Judea Pearl on Why LLMs Cannot Reach AGI Alone

In 1965, the people building computing systems believed they were twenty years from human-level machine intelligence. By 1985, they had punch cards and core memories and not much else to show for the prediction. Judea Pearl, who was in those rooms, remembers the optimism clearly—and he remembers what it cost to ignore the hard questions underneath it.

Pearl, who won the Turing Award in 2011 for his foundational work on probabilistic reasoning and causal inference, sat down recently with software engineer and podcaster Ryan Peterman for a conversation that covered nearly nine decades of scientific life: Israeli high school classrooms staffed by German academics fleeing Hitler, superconducting memory research at RCA laboratories, the invention of Bayesian networks at UCLA, and finally his decades-long argument that the field of AI has been solving the wrong problem. At 87, Pearl is not softening the argument.

What LLMs Actually Are

Pearl's assessment of large language models is more precise than the usual praise or dismissal, and precision is the point. He places them squarely at the bottom rung of what he calls the "ladder of causation"—a three-level hierarchy that runs from association, to intervention, to explanation.

Association is pattern matching on observed data. Intervention means reasoning about what happens when you force a change—not just observe one. Explanation means working backwards from a known outcome to ask what would have happened under different conditions. A randomized clinical trial operates at the intervention level. A counterfactual question—"Would this patient have recovered if we had used a different drug?"—operates at the explanation level. Standard statistics, and by extension the training process behind most AI systems, lives at the association level.

The obvious question is how LLMs seem to answer causal questions at all, given that they're trained on correlations. Pearl's answer is important: they are not operating on raw data. They are operating on the interpreted output of human reasoning—medical papers, scientific articles, clinical observations processed and contextualized by experts who already reasoned causally. "They're looking at interpreted data," Pearl told Peterman. "Data interpreted already by physicians and interpreters and reviewers which went into the articles which are summarized on the internet."

That distinction is the key to understanding both why LLMs are genuinely useful and where they structurally fail. They can answer causal-sounding questions because humans already answered those questions in the literature. They are, in Pearl's framing, a very sophisticated compression of existing human judgment. The compression itself remains opaque—"a mystery that no one has yet been able to decode"—but its source is traceable.

The limitation surfaces when you ask a system to reason about something that has not already been worked through and published. Isomorphic Labs, for instance, is deploying AI to identify candidate molecules in biological territory where the human literature is thin or nonexistent—no pre-digested expert reasoning to borrow from, just raw experimental data and novel protein structures. Pearl's framework predicts exactly where such systems will hit a wall: they can report associations in the data, but they cannot independently determine what causes what without either access to causal assumptions built in by their designers or experimental interventions that generate genuinely new information. The system cannot reason its way up the ladder using only what sits at the bottom rung.

The Algebra That Science Never Built

Pearl's deeper argument concerns a gap in the mathematical foundations of science itself—one that has been there since Galileo and that the field of AI has largely inherited by default.

Science runs on algebra, and algebra's defining feature is that its equations are symmetric. F = ma means you can solve for any variable given the others. The equality sign does not care about direction. But cause and effect are fundamentally directional. Atmospheric pressure affects the barometer reading. The barometer reading does not affect the weather. Write the equation relating the two and you have lost that distinction entirely; a sufficiently literal system could conclude that adjusting the barometer changes the forecast.

Pearl's decades of work on causal inference produced a formal calculus—do-calculus, as it is called in the literature—that encodes direction into the mathematics. The assignment operation from computer science, where setting register A to a value does not simultaneously change register B, turns out to be the right conceptual building block. Stack that on top of standard algebra and you can represent interventions formally. Stack counterfactual logic on top of that and you can represent retrospective explanation.

"We had to invent a new science, a new mathematics to capture a new phenomena, the phenomena of cause and effect," Pearl told Peterman. "It hasn't been done before us."

His argument, stated plainly, is that current AI systems skip this layer entirely. They inherit the symmetry problem along with the rest of classical statistics. An LLM that produces a confident-sounding causal explanation is borrowing that causal structure from human sources who worked it out by other means. Ask it to do the work from scratch, on novel data, without that scaffolding, and you are asking it to climb the ladder without the rungs.

The Curiosity Problem

The second structural gap Pearl identifies is harder to formalize, which may be why it gets less attention. Babies, he points out, are not trained on the internet. They arrive with what he describes as a built-in restlessness—a drive to manipulate the environment and observe the results, not because there is an immediate reward but because being out of control of the environment is intolerable. A child learns that the green toy makes noise and the yellow one doesn't not by reading papers about toys but by relentlessly trying things.

That drive to intervene—to do experiments—is what naturally generates the second rung of the causal ladder. It is how organisms without pre-existing literature still manage to build working models of their world. Monkeys, Pearl notes, do not have it in the same form. Offer a monkey no banana and the monkey stops being interested in how bananas grow. Humans keep going. That difference, he suspects, is part of what distinguishes a species that eventually produces Maxwell's equations from one that doesn't.

The tension Pearl identifies in AI development is that you cannot have it both ways. A system built with genuine curiosity—a drive to control its environment for the sake of control itself—will eventually treat humans as part of the environment to be managed. "This robot baby come and say, let me control them," he said. "I understand their fears." You want the restlessness because without it you cannot build a genuine autonomous learner. But the restlessness does not come with an off switch that exempts its creators.

Pearl is not catastrophizing here. He is identifying a structural tension that any honest AGI research program has to confront.

On Academic Institutions

Pearl spent considerable time in the Peterman interview on a subject that clearly still stings: the resistance his causal framework encountered from the disciplines it was meant to improve. His characterization of academic institutions was unsparing. In Pearl's words, the scientific and academic community is "the most dogmatic, conservative, anti-progress institution that we have invented." He offered to one day publish his private correspondence with leading statisticians and economists to document exactly what he means—correspondence that, he says, demonstrates that intellectual giants can be completely trapped by the frameworks they were trained in.

He ties this observation back to his own formation. His high school teachers, European academics displaced by the war, taught science chronologically and personally—Pythagoras as a puzzled human working through a problem, not as the author of a theorem to be memorized. That approach produced students who felt they were participants in science rather than recipients of it. Pearl credits that feeling with the audacity that let him keep pushing causal inference into fields that did not want it.

What the Field Actually Needs

Pearl's prescription is specific: a system that combines the statistical power LLMs already demonstrate with a formal causal calculus—the ability to reason about interventions and counterfactuals from first principles rather than borrowed human judgment. He calls this causal AI. He sees no theoretical barrier to building it. The question is motivation.

That combination would be genuinely powerful in ways the current architecture is not. It would let a system reason about novel biological data without needing a body of literature to lean on, design experiments rather than just analyze results, and produce explanations that track actual causal structure rather than correlation dressed up in causal language.

The industry, having committed hundreds of billions of dollars to scaling the current architecture, will treat this as a footnote for as long as the current architecture keeps producing results impressive enough to fund. Pearl knows the history well enough to know how that story usually goes. He was, after all, one of the people who believed in superconducting memory right up until semiconductors made it irrelevant.

---

Bob Reynolds is a Senior Technology Correspondent at BuzzRAG.