Artificial Intelligence 2018 Daily Pick
The Book of Why
《为什么》
- Published
- 2018
- Category
- Artificial Intelligence
- Original language
- en
- Source list
- Daily Pick
Returning causal reasoning to machines: the three-rung ladder from association to intervention to counterfactual
Imported from a third-party reading list or added as a daily pick — bibliographic facts, the source framing, and a full reading guide.
Source theme
Returning causal reasoning to machines: the three-rung ladder from association to intervention to counterfactual
What is this book about?
This is the Turing laureate Judea Pearl's systematic statement of the science of causation, and his popular account of half a century of his own work. Pearl's central claim is that statistics and machine learning have long substituted correlation for causation, and that this is not enough: a machine can predict but cannot answer what would happen if we acted. He sets out a three-rung ladder of causation, from association (seeing) through intervention (doing) to counterfactuals (imagining), and argues that only the upper rungs support anything like human causal understanding. Through cases such as smoking and lung cancer and Simpson's paradox, the book explains Bayesian networks, the do-operator, the back-door criterion, mediation analysis and instrumental variables, turning a difficult mathematical language into a climbable ladder of inference. Pearl received the 2011 ACM A. M. Turing Award for his foundational contributions to probabilistic and causal reasoning.
Why read it?
In a period that repeatedly mythologises big data and deep learning, this book supplies the most substantial counter-argument available: pattern matching over data alone cannot generate causal knowledge. Pearl is not an outsider critic but the person who laid the foundations of the field, since Bayesian networks and structural causal models are his work, and his objection therefore carries technical weight. For those working in AI, in data analysis or in investment decisions it offers criteria rather than slogans: which questions require experiments, which can recover causation from observational data, and which data can never answer. The Chinese edition was published by CITIC Press in 2019.
Core Ideas
- Correlation is not causation, and the gap between them can be characterised precisely. The ladder of causation has three rungs, association, intervention and counterfactual, and most machine learning sits on the lowest one, which is why it can predict but neither explain nor reliably guide action.
- Causal assumptions cannot be conjured from data. Establishing causation requires prior assumptions about how the world works, expressed as a causal diagram, and data can only test or estimate parameters within those assumptions. This is the deepest limit of data worship.
- The do-operator makes intervention computable. Through tools such as the back-door and front-door criteria, one can determine without running an experiment which observational data suffice to estimate the effect of an intervention. This is the technical core of the causal revolution.
- Counterfactuals are the distinctly human rung. The third level requires imagination and a model, not merely curve fitting, and Pearl treats the ability to handle counterfactuals as the line separating a predicting tool from an understanding mind.
What questions does this book try to answer?
- Why can causation never be read directly out of data without external assumptions, and what does that limit imply for artificial intelligence?
- What capacities must a machine acquire in order to move from prediction to understanding, and to answer questions of intervention and counterfactual reliably?
Who should read it?
Practitioners of machine learning, data science and causal inference; product, operations and policy analysts who must judge from observational data whether an intervention actually worked; and researchers or investors concerned with the limits and interpretability of AI. No advanced mathematics is assumed, but the reader must be willing to follow an argument rather than collect conclusions.
How to Read It
The Chinese edition was published by CITIC Press in July 2019. Read the chapters in order rather than skipping, because Pearl builds the ladder rung by rung: the opening chapters set the framework of the ladder and Simpson's paradox, the chapters on intervention and the do-operator mark the turning point, and the material on counterfactuals and mediation becomes markedly harder and is precisely where the book's value lies. Readers wanting the mathematics can continue with his Causality; those wanting engineering practice can pair it with the relevant chapters of Artificial Intelligence: A Modern Approach, already in this library.
Reading Notes
2026-10-19
Data can tell you only what occurred together, never what would change if something were changed; the second and third rungs of the ladder — intervention and counterfactual — can never be read off observation, only climbed with an explicit model.
The key move of this book is to split causal questions into a ladder of three rungs: at the bottom, association — seeing two things occur together; in the middle, intervention — what happens to one thing when another is deliberately changed; at the top, counterfactual — what would have happened had something not been done. What it genuinely changes is the standing of the phrase "let the data speak": Pearl argues repeatedly that statistical learning, however powerful, stays at the bottom rung, because correlation carries no information about what would happen under a change, and no sample size will grow that layer on its own. The easily missed point is that climbing to the middle and top rungs requires not more data but something else entirely — an explicit model of how the world works; the same observations paired with different causal graphs yield different conclusions about intervention, which is exactly what makes Simpson's paradox unsettling. The caution worth keeping is that the framework is often read as "draw the arrows and you have causation," whereas Pearl's own condition is the reverse: the model must be testable and refutable, or it is merely prejudice translated into a diagram. It forms two halves with this site's "The Evidence That Isn't There": that piece examines how the sample is quietly filtered by the outcome, and this book explains that even a complete sample leaves observation itself at the foot of the ladder.