A model that's 98% accurate on your test set can still be completely wrong about why. A famous, real example: a classifier trained to distinguish huskies from wolves turned out to be detecting snow in the background, not the animal — accurate on a test set that happened to reflect the same correlation, and silently broken the moment that correlation didn't hold. Interpretability tools exist to catch exactly this before deployment, not after.
Every other model in this section requires some statistical literacy to interpret. A decision tree doesn't — you can hand the diagram to someone with no ML background and they can trace a prediction themselves, one yes/no question at a time. That transparency comes at a real cost: trees are greedy, and greedy is not the same as optimal.
Global methods describe how a model behaves across the whole dataset: which features it relies on, and what shape each relationship takes.
Local methods explain a single prediction: not "what does this model rely on" but "why did this row come out the way it did". That is the question a declined applicant, a flagged transaction, or a debugging session actually asks.
Explanation methods produce a plot for any model, on any data, whether or not the result means anything. Knowing when the plot is lying is most of the skill.
A model that is accurate on your test set can still be unusable — because it is illegal to deploy without an explanation, because it is right for a reason that will not survive next quarter, or because nobody will act on a number they cannot interrogate.