Can a Language Model Understand Cause and Effect?

A field guide to LLMs and causality — the trap, the theory that explains the trap, and the practical, honest way out of it.

There is a particular kind of magic trick that large language models perform so smoothly that we rarely notice it is a trick at all. You ask a modern model whether smoking causes cancer, whether raising interest rates cools inflation, or whether a certain drug lowers the risk of stroke, and it answers with the calm authority of a seasoned expert. The sentences are fluent, the reasoning sounds sound, and the conclusion is very often correct. It feels, unmistakably, like the machine understands cause and effect.

It does not. And the gap between what these models appear to do and what they actually do turns out to be one of the most important and misunderstood topics in modern artificial intelligence. Getting it right matters enormously, because the moment we start using these systems to make consequential decisions — in medicine, in economics, in public policy — the difference between a model that knows about causation and a model that merely recites causal-sounding text becomes the difference between a tool and a liability.

This article is a tour through that gap. We will build up, from first principles and with plenty of everyday examples, an understanding of what causality actually is, why it is fundamentally harder than the pattern-matching that language models excel at, and why no amount of scale can close that gap on its own. Then we will do something more hopeful. We will show that although a language model can never be a causal reasoner, it can become an extraordinarily valuable component of a causal pipeline — provided we understand exactly what job to give it and, just as importantly, what job to keep firmly in human hands.

1The oldest confusion in science

Let us begin with the confusion that has plagued human reasoning for as long as humans have reasoned: the difference between things that happen together and things that make each other happen.

In the summer, ice-cream sales rise. In the same summer, the number of people who drown also rises. If you plotted these two numbers over the course of a year, they would track each other beautifully. A naive analyst, or a naive algorithm, might conclude that ice cream is dangerous — that eating a cone somehow leads people to their watery doom. Of course, the real story is that hot weather independently drives both: heat sends people to the ice-cream stand and, separately, into lakes and swimming pools where some of them, tragically, drown. The weather is a hidden common cause, and it manufactures a correlation between two things that have no direct relationship whatsoever.

Statisticians call that hidden common cause a confounder, and it is the villain of nearly every story about causation. The phrase “correlation does not imply causation” is really a compressed warning about confounders. Whenever you observe that two things move together, there are always at least three possibilities lurking behind the pattern. Maybe the first thing causes the second. Maybe the second causes the first. Or maybe some third thing you haven’t even measured is quietly causing both. The data alone — the raw record of what happened — usually cannot tell you which of these three worlds you are living in.

Here is a more mischievous example, and a real one. A famous analysis once showed that a country’s chocolate consumption per capita is strongly correlated with the number of Nobel laureates it produces. Switzerland, a nation of chocolate lovers, tops both charts. Should we conclude that feeding a population more chocolate will win them Nobel Prizes? Obviously not. The far more sensible explanation is that wealthy, highly educated countries tend to both consume more luxury goods like chocolate and invest heavily in the universities and research institutions that produce laureates. National wealth and educational infrastructure is the confounder. The chocolate is a bystander that merely happens to rise and fall alongside the thing we actually care about.

2Building the language of causes

To reason carefully about cause and effect, scientists use a tool called a Structural Causal Model, usually abbreviated SCM. The name sounds intimidating, but the idea is one you already hold intuitively. An SCM is simply a precise way of writing down what causes what, and by how much.

It has two parts. The first is a picture: a diagram of arrows, where each arrow points from a cause to its effect. If altitude affects temperature, we draw an arrow from altitude to temperature. This diagram is called a directed acyclic graph, or DAG — “directed” because the arrows have a direction, and “acyclic” because you are not allowed to have causal loops (a thing cannot ultimately cause itself). The graph answers the qualitative question: which variables matter, and which way does the influence flow?

The second part is the set of equations that put numbers on those arrows. Knowing that altitude affects temperature is one thing; knowing that temperature drops by roughly six and a half degrees Celsius for every kilometer you climb is another. The graph tells you what affects what; the equations tell you how much.

Why bother with all this machinery? Because it forces you to make your causal beliefs explicit, and once they are explicit, you can reason about them rigorously. An SCM is not just a description of the world as it is. It is a description of what would happen if you reached in and changed something — if you intervened. And that ability to reason about interventions, about hypothetical actions you have not yet taken, is precisely what separates genuine causal understanding from mere pattern recognition. This brings us to the single most important idea in the entire field: Judea Pearl’s ladder of causation.

3The three rungs of the ladder

Pearl, who won computing’s highest honor in large part for this work, observed that causal questions are not all created equal. They come in three levels, arranged like the rungs of a ladder, and each rung is strictly harder to reach than the one below it.

The bottom rung is observation. This is the world of seeing. Here we ask questions like: “What is the recovery rate among patients who happen to take this drug?” or “How does blood pressure typically vary with age?” These are questions about associations, about what tends to occur alongside what. This is the natural home of statistics, of machine learning, and — as we will see — of language models. It is the rung of what is.

The middle rung is intervention. This is the world of doing. Now we ask: “If I give this drug to a patient, will their recovery improve?” Notice how different this question is from the one below it. The observational question asked about patients who chose to take the drug, or were selected to take it by their doctors. The interventional question asks what happens when we decide, when we reach in and act, potentially overriding whatever process normally determines who gets treated. This is the rung of what if I act — and it is the rung on which almost all important decisions actually live.

The top rung is counterfactual. This is the world of imagination, of roads not taken. Here we ask the most delicate questions of all: “This particular patient took the drug and recovered. Would they have recovered anyway, even without it?” We are now reasoning about a hypothetical alternate history for a specific individual, a history we can never actually observe because it did not happen. This is the rung of what would have been.

Now here is the crucial, non-obvious fact, the fact that everything hinges on: you cannot climb this ladder for free. Data collected by watching the world — no matter how much of it you have — lives on the bottom rung. It can tell you what tends to happen alongside what. But it cannot, by itself, tell you what would happen if you intervened. To make that leap from seeing to doing, you must add something to the data: knowledge, structure, assumptions about how the world’s causes are wired together.

The intuition is worth pausing on. Imagine you observe a million people, and you notice that those who take daily vitamin C supplements are healthier than those who don’t. Can you conclude that vitamin C makes people healthier? You cannot — not from this observation alone. Because the people who choose to take daily supplements are probably also the people who exercise, eat well, have the income to afford supplements and gym memberships, and generally take care of themselves. Their health might have nothing to do with the vitamin and everything to do with the kind of person who takes vitamins. To pull them apart, you would need either to run an experiment — randomly assign vitamin C to some people and a placebo to others — or to make explicit assumptions about the causal structure and reason from there. Watching is not enough. It never is.

4The theorem that says scale won’t save you

Pearl and his collaborators did not leave this as an intuition. They proved it. The result, known as the Causal Hierarchy Theorem, states, in plain language, that you cannot determine an interventional answer from purely observational data without adding causal assumptions from outside the data. The rungs of the ladder are genuinely separate; you cannot smuggle yourself up from one to the next using data alone.

The reason this theorem matters so much in the age of large models is that it applies to any model, of any size. It does not care whether you are using a simple linear regression or a neural network with a trillion parameters. It is a statement about information, not about architecture. If the information required to answer a causal question is not present in observational data — and the theorem proves that it fundamentally is not — then no algorithm, however sophisticated, can conjure it out of that data. A bigger telescope can show you more distant stars, but it can never turn a photograph into an experiment. More data and more parameters give you a sharper, more detailed picture of the correlations. They do not give you causation.

And yet. If you actually go and ask a large language model a causal question, it very often gets the answer right. This creates a genuine puzzle, and resolving that puzzle is the key that unlocks everything.

5The paradox of the causal parrot

Let us state the puzzle sharply. On one hand, the theorem we just discussed says that no model can extract causation from observation alone. On the other hand, a language model is trained almost entirely by observation — it reads enormous quantities of text and learns to predict what word comes next. That is about as pure a form of “learning by watching” as one could imagine. So by the logic of the theorem, a language model should be causally clueless. And yet, empirically, these models answer causal questions correctly all the time. How can both things be true?

A group of researchers — Zečević, Willig, Dhami, and Kersting — gave this puzzle a memorable name and a satisfying resolution in a paper titled Causal Parrots: Large Language Models May Talk Causality But Are Not Causal. Their argument builds on an earlier and equally memorable idea: the “stochastic parrot,” a term coined to describe how language models produce fluent, convincing text by echoing the statistical patterns in their training data, without any genuine understanding of meaning. The parrot says “hello” at the right moments not because it understands greeting, but because it has learned that certain sounds tend to follow certain situations.

Zečević and colleagues extend this metaphor into the realm of causality. Language models, they argue, are not only stochastic parrots but also causal parrots. When a model correctly tells you that smoking causes cancer, it is not because it has reasoned its way from evidence to a mechanism. It is because, somewhere in the ocean of text it was trained on, that causal statement appeared — many times, in many forms — and the model has learned to reproduce it. It is reciting, not reasoning.

To see why this distinction is so sharp, consider a single fact: at higher altitudes, the temperature drops. This is a genuine causal truth, and it exists in two completely different forms. In the first form, it exists as data: thousands upon thousands of paired measurements of altitude and temperature, collected by weather stations on mountainsides and in valleys around the world. A scientist who gets hold of this data can run a causal-discovery procedure on it, work out that altitude drives temperature and not the other way around, and arrive at genuine understanding — an understanding grounded in evidence about the underlying mechanism, the thinning atmosphere and its reduced capacity to hold heat.

In the second form, the very same fact exists as text: the sentence, sitting in an encyclopedia or a textbook, “temperature normally decreases with altitude.” Anyone who reads this sentence can repeat it. They now know the fact. But knowing the fact in this way is a hollow sort of knowing. It carries no understanding of why, no ability to reason about related situations, no grounding in evidence. It is a memorized answer, not a comprehended truth.

The language model only ever encounters the second form. It never sees the mountainside measurements; it only reads the sentences people have written about them. So when it tells you that temperature falls with altitude, it is repeating a sentence, not describing a mechanism it has grasped. And this has a disturbing consequence. Because the model learns from all the text it reads, and not everything written down is true, it will just as confidently repeat false causal claims as true ones. It has no way to tell the difference, because it never had access to the evidence that would let it check. It is a student who memorized the answer key: flawless on the exact questions the key covers, hopeless on any variation, and utterly unable to notice when the key itself contains a mistake.

6The loophole: where the causal knowledge really comes from

We still have not resolved the paradox. If language models are just parrots, and if the theorem forbids extracting causation from observation, then where does the correct causal knowledge come from? A parrot that reliably says true things is not saying random things — the truth is coming from somewhere. The resolution is subtle and elegant, and once you see it, the whole picture clicks into place.

The theorem forbids inferring causation by observing a system. If you watch altitude and temperature vary, you cannot, from that observation alone and without assumptions, prove the causal direction. That much is settled. But — and here is the loophole — the theorem says nothing at all about the situation where the causal answer has already been figured out by someone else and simply written down as text that you then read.

Think about how causal knowledge actually enters the world’s stock of written text. A scientist runs a randomized controlled trial. She randomly assigns some patients to receive a drug and others to receive a placebo, thereby intervening — climbing, through the deliberate act of randomization, from the observational rung to the interventional rung. She discovers that the drug reduces mortality. And then she writes it down, in a paper, in a review, eventually in a textbook and a hundred news articles. The causal fact, established through genuine intervention, is now recorded in text.

The researchers behind the causal-parrots paper formalize this with a concept they call a meta-model — a causal model whose observable output is text describing the causal facts of some other system. And they point out that the whole of nature, together with the whole enterprise of human science, functions as exactly such a meta-model. Nature produces physical quantities, yes — but the scientific community, by studying nature through experiments, produces text about nature’s causal structure. And that text is what language models train on.

This is why we can say, all at once and without contradiction, that language models are not causal, that the theorem holds, and that models are frequently right about causation. They are right precisely and only to the extent that the right answer was written down by someone who genuinely knew it. When the answer is in the text, the parrot recites it correctly. When it is not, the parrot either falls silent or, worse, confidently makes something up.

7Three reasons a language model can never truly be causal

It is worth pausing to gather, in one place, the reasons why a language model cannot be a genuine causal reasoner. There are three, and they reinforce one another.

• The raw material. Language models learn from text, and text alone. They never observe the physical process that generates the data — never see the mountainside measurements, never watch the drug act on the body. And you cannot perform inference about a mechanism you have never observed. You could read a thousand vivid descriptions of riding a bicycle and still be unable to balance on one, because the skill lives in the physical interaction, not in the words about it.
• The machinery. Purpose-built causal models contain explicit apparatus for representing causation: structural equations, a graph of dependencies, an operator that represents the act of intervening. A language model contains none of it. It has no graph, no notion of intervention, no representation of “setting a variable” as opposed to “seeing a variable.” It has only a vast, intricate model of which words tend to follow which other words. There is simply no place inside it where a causal structure could live.
• The theorem itself. Even setting aside the first two practical limitations, even in principle, observation cannot yield causation without external assumptions. This is not a temporary shortcoming to be fixed with better engineering; it is a permanent feature of the mathematics of information. It binds every model equally, the largest as much as the smallest.

Put these three together and the conclusion is inescapable. A language model, asked directly to reason about cause and effect, is a parrot: fluent, frequently correct, but fundamentally reciting rather than reasoning, and offering no reliable way to distinguish the two. If that were the end of the story, it would be a rather deflating one. But it is not the end. It is the setup for the genuinely useful part.

8Two ways to use a language model for causality

Everything we have said so far describes one particular way of using a language model for causal questions — the naive, direct way, where you simply ask the model for the causal answer and trust what it says. Let us call this the parrot regime. In this regime, you type “Does this drug reduce mortality?” and take the model’s reply at face value.

The parrot regime is fragile for all the reasons we have explored. It only works when the exact fact happens to have been written down somewhere in training. It is inconsistent across the ladder of causation — often reasonable at the observational level, shaky and unreliable when you push it toward genuine interventional or counterfactual reasoning. It hallucinates, inventing plausible-sounding causal links that are simply false. And most dangerously of all, it gives you no signal about which of its answers are recalled truths and which are confident fabrications. Everything comes out in the same authoritative tone. You cannot tell the memorized answer key from the invented one.

But there is a second way to use the very same model, and it changes everything. Call it the structure-proposer regime. In this regime, you do not ask the model for the causal answer at all. Instead, you ask it to propose the causal structure — the graph of what causes what — by synthesizing what the scientific literature says. Then you hand that proposed structure to human experts and to real data for verification, and once it survives that scrutiny, you use it as the foundation for proper causal inference, computed not by the model but by trusted, transparent statistical machinery.

Why does this second regime work when the first one fails? Because it plays to the model’s actual strength and away from its actual weakness. The model’s genuine superpower is not reasoning; it is reading. It has ingested more scientific papers than any human being could read in a hundred lifetimes. Its weakness is that it cannot reliably reason about causes or tell truth from confident fiction. The structure-proposer regime asks the model only to do the thing it is good at — survey and summarize the causal claims scattered across a vast literature — and reserves the things it is bad at — verification, and the final causal inference — for humans and for mathematics.

There is a beautiful subtlety here that is easy to miss. Did we not just spend several sections arguing that a model’s causal knowledge is merely recited, merely borrowed from text? Yes. So how can that same recited knowledge suddenly become trustworthy? The answer is that the two regimes differ in where the knowledge lives and how much of it the model must combine. In the parrot regime, the causal claim typically rests on a single source — one sentence, one article — and the model simply repeats it, opaquely and unverifiably. In the structure-proposer regime, the relevant causal knowledge is distributed across thousands of independent studies, no single one of which contains the whole picture. Each individual study contributes, in effect, one edge of the graph. The model’s real value is synthesis at scale — assembling thousands of these one-edge findings into a single coherent candidate structure, doing in minutes what a domain expert does slowly and incompletely from memory.

And here the “meta-model” idea from earlier becomes wonderfully concrete. The scientific literature is the meta-model. Every randomized trial and careful cohort study establishes a causal fact through genuine intervention and writes it down. The model harvests those written facts. It never discovers causation from raw data — the theorem still forbids that, and always will. Instead, it rediscovers and organizes the causal claims that humans have already proved experimentally. A useful way to picture it: no single physician has read every relevant paper, and none ever will. The language model is like a tireless research librarian who has read them all and can hand you, in an afternoon, a draft map of everything the literature claims causes everything else. You would never bet a patient’s life on that draft without a senior clinician checking it and without testing it against your own data — but the librarian has just compressed months of literature review into an afternoon.

9Letting the model draft the map

Let us now walk through what the structure-proposer regime actually looks like in practice, because the details are where the value lives.

The first thing to understand is why proposing the causal structure is such a valuable service. In the standard practice of causal inference, the genuinely hard, slow, and expensive step is not the mathematics. It is building the graph — deciding which variables exist, which ones influence which others, and in what direction the influence flows. Traditionally, this graph is elicited from domain experts in long meetings, drawn on whiteboards, argued over, and revised. It is subjective, it depends on which experts happen to be in the room and what they happen to remember, and it does not scale. And it is the foundation on which everything else rests: get the graph wrong, and every number you compute afterward is quietly, invisibly corrupted.

This is precisely the bottleneck a language model is positioned to relieve, because it has effectively read more of the relevant literature than any single expert could hold in their head. Used well, it can propose the edges of the graph and their directions, assembling a first-draft structure in a fraction of the time.

The phrase “used well” carries a lot of weight. There is a right way and a wrong way to query a model for causal structure. The wrong way is to ask it, in one breath, to produce the entire graph — an invitation for it to generate a confident, elaborate, and possibly fictional diagram. The right way is more disciplined. You ask it pairwise questions: does this specific variable cause that specific one, does the influence run the other way, or is there no direct link at all? You give it clear definitions of each variable so it is not guessing at your meaning. You ask it to name the mechanism or the study behind each proposed edge, so that every claim it makes is something a human can go and check. And you push it especially hard on direction, because getting the arrow pointing the right way is both the most important and the most error-prone part of the whole exercise.

Consider a concrete medical example. Suppose we are studying whether a blood-pressure medication reduces the risk of stroke, and our variables are age, smoking, blood pressure, the treatment itself, and stroke. Drawing on the vast cardiovascular literature, the model might propose that age raises blood pressure, that smoking raises blood pressure, that age and smoking and blood pressure all raise stroke risk, and — critically — that the treatment reduces stroke risk by acting through blood pressure. That last relationship, where the treatment influences the outcome by way of an intermediate variable, marks blood pressure as a mediator, and whether a variable is a mediator or a confounder completely changes how you must handle it in the analysis. Getting that distinction right is exactly the sort of thing a well-queried model, synthesizing the literature, can help surface in minutes.

Now, this drafted graph is not the theorem’s enemy; it is the theorem’s answer. Recall that the theorem said you cannot get causation from data without adding external causal assumptions. The graph is those assumptions. So when the model proposes a structure, it is not defying the theorem — it is supplying, at scale and from the collective memory of science, precisely the ingredient the theorem demands. You are using the meta-model loophole deliberately and transparently, rather than crossing your fingers that a single recited fact happens to be correct.

But — and this “but” is non-negotiable — the human stays in the loop, for three concrete reasons:

• The model hallucinates. It will, with total confidence, assert links that do not exist, or draw arrows backwards. And because a single wrong edge silently corrupts every downstream calculation, these errors are not cosmetic; they are potentially fatal to the analysis.
• A graph is an assumption, not a proven fact. This follows directly from the theorem: no quantity of data and no degree of model confidence can prove a causal edge. Every arrow must ultimately be justified by mechanism, by experiment, or by expert judgment — never by the model’s say-so alone.
• Verification is both cheap and mandatory. You verify that the graph is acyclic, containing no impossible causal loops. You run conditional-independence tests against your own data: a causal graph makes specific, testable predictions about which variables should be statistically independent of which others, and if your data flatly contradicts those predictions, you have caught a wrong edge. And you have a human expert review the whole structure for plausibility. The model proposes; the human and the data dispose.

10Turning the map into an answer

Suppose we have done the work of the previous section. We have a candidate causal graph, drafted by the model, scrutinized by an expert, and tested against our data. It has survived. Now comes the payoff, and it is a big one: with a trusted graph in hand, we can legitimately compute a causal effect from observational data alone — no randomized trial required.

The first gift the graph gives us is an answer to the question that trips up more analyses than any other: which variables should I adjust for? In everyday statistical practice, people often deal with confounding by “controlling for” other variables — throwing them into a regression to hold them fixed. The trouble is that controlling for the wrong variables does not merely fail to help; it actively introduces bias. Without a causal graph, deciding what to control for is little more than educated guesswork. The graph replaces the guesswork with a rule.

To apply the rule, you need to recognize three kinds of variables, and treat each completely differently:

• Confounder — a common cause of both the treatment and the outcome. In our medical example, age causes both who gets treated and who has a stroke, so age is a confounder. The rule is firm: you must adjust for them. If you fail to, age’s effect on stroke will masquerade as the treatment’s effect, and you will draw a false conclusion.
• Mediator — a variable that sits on the causal pathway between treatment and outcome. Our treatment reduces stroke risk by lowering blood pressure, so blood pressure is a mediator. Here is the counterintuitive part: you must not adjust for a mediator if you want the treatment’s total effect. Since the treatment works precisely by changing blood pressure, holding it fixed would mask the very effect you are trying to measure — you would conclude, wrongly, that the treatment does nothing.
• Collider — a common effect of two others, and the most treacherous of the three. Consider fame among actors: it helps enormously to be either exceptionally talented or exceptionally good-looking. If you look only at famous actors and examine talent versus looks, you will likely find a negative correlation — not because the two are at odds in the general population, but because restricting attention to the famous conditioned on the collider and conjured a correlation out of thin air. The rule: you must not adjust for colliders, because doing so invents relationships that are not real.

The reason all of this matters so much is that these three types look identical in raw data. A confounder, a mediator, and a collider are all just “some other variable correlated with the ones you care about.” Nothing in the data itself tells you which is which. Only the causal graph, with its arrows, reveals their true roles — and therefore only the graph can tell you the correct set of variables to adjust for.

Once the graph has told you the correct set of confounders to adjust for, a formula called the backdoor adjustment computes the causal effect. The logic is intuitive once you have the confounders in hand. To estimate the effect of a treatment while accounting for a confounder like age, you do not compare all treated patients against all untreated patients — that comparison is contaminated, because the treated group might be systematically older or younger. Instead, you compare treated against untreated patients within each age band separately, then average those within-band effects together, weighted by how common each age band is in the population. By comparing like with like within each stratum of the confounder, you close the “backdoor” through which age was leaking its influence into your estimate. What you are left with is an estimate of the treatment’s genuine causal effect — you have climbed from the observational rung to the interventional rung, and the graph’s assumptions are the ladder that let you do it.

With the confounders settled by the graph, the choice of statistical estimator becomes a downstream, swappable matter of taste and robustness. You can use plain regression, reweighting methods that make the treated and untreated groups comparable, doubly-robust methods that give a correct answer as long as either of two underlying models is right, or flexible machine-learning estimators that capture how the effect varies across subgroups. The hard causal decision — what to adjust for — was made by the graph; the estimator merely does the arithmetic afterward.

And the payoff, once those assumptions are met, is genuinely transformative. You obtain a causal effect estimate from observational data alone, without ever running a randomized controlled trial. In medicine this is enormous, because trials are frequently impossible, unethical, prohibitively expensive, or simply too slow to help the patients in front of you today. You cannot randomize people to smoke. You often cannot ethically withhold a promising treatment. But you can, with a trustworthy causal graph and a body of observational data, extract a defensible causal answer where a trial is out of reach.

11The whole story, told straight through

Let us draw the threads together into a single narrative, because the individual pieces are most convincing when you see how they lock into one another.

Correlation is not causation, and the reason is that observation alone cannot distinguish a genuine cause from a hidden common cause. This is not a practical limitation to be engineered away; it is a mathematical fact, proven as the Causal Hierarchy Theorem, and it binds every model regardless of its size. So a language model, which learns purely by observing text, cannot in principle be a genuine causal reasoner — and when you ask it a causal question directly, it behaves as a parrot, reciting causal-sounding statements it absorbed from its training text, fluent and often right but fundamentally unable to reason and unable to signal when it is fabricating.

The only reason such a parrot is ever reliably correct is that human scientists, through actual experiments, established causal facts and wrote them down, and the model read them. The scientific literature functions as a vast recording of humanity’s hard-won causal knowledge, and the model is, at bottom, a spectacularly capable reader of that recording.

That reframing is what turns a limitation into an opportunity. Instead of asking the model for causal answers, we ask it to survey the entire literature and propose the causal structure — the graph of what causes what — which is precisely the external assumption the theorem tells us we need and cannot get from data. A human expert and our own data then verify that structure, catching the model’s hallucinations and confirming its proposals against reality, because a graph is an assumption that must be earned, never a fact the model can decree. And once we hold a trusted graph, established causal machinery — the backdoor adjustment and its relatives — converts ordinary observational data into a real causal effect estimate, with no randomized trial required.

The model, in this final and correct picture, never becomes the causal reasoner. It becomes something both more modest and more genuinely useful: the fastest and most widely read librarian of causal knowledge the world has ever had, feeding a disciplined pipeline in which humans and mathematics do the reasoning it cannot do itself.

What to take away

If you remember nothing else, remember this. A language model asked to reason about cause and effect is a parrot — impressive, fluent, frequently right, and fundamentally untrustworthy, because it recites rather than reasons and cannot tell you which is which. But a language model asked to read the world’s science and propose the structure of causes, with humans and data standing guard over its output, is a genuine accelerant to causal discovery. The difference between these two uses is not a matter of prompt engineering or model size. It is a matter of understanding what the tool actually is, and having the discipline to give it the job it can do while keeping the jobs it cannot do — the reasoning, the verification, the final inference — firmly in human hands.

Conceptual sources: Zečević, Willig, Dhami & Kersting, “Causal Parrots: Large Language Models May Talk Causality But Are Not Causal,” Transactions on Machine Learning Research, 2023; Bareinboim, Correa, Ibeling & Icard, “On Pearl’s Hierarchy and the Foundations of Causal Inference,” 2022; Judea Pearl, “Causality: Models, Reasoning, and Inference,” 2009. All examples and explanations have been rewritten and synthesized for a general audience.

← Back to Research