Judea Pearl

In the [YOU] on AI Field Guide

The cycle that began with [YOU] on AI asks what it would mean to take the orange pill—to see the machine clearly, without the narcotic of hype or the paralysis of fear. Pearl is the clearest seer the cycle could summon of the gap between what these systems do and what intelligence requires. To read him is to be handed a measuring instrument: a ladder, three rungs high, and a claim about where our machines stand on it. The claim is not flattering, and it is not meant to be—it is meant to be true, and Pearl is unusual among critics of AI in that he can prove much of what he says rather than merely assert it.

His lens reframes every question the cycle asks. The issue is never how impressive a model's performance is, but which rung it is performing on. A system can become superhumanly good at detecting patterns and remain, in the strict sense Pearl gives the word, unintelligent—unable to reason about its own actions, unable to imagine alternatives, unable to ask why. He is not a pessimist; he believes machines may one day climb the ladder, reason about interventions, perhaps possess something like free will. But they will get there, he insists, only if we stop confusing the foothills for the peak.

The diagnosis explains a pattern that has puzzled many observers of contemporary AI: the coexistence of superhuman fluency and absurd, basic error. The same system that drafts a competent legal brief will invent a case that does not exist and cite it with perfect confidence. To a first-rung machine these are the identical act—the generation of text that fits the patterns of its training data—because telling a real case from a fabricated one would require a model of what is real, and a curve-fitter has no model of what is real. It has only a model of what is typical. The same decorrelation of fluency from authority that the cycle treats as the signature hazard of the age, Pearl explains from first principles.

He thus stands in the cycle's gallery of thinkers as the one who supplies the rigorous instrument the others reach for in their own idioms. Where Gary Marcus argues from cognitive science that pattern-matching cannot deliver reliable generalization, and Stuart Russell argues from control theory that we have been building intelligence the wrong way, Pearl argues from the logic of information itself—and even Geoffrey Hinton, who built the curve-fitting machines and disputes Pearl's verdict, is most precisely opposed when measured against Pearl's ladder.

Gary Marcus

Origin

Born in Tel Aviv in 1936 and trained as an electrical engineer at the Technion before earning his doctorate in the United States, Pearl built his reputation by asking a question his entire discipline had agreed to forbid. Karl Pearson, who founded the modern statistical apparatus, regarded causation as a relic of metaphysics—a word real scientists should expunge. The honest researcher reported associations and left the rest to philosophers. This refusal was a genuine achievement; it protected generations from the oldest fallacy in reasoning. But it came at a cost the field preferred not to count: by banishing causation, statistics had also banished the very questions science exists to answer. Does the drug cure the disease? Did the cigarette cause the cancer? A science that cannot ask them, Pearl observed, is not modest. It is mute.

Stuart Russell

His intervention was to notice that the taboo rested on a confusion of its own. The problem was never that causation was unscientific; it was that no one had given it a mathematics. Probability theory could express that two events occur together, but had no symbol for the claim that one produces the other—no way to write that the coming storm makes the barometer fall and not the reverse. So Pearl set out to invent the notation. The work occupied decades and produced two revolutions: in the 1980s, Bayesian networks, which let a machine propagate probabilities through a web of dependencies the way water finds its level; and later, the calculus of intervention and counterfactuals for which he won the 2011 Turing Award, his field's highest honor.

Pearl came to regard the Bayesian-network achievement, which alone would have secured his place in the field's history, as a way station. The networks were brilliant at the first task—tracking what goes with what—and that very brilliance exposed the deeper lack. They could tell you that wet grass and rain go together. They could not tell you that the rain wets the grass and the grass does not summon the rain. To capture that, you needed a different kind of object: a model not of probabilities but of mechanisms. The pursuit of that object defines the rest of his career—and the rest of his quarrel with the present moment in AI.

Geoffrey Hinton

Key Ideas

The Ladder of Causation. Pearl organizes all of intelligence around a three-rung hierarchy—association (seeing), intervention (doing), and counterfactual (imagining). The rungs are nested, and the gaps between them are differences of kind, not degree: you cannot climb from one to the next by collecting more data, only by adding a new kind of knowledge. The whole of the ladder is his answer to the question of what intelligence requires.

The do-operator. His central technical contribution is a notation that distinguishes intervening on a variable from merely observing it—the difference between what did happen when prices rose and what would happen if you set the price by decree. No amount of observational, rung-one data, he proves, can in general answer an interventional, rung-two question. The do-calculus is the apparatus that marks exactly where that boundary lies.

You are smarter than your data. Causal questions cannot be answered from data alone; the missing ingredient is a causal model, an account—supplied by the modeler—of how the world that generated the data actually works. Data do not understand causes and effects; humans do, and they bring that understanding to the data rather than draw it out. This conviction runs directly against the model-free instinct of modern deep learning, which treats prior knowledge as contamination to be eliminated.

Counterfactuals and the human difference. The third rung—what would have happened had I acted otherwise—is, by Pearl's analysis, the cognitive engine behind science, law, and morality. Blame, regret, and responsibility all depend on imagining a world that did not occur. Counterfactual reasoning is the rung furthest beyond any machine we have built, and the one the cycle ties most closely to consciousness and the sense of self.

The curve-fitting critique. Pearl treats "curve fitting" not as an insult but as a classification. Modern machine learning, however dazzling, lives entirely on the first rung; scaling makes a system better at that rung and nothing else. The wall it approaches is built not of insufficient data or compute but of the difference between seeing and doing—and it is why a curve-fitter fails, confidently and without insight, the moment the world departs from its data.