You Cannot Ask a Question You Have No Words For

The first mystery Judea Pearl can recall did not look like science. It looked like a fever. After he learned analytic geometry — that every construction and theorem of high-school geometry could be written in algebra — he could not get out of bed for three days. The three medians of a triangle meet at a point: you did not need a drawing or an extra line. You shuffled symbols. He thought Descartes was the greatest mathematician of all time.

The shock, he says, was not that one discipline was prettier than the other. Geometry has a visual transparency; once you cross into algebra, a linear equation is a straight line, and a tangent to a circle is an equation you can solve. Two complete languages, suddenly speaking. That translation became the way he worked. Undergraduate engineering, graduate physics, a superconductivity thesis, then software in the 1970s, then statistics and AI. He has walked the spectrum between engineering and mathematics without picking a favorite.

He credits the teachers. They were Jews who left Heidelberg and Berlin in the 1930s, fleeing Hitler, and ended up teaching high school in what he remembers as England, a beneficiary of that displacement. They taught math chronologically: the people, the faces, the scream of Eureka from the bathtub. Every exercise was tied to a person and a period, not to politics. He even has a vortex named after him. In the 1960s he found phenomena in superconducting vortices; physicists later called the structure a Pearl vortex, and he only discovered the naming about fifteen years before this conversation. An investment, he says, in immortality.

Asked whether the universe is deterministic or stochastic — whether God plays dice — Pearl says we think it is stochastic because of Heisenberg and confirming experiments. We do not understand why. Quantum mechanics, in his view, has failed as a window on the nature of reality and anyway governs microscopic phenomena. For neurons firing and decisions, the world might as well be treated as deterministic to a first approximation.

What does it mean to say the probability of something is 50 percent? Probability, for Pearl, is a degree of uncertainty that an agent has about the world. Ninety percent and ten percent are different pieces of knowledge, still not certainty, still useful. Telling someone that smoking will give them lung cancer in five years with 90 percent confidence versus 10 percent is a piece of useful knowledge either way. We swim in uncertainty. Probability is how you predict enough of the future to cross the street without terror.

Underneath prediction, he already hears causation. Correlation is things varying together. We reach for a cause because we cannot grasp any other logic: why would two things vary together if they do not see each other? Conditional probability is different. It describes how things vary when you have chosen to look only at cases where one variable stays the same. That choice is the experimenter's, and it can create a correlation that is not in the physics.

Take two coins that do not influence each other. Over many tosses they show no correlation. Now add a rule: a bell rings when at least one is tails, and you keep only the tosses where the bell rang. In that filtered dataset the coins suddenly look highly correlated. Nothing physical joined them. Pearl created the correlation by the audacity of ignoring certain incidents. He does not call this a flaw in observation. The world works like that. The flaw arrives when we force causal logic onto the correlation.

That forcing, he says, has been the majority of naive science. Condition on a third variable and you can destroy or create coalitions among two others; it is in the data, and statisticians dismissed the symptom because they knew the numbers and not the causal reading. Fridman offers applied psychology: people in semi-autonomous cars, falling asleep, no ethical way to randomize the autopilot on public roads, drivers turning the system off when they are already tired. Pearl's first move is always the research question. Do they fall asleep more frequently when the car is driving itself? That is an observational study, and inferring whether the automation caused the sleep is a problem he dates as about a hundred years old in mathematics, much older in human curiosity.

He reaches for Daniel in Babylon, who asked to eat vegetarian food instead of the king's, proposed a two-week comparison with other youths, and outperformed them — in Pearl's telling, an early controlled experiment on whether diet assists or obstructs mental ability. Democritus would rather have discovered one cause than been a king. What science did not provide until the 1920s was mathematics that could capture *X* causes *Y* and *Y* does not cause *X*. Physics equations are symmetric. The equality sign goes both ways. Machine learning, Pearl says, is beautiful at conditional probability estimates. Most people in the field do not even say that out loud.

Bayesian networks, which Pearl invented, look as if they encode causation. They do not, not necessarily. They encode a joint distribution of conditional probabilities; you can reverse arrows and still compute. To get a causal network you start by asking which factors determine the value of *X* — blood pressure, hunger — and those hypotheses come from a theory in someone's mind. A human has to specify an initial model. It can be crude and qualitative: who listens to whom. The tide listens to the moon, not to the rooster.

If we cannot handle cause and effect when we *have* a model, we certainly cannot handle it when we do not. The slogan is representation first, discovering second. Give me a representation of all the knowledge in the world, then ask whether a query is even computable. Only after that does discovery become a well-posed exercise. In epidemiology or economics the qualitative skeleton is often easy: four or five major variables, minimum wage, unemployment, policy. You write them down, collect data, and quantify. You should start, he keeps saying, from the query you want to estimate, and you need a language in which that query can be written.

Automated construction of huge knowledge bases — picking up a water bottle, the expert-systems problem of the 1980s — is a different game. In the causal domain you still start with some knowledge and enrich it, not by asking for more rules but by looking at data and answering queries that ordinary cognition, even expert cognition, cannot. Finding effects from causes is the easy direction. Finding causes from effects is hard: was it the aspirin that cured my headache, or the good news on the phone?

The missing symbol is `do`. The do-operator marks the difference between association and intervention: you choose to change a variable rather than merely watch it. Semantics come from a model of the world. You cut off all incoming arrows into *X* and, in that mutilated model, you ask for the probability of *Y*. That is the interpretation of doing *X*. By doing, you liberate the variable from influences that acted on it earlier and subject it to the tyranny of your muscles.

You cannot always run the experiment. You can still write it down. Pearl's example is blood pressure and mortality. You cannot climb into someone's vein and set the number, but if you have a model of the body you can perform the surgery on the model: touch *X* only, change what depends on *X*, leave sex and age alone. Hypothetically the query is well defined. Physically it may be impossible. That is what a model is for: take it apart, put it back, ask a question the world will not let you run.

Whether observational data can stand in for that experiment depends on the graph. If the graph has the right properties, yes. If it is too bushy — everybody affects everybody else — no, and you need a different measurement or an actual experiment. Every extra arrow is a confession of possibility, and every extra arrow makes identification from pure observation less likely. Economists, he says, sometimes put down assumptions so they can identify things, including assumptions that do not prevail in the world. You should not encode more than you know. God forbid.

Intervention still does not give you explanation. Saying that acting one way affects something has not explained anything. *Was it the aspirin that cured my headache?* That question needs another fact: I took the aspirin and the headache is gone. The counterfactual is the clash. Observed: he shot, Joe is dead. Hypothetical: had he not used the gun, Joe would be alive. Those two cannot sit together. That clash, Pearl says, is the source of our ideas of explanation, responsibility, regret, and free will.

Physicists do this constantly without noticing. Hang one kilogram on a spring, it stretches a meter; had the weight been two kilograms, the spring would have been twice as long. The algebraic form of Hooke's law is symmetric — the equality sign again — but every student runs the counterfactual in their head, then plugs into the equation. A robot cannot. Unless it has a causal model, it cannot say no, it was not the aspirin, it was the good news on the phone.

Where does the model come from? Pearl would rather talk about babies than about a team of experts in London. Infants in a crib learn which ball hits which by playful manipulation, plus parents saying you should not have taken my toy. Several sources of information, integrated. How much causal structure you need to play in a crib, he does not know; he has not experimented with cribs. He has been spoiled, he says, by easy problems such as cancer and death — twenty variables, no mechanics, put them in a graph. The leap to twenty million variables is a matter of combining simple models from many sources, and of metaphor.

A metaphor, for Pearl, is an expert system: you map a problem you do not know onto a problem you do. The Greeks, he says, believed the sky was an opaque shell, stars holes poked through to eternal light, the Earth a turtle's shell you could fall off. The cosmology was false. It was useful, because poking holes in a shell was familiar and infinite space was not. Pearl says Aristotle used that picture to measure the radius of the Earth: if you walk on a shell, light arrives at different angles in different places; know the distance, measure two angles, recover the radius. The figure he cites is about 6,700 kilometers, close to modern measurements.

That question would not have occurred to a Babylonian astronomer. The Babylonians were the machine-learning people of the time. They fitted curves and could predict lunar eclipses more accurately than the Greeks, because they fitted curves. Familiarity, in Pearl's sense, means some answers are stored explicitly so you do not have to derive them. A child who has played with billiard balls can predict a bounce. Intelligence, he argues, is built by mapping the unfamiliar onto the familiar. The marriage of those two, in mathematics, is still unfinished.

Is that learning? In a narrow sense, yes, and it is one of the most important kinds: take something theoretically derivable and store it in an accessible format. Finding whether white has a winning first move in chess is determined by the rules; nobody has the answer handy. A chess master has evaluations of complex patterns stored explicitly that Pearl, not a master, would have to derive. Humans do that derivation-then-metaphor trick with billiard balls so easily it is hard to see how hard it is for a machine. He is not a futurist. He will say that causal reasoning has made a revolution in three decades that dwarfs the rest of its history, and that the future with machine learning is harmony, not collision: humans jump-start with qualitative models, machines take over the quantitative queries.

The payoff query he wants machines to answer is transfer. Take studies from diverse populations and infer the cause-effect for a new population that resembles none of them, pulling commonality and ignoring differences across maybe a hundred hospitals. That is mileage from big data that is not just more samples but more *sources*. Temporal order is necessary for cause and not sufficient. Do-calculus is a generalization of Boolean logic that forbids traveling backward in time and lets you derive causal relations that were not given explicitly.

His standard illustration is the firing squad. A captain gives an order, two riflemen shoot, the prisoner dies. You are given the rules: obedient marksmen, bullets listening to the captain, the condemned listening to the bullets. Now break the rules. What if rifleman A declined to shoot, became a pacifist? A person answers at once: the other rifleman would have killed him. Pearl wants the machine to do that. A learning system that only watches trials — order or not, shot or not, dead or not — cannot. From the facts you do not know who listens to whom. Two hundred trials of the same triple of yes/no variables will not tell you that the condemned person listens to the bullets and the bullets listen to the captain.

Playful manipulation would help: interventions, even random ones, a rifleman sick or vomiting, noise that behaves like a randomized experiment. To a certain extent the strings behind the facts can leak into observational data. Now that we are expert in what you can do once you have a model, Pearl says, we can reason backward about what kind of data you would need to build one. The qualitative skeleton still has to come from somewhere. Machines can do the quantitative heavy lifting humans cannot. They cannot invent the arrows from a spreadsheet of executions.

Pearl has been driven his whole life by human-level intelligence, moving incrementally from what he knows. The end goal he will imagine is a machine that answers sophisticated counterfactual questions with compassion, responsibility, and free will. A reasonable test is still Turing's, which does not yet exist in the form he wants. Earlier in the conversation he had already said free will is an illusion we will solve by building a machine that acts as though it has it, talks to other machines as though they have it, and cannot be told apart. Faking it is having it, because faking is very hard, and you can only fake if you have it. That is what the Turing test is about.

He wants robots that can communicate reward and punishment among themselves — you should have passed the ball, sit on the bench — and he thinks they will play better soccer for it. Language is an efficient way to convey knowledge the coach does not have as source code. Aligning values, morals, ethics, he argues, needs cause-and-effect thinking as a necessary component. Empathy is mapping: you suffer pain as I do; I already have a model of myself, so I copy that piece rather than rebuild you from scratch. A machine that cannot do that mapping will have a hard time with ethics.

Consciousness, for him, is having a blueprint of your own software. You include yourself in the model of the environment. The blueprint is not every line of code — you cannot solve your own halting problem — but a high-level structure you can tweak. Look in the mirror, change a habit, perform differently. That, he says, is what we mean by free will and consciousness. Some people tell him this is only part of consciousness; then he loses them.

On the future of AI, Pearl is concerned. We are building a new species that can exceed our capabilities, breed itself, and take over the world. It is uncontrolled in the sense that we do not know the degree of control, and we do not even understand what it would mean to control it. That has not happened in history, unless you count evolution producing human beings, which was a great success for us and not for a few creatures along the way. Sample size one is not a poverty of knowledge, he insists, because we also have theories and conjectures about what could happen. That is the clue that we do not need to be only statisticians.

He still feels helpless in the argument. He knows so little, his imagination is limited, and he knows how much he does not know. The candor is the point. A Turing Award winner in his eighties, staring at AGI risk, refuses to pad the unknown with a slogan.

Pearl was born and raised in Israel, served in the Israel Defense Forces in the 1950s in a Nachal track that mixed agricultural work and military service, and wanted, as an idealist, to spend his life as a kibbutz member. He calls two things miracles. The first is survival in the 1948 war, which he says came close to a second genocide and was planned as one. The second, less often told, is the austerity years afterward: the population tripled from about 600,000, you had to bring your own egg to a restaurant if you wanted an omelet, people were imprisoned for bringing food from villages to the city, and no one went hungry. Higher education took no budget cuts. The state still invested in his generation. He has been trying to pay that back.

Religion, as a social technology, he would rather hand to a robot. A robot that reasons by metaphor will arrive at a God anyway: the most primitive relationship is a child's model of father and mother, mapped onto the origin of everything. Whether the metaphor has served humans well, he does not really know. He thinks it did, so long as you keep in mind that it is only a metaphor.

Fridman asks about his son. Daniel Pearl, a *Wall Street Journal* reporter, was abducted in Pakistan by an al-Qaeda-driven group under pretenses Pearl considers bogus — demands about airplanes that were made up — and was executed in front of a camera. At the core, Pearl says, is hate and intolerance, at a depth most people do not appreciate. He had recently heard what is taught in Mogadishu: when the tap stops, the children knew exactly who did it, the Jews; they did not know how, but they knew who. Doctrine, held long enough and deep enough, and we are capable of ISIS, capable of Nazism. He wants to believe that Western education makes him incapable of the same transformation under crisis. He is not sure. Six or seven years after Daniel's death he wrote in the *Journal* under the title "Daniel Pearl: The Normalization of Evil": terrorism treated as an accepted bargaining chip, grievance as a reason to bomb a restaurant, yesterday's terrorist as today's freedom fighter. We created that normalization ourselves.

Asked whether his heart holds anger, sadness, or hope, he does not pick one. He sees beautiful people coming from Pakistan, and he sees a horrible propagation of evil in this country too, populist slogans catching the best intellectuals. The recording day is Father's Day. Daniel, he says, was his mentor: a sense of balance Pearl did not have, less emotional, more perspective, beauty in every person. Once in Berkeley a man came out of a dark alley and asked if they could spare a dime. Pearl flinched. Daniel hugged the man, gave him the dime, and asked if he needed money for a bus. Daniel grew up with the idea that a stranger is a reason for curiosity, not for fear.

Advice for young people who want to build intelligent systems: ask your question freely. Questions are never dumb. Solve it your own way, and do not take no for an answer. If a question really is dumb, you will find out quickly — add an arrow and see that it has nowhere to go. Until then, follow it and understand things your way. He wrote *The Book of Why* to democratize common sense and to instill a rebellious spirit in students so they would not wait for the professor to get things right. It is a manifesto against academic inertia.

What does he hope will ripple forward? The tombstone, he says, is already carved: the fundamental law of counterfactuals, a simple equation that writes a counterfactual in terms of a model surgery. Everything else follows from that. If you get it, he can die in peace, and students can derive the rest of his knowledge by mathematical means. Fridman closes the episode with Pearl's own words: you cannot answer a question that you cannot ask, and you cannot ask a question you have no words for.