Model explainability is a field with an unusually high ratio of tooling to understanding. Everybody has SHAP installed. Far fewer people can tell you what the Shapley values are actually attributing, what assumption breaks when your features are correlated, or why two runs of LIME on the same instance gave different explanations. This book is the reason to close that gap, and the thing that makes it better than the alternatives is structural.
Every method chapter ends with a disadvantages section, written plainly, and Molnar does not soften them. Permutation feature importance is misleading with correlated features. Partial dependence plots hide heterogeneous effects and can describe regions of feature space that contain no real data. LIME's explanations are unstable under resampling.
Shapley values are expensive and the common approximations break their theoretical guarantees. This is the material that vendors and tutorials skip, and it is exactly the material you need, because the risk in explainability is not failing to produce an explanation, it is producing a confident wrong one and having somebody make a decision on it. The organisation is sensible. Interpretable models first, covering linear regression, logistic regression, decision trees, rule-based approaches and generalised additive models, which is the right starting point because the honest first question about explainability is whether you needed a black box in the first place.
A great many production models would lose almost nothing in accuracy as a boosted tree or a GAM and would be dramatically easier to explain, audit and defend. Then local explanation methods, then global ones, then a section on neural networks. The third edition is a real improvement, adding leave-one-feature-out and ceteris paribus methods, reworking the introductory chapters, moving to the Palmer penguins dataset instead of the tired alternatives, and generally tidying up. Molnar has also maintained this for years while publishing his own workshops and consulting alongside it, and the free version has never been degraded into a lead magnet, which deserves credit given how many free technical books turn into a chapter and a form.
Where it is thin is deep learning and language models. There is neural network interpretation material, covering feature visualisation, saliency and concept-based approaches, and it is less developed than the tabular material. On LLMs there is essentially nothing, which is where the most consequential interpretability questions now sit. Why did the model refuse.
What in the context drove that answer. Which training data shaped this behaviour. Mechanistic interpretability has become a serious research field since this book took its shape and the book has not absorbed it. That is a gap in the field's textbooks generally rather than a failure of this one, and it does mean you should not come here expecting answers about transformers.
The other thing to know is that this is a reference rather than a course. There is no problem set, no project, no path. You read the chapter for the method you are about to use. That is the right form for the content and it does mean you need discipline to get through it without a deadline.
Four point six. The best free technical book in its area, honest in a way technical writing usually is not, and behind the frontier on language models.