causal claim are different tasks. A plausible-looking explanation is not evidence that the model learned a scientifically correct mechanism.