Judea Pearl won the Turing Award for work on probabilistic and causal reasoning, and this is his attempt to explain to a general audience why he thinks the entire field of machine learning has been solving the wrong problem. Written with Dana Mackenzie, it is a popular book with a serious argument underneath it, and both halves deserve attention. The central idea, and the reason to read it, is the ladder of causation. Rung one is association, seeing that two things co-occur, which is what statistics and virtually all machine learning does.
Rung two is intervention, knowing what happens if you do something, which requires a causal model and cannot be extracted from observational data alone. Rung three is counterfactuals, knowing what would have happened had you acted differently, which requires more still. Pearl's claim is that systems operating at rung one cannot reason about rungs two and three no matter how much data you give them, because the information simply is not present in the correlations. Once you have this framework you cannot stop seeing it.
The model that predicts hospital readmission but cannot tell you which intervention reduces it. The recommendation system that cannot distinguish between causing engagement and being shown to people who were going to engage anyway. The credit model that learned a proxy for a protected attribute and will happily keep doing so at any scale. These are not data problems or architecture problems.
They are rung problems, and no amount of additional training fixes something that is structurally absent. For anyone building models that inform decisions, which is most applied machine learning, this is the most useful conceptual tool in the book and possibly the most useful one available. The teaching of causal diagrams is good enough to be practically useful. You come out able to draw a DAG, reason about confounders, understand what conditioning on a collider does to your estimate, and recognise Simpson's paradox when it appears.
The Simpson's paradox treatment alone is worth the price, because it demonstrates concretely that the same dataset supports opposite conclusions depending on a causal assumption that the data itself cannot adjudicate. Now the problems. Pearl's self-regard is extraordinary and sustained. The book is substantially a narrative of Pearl being right while the establishment resisted him, and by the halfway point the pattern is wearing.
Related, and worse, is his treatment of statistics as a discipline. He has a long grudge, rooted in real historical resistance to causal language in statistics, and he prosecutes it unfairly, presenting a caricature of a field that has done far more careful work on causality than he allows. Readers with a statistics background will find this irritating and it is worth pushing through. His dismissal of deep learning has aged in a complicated way.
Some of it looks prescient, since large models do exhibit exactly the brittleness under distribution shift and intervention that his framework predicts. Some of it looks like a man who underestimated what scale would do. The honest position is that he identified a real limitation and was wrong about how far rung one could be pushed before hitting it. Four point two.
An essential framework, taught well, inside a book whose author cannot get out of its way.