Philosophy · 10 min read
What We Know and Don't Know About Causation
Every court verdict, drug approval, and AI system rests on a concept that philosophers have spent 2,000 years trying to define — and still can't. The word "cause" feels like bedrock, but the closer you look, the more it dissolves beneath you.

A few years ago, a jury in a toxic-tort case involving asbestos exposure was asked to decide something deceptively simple: did the company's product cause the plaintiff's mesothelioma? Lawyers on both sides wielded the word "cause" with absolute confidence. Expert witnesses testified about it. The judge instructed the jury on it. Billions of dollars turned on it. Yet had anyone pressed any of those participants — under oath, with philosophical precision — to define exactly what they meant by causation, the room would have gone very quiet very fast.
This isn't a rhetorical point about courtroom drama. It's the starting gun for one of the most consequential unsolved problems in human thought. The same concept that determines who pays for industrial poisoning also decides which drugs get approved by the FDA, how we interpret climate science, and how we try to prevent AI systems from going haywire. We use "cause" the way we use the floor — constantly, without thinking, trusting it completely. And philosophers have been quietly discovering that it might not hold our weight.
Hume's Bombshell: We've Never Actually Seen a Cause
David Hume fired the opening shot in 1739, and the field has never fully recovered. In his Treatise of Human Nature, the Scottish philosopher made an observation that sounds trivial until it doesn't: when you watch one billiard ball strike another, what do you actually see? You see the first ball move. You see the collision. You see the second ball roll away. What you do not see — what no one has ever seen — is the causal connection between these events. You observe sequence. You observe proximity. You never observe the invisible glue that supposedly binds cause to effect.
Hume offered two candidate definitions to fill the gap, and the fact that he couldn't choose between them is itself telling. The first: a cause is an object followed by another, where all objects similar to the first are followed by objects similar to the second — what he called "constant conjunction." The second: a cause is something that, if it hadn't existed, the effect wouldn't have followed. These are not the same definition. The first is about regularities in the world; the second is about hypothetical scenarios that may never happen. Hume knew this was a problem and essentially threw up his hands, concluding that our belief in causation is a psychological habit, not a metaphysical discovery.
The psychological dimension has since been confirmed in ways Hume couldn't have imagined. In the 1940s and 50s, Belgian psychologist Albert Michotte ran a series of now-classic experiments in which subjects watched geometric shapes move across a screen. When one shape moved toward another and the second began moving at the moment of apparent contact, people reported "seeing" one shape push the other — even though nothing was pushing anything, and even when the shapes were abstract rectangles. We don't just infer causation. We perceive it, hallucinate it, experience it as directly as we experience color. This is precisely what makes Hume's point so difficult to absorb: the illusion is phenomenologically identical to knowledge.
The Counterfactual Rescue — and Its Spectacular Failures
For much of the 20th century, philosophers thought they had Hume's problem under control. Hume's second definition — that if C hadn't occurred, E wouldn't have — was formalized by David Lewis into what became the dominant modern theory of causation. In Lewis's 1973 account, causal relationships are analyzed in terms of "possible worlds": C caused E if, in the nearest possible world where C didn't happen, E didn't happen either. It's elegant. It's intuitive. It's how lawyers argue in court and how epidemiologists design studies.
Then came the counterexamples, and they were brutal. Consider the firing squad problem. Two marksmen fire simultaneously at a prisoner. Either bullet alone would have been fatal. Both hit. The prisoner dies. Now apply Lewis's test: if Shooter A hadn't fired, would the prisoner have survived? No — Shooter B's bullet would still have killed him. If Shooter B hadn't fired? Same result. By Lewis's definition, neither shot caused the death. Common sense, law, and basic moral intuition scream that both shots caused the death, or at least that each was a cause. The counterfactual theory gives us an answer that is technically clean and obviously wrong.
This is called the problem of overdetermination, and it fractured into a family of even thornier variants. Preemption cases — where one cause gets there first and prevents a backup from operating — expose further holes. Late preemption cases, where the timing is ambiguous, generate yet more complications. Each fix philosophers proposed introduced new failures, often more elaborate than the ones it solved. By the 1990s, the literature on causation had taken on a quality of baroque desperation: towering edifices of possible-world semantics, each built to patch the last one's leaks, each leaking in new places. Lewis himself spent decades revising his theory and never produced a version that satisfied him or his critics.
Mechanisms, Probabilities, and the Scientific Workarounds
Practicing scientists, to their credit, largely stopped waiting for philosophers to sort this out. Biologists and physicians developed mechanistic accounts: a cause is something that transmits energy or information along a continuous physical pathway to its effect. This approach works well when you have a virus invading a cell or a toxin disrupting a metabolic process. It has the satisfying quality of pointing at something, of tracing a physical thread.
But mechanistic accounts choke on causation-by-absence, which turns out to be ubiquitous in medicine. What caused the sailor's scurvy? The absence of vitamin C. What caused the patient's immune deficiency? The lack of a particular enzyme. Absences don't transmit energy. They don't travel along pathways. They aren't things at all — they're holes in things. If your theory of causation can't handle some of the most important causal claims in medicine, it has a serious problem.
Probabilistic causation offered another route out. Causes, on this view, raise the probability of their effects. Smoking raises the probability of lung cancer; therefore smoking causes lung cancer. This framework proved enormously useful in epidemiology and clinical research. But philosopher J.L. Mackie showed in 1965 that the relationship between probability-raising and causation is messier than it looks, through what he called INUS conditions — an Insufficient but Necessary part of an Unnecessary but Sufficient condition. Striking a match is an INUS condition for a fire: it's not sufficient on its own (you need oxygen, dry wood, and no wind), and the whole combination isn't necessary (a spark or a lighter would also do), but within that combination, the strike is necessary. Probability-raising doesn't capture this structure. Causes can fail to raise probabilities; probability-raisers can fail to be causes. The tool is useful, but it isn't foundational.
The most honest solution science found was also its most powerful: the randomized controlled trial. By randomly assigning subjects to treatment and control groups, researchers can detect the statistical fingerprint of a causal effect without ever defining causation. If the treatment group systematically differs from the control group, and if the randomization was done correctly, something in the treatment caused that difference. It's an epistemic hack of extraordinary power. It has saved millions of lives. It also explains precisely nothing about what causation actually is — which is remarkable when you think about it. Our gold standard for detecting causes is agnostic about their nature.
Where the Confusion Costs Us: Law, Medicine, and AI
In law, the conceptual instability produces consequences that are concrete, expensive, and occasionally absurd. American tort law uses "but-for" causation as its baseline — essentially Lewis's counterfactual test dressed in legal language. But whenever that test yields obviously wrong results — as in the firing squad, or in mass-exposure cases where thousands of sources contributed to a plaintiff's injury — courts quietly abandon it and switch to a "substantial factor" test. This is not a principled transition. There is no agreed criterion for when to switch. The result is that identical facts can produce opposite verdicts depending on which jurisdiction the case is tried in, or which judge draws the assignment.
Medical causation operates under similar improvised conditions. The Bradford Hill criteria, developed by epidemiologist Austin Bradford Hill in 1965, provide a checklist for assessing causal claims: strength of association, consistency across studies, biological plausibility, dose-response relationship, and several others. They are explicitly, in Hill's own words, not a definition of causation. They are heuristics — educated intuitions about what causal evidence looks like. For well-established disease mechanisms, they work remarkably well. For newer, more complex phenomena — long COVID's multi-system effects, the precise mechanisms of statin side effects, the causal role of the gut microbiome in mental health — they produce fierce disagreement among experts who all believe they are applying the same criteria, because they are, in fact, applying incompatible unstated assumptions about what causation is.
Artificial intelligence may be where the stakes become most difficult to calculate. Judea Pearl, the computer scientist and statistician who won the Turing Award in part for his work on causal inference, has built an influential framework called the do-calculus — a formal language for representing and reasoning about interventions in causal systems. Pearl has been explicit that he considers his framework philosophically fundamental, not merely practically useful. But his framework assumes that the world has an underlying causal structure that can be represented as a directed acyclic graph — a network of nodes and arrows depicting what causes what. If our foundational concept of causation is genuinely incoherent, or merely perspectival, or a useful fiction, then the edifice Pearl has built — and that is now embedded in machine learning systems making medical diagnoses, informing policy recommendations, and guiding autonomous vehicles — rests on an assumption that may not be warranted. We are building systems that optimize for causal relationships in the world. We do not know if the world, at bottom, contains such things.
The Deepest Cut: Is Causation Even in the World?
The most radical position in this debate was staked out not by an obscure metaphysician but by Bertrand Russell, in a 1913 paper called "On the Notion of Cause." Russell's argument was simple and devastating: look at mature physics. The fundamental equations — Newton's, Maxwell's, Einstein's, Schrödinger's — describe functional relationships between quantities. They are, in their basic form, time-symmetric: they work equally well run forward or backward. Nowhere in those equations does the word "cause" appear. Nowhere is there a privileged direction of influence, a distinction between the pusher and the pushed. The concept of causation, Russell concluded, is a relic of a pre-scientific worldview, about as useful to modern physics as "the notion of fate." We kept the word long after the universe stopped needing it.
This eliminativist view has never quite gone mainstream, partly because it is difficult to live with and partly because it faces a genuine challenge: if causation is an illusion, why does the concept work so well? Why do causal explanations succeed? Why can we manipulate things so effectively using causal reasoning? James Woodward, in his 2003 book Making Things Happen, offered what may be the most sophisticated answer: causation is real, but it is perspectival. It is not a fundamental feature of the universe the way mass or charge might be. Instead, it is the structure of what would happen under interventions — what changes when an agent reaches in and manipulates a variable. Causation is, on this view, inherently tied to the possibility of action. It is not in the world as a brute fact; it is in the relationship between the world and agents who can do things in it.
This is a satisfying position in many ways. It explains why causal reasoning is so useful (we are agents; we need to know what to do), why it is perspectival (different possible interventions reveal different causal structures), and why it might disappear from the most fundamental physics (physics doesn't describe agents intervening). But it also means that causation, the concept on which we hang our legal system, our medical decisions, our scientific conclusions, and our AI architectures, is not a description of how reality is built. It is a description of how reality responds to us.
There is something vertiginous in that conclusion — and something clarifying. We are not discovering that causation is wrong or useless. We are discovering that it is ours: a tool we developed, not a feature we found. The discomfort comes from having spent millennia treating our most indispensable tool as if it were the floor rather than the hand holding it. Perhaps the deepest lesson is not that we should abandon causal thinking, but that we should hold it with exactly the kind of self-awareness we have so far refused to grant it — especially now, when the systems we're building are inheriting our concepts at scale, with no idea what they're actually inheriting.
