Judea Pearl, 2018

“Data are profoundly dumb.”

Computer scientist Judea Pearl and science writer Dana Mackenzie used this sentence in their 2018 book “The Book of Why.” Its bluntness serves a precise purpose. A dataset records observations, but it does not announce the causal structure that produced them. Statistical association can reveal that two variables move together and can support useful prediction. It cannot, by itself, tell whether changing one variable will change the other, whether both share a hidden cause, or whether the relationship will survive an intervention. Those are questions about the world, not properties that raw numbers explain unaided.

The distinction matters throughout machine learning. A model trained on past hiring may predict who resembles previous employees, but that does not show which qualifications cause success or whether old decisions were fair. A health model may learn that spending is associated with need, even when unequal access makes spending a distorted proxy for illness. Prediction can be accurate according to its target and still mislead the decision-maker. Choosing a target is therefore a scientific and moral act. It requires a theory about what the data represent and how the output will be used.

Pearl’s work on causal inference formalizes some of these questions with diagrams, assumptions, interventions, and counterfactuals. The mathematics does not make causality automatic. Researchers must state which relationships they assume, identify what can be learned from the available evidence, and test the implications where possible. That explicit structure is valuable because it exposes disagreements that a large predictive model may bury. It also clarifies when an experiment, natural experiment, domain expertise, or additional measurement is needed before recommending a policy or treatment.

In the era of vast datasets and generative models, the quote is a defense of explanation. More data can improve coverage and reduce some forms of uncertainty, but quantity cannot answer a question the data were never designed to identify. People naturally ask causal questions: What made this happen? What would have happened under another choice? Which intervention will help? AI that responds with correlations may sound authoritative while missing the request. Responsible use begins by matching the method to the question and by disclosing when a prediction cannot justify the action someone wants to take. Pearl’s provocation is not an argument against collecting evidence. It is an argument for reasoning with evidence. The strongest AI systems will know when a pattern supports prediction, when a decision requires a causal model, and when uncertainty should prevent a confident recommendation.