The Problem of Induction

In brief

You have seen the sun rise thousands of times, so you expect it to rise tomorrow. You have released stones and watched them fall, so you expect the next one to fall too. Every such expectation runs from cases you have observed to cases you have not. David Hume (1711 to 1776) asked the simple, ruinous question: what justifies that step? Not logic, because there is no contradiction in supposing the sun fails to rise. Not experience, because experience only ever tells you what has happened, never what will. The bridge from past to future is a principle, that nature is uniform, that we can neither prove nor observe. Any attempt to support it by past success ("induction has worked before") already uses induction, and so begs the question. This is the problem of induction, and it sits underneath every empirical claim science makes.

The full treatment

The problem it answers

Deductive reasoning is safe but sterile. If all men are mortal and Socrates is a man, then Socrates is mortal, and the conclusion adds nothing not already contained in the premises. Ampliative reasoning, by contrast, is reasoning that reaches beyond its evidence: from "every observed raven is black" to "all ravens are black," or from "this drug helped these patients" to "it will help the next." This is the reasoning that lets us predict, generalize, and act. It is also, Hume showed, the reasoning we cannot vindicate. The problem of induction is the question of whether ampliative inference is ever rationally justified, and if so, how.

How Hume's argument works

Hume's argument, given in A Treatise of Human Nature (1739) and restated in An Enquiry Concerning Human Understanding (1748), is a dilemma. Take any inductive inference: the future will resemble the past in some respect. What could justify it?

The justification must be either demonstrative (a matter of pure logic) or probable (a matter of experience). It cannot be demonstrative, because there is no contradiction in imagining the course of nature changing: bread that has always nourished might, tomorrow, poison. A world where the future breaks with the past is perfectly conceivable, so no logical proof rules it out. So the justification must come from experience. But every appeal to experience presupposes exactly what is in question, namely that the future will resemble the past. To argue "the uniformity of nature has held so far, therefore it will keep holding" is to use an inductive inference to justify induction. The support is circular.

Hume's conclusion is not that induction is irrational and we should stop. It is that induction has no rational foundation and we do it anyway. We are led to expect the future to resemble the past not by reason but by custom or habit, a "gentle force" in the mind that forms the expectation once we have seen events repeatedly conjoined. Belief in the unobserved is a fact about human psychology, not a conclusion of argument.

The key example and text

The cleanest illustration is the one Bertrand Russell (1872 to 1970) gave in The Problems of Philosophy (1912). A chicken is fed by the farmer every morning. Each feeding strengthens, on impeccable statistical grounds, its expectation of being fed. Its confidence peaks on the very morning the farmer wrings its neck. The chicken reasoned correctly from its record; the record simply did not contain the one event that mattered. The point is not that induction usually fails. It is that no amount of confirming evidence, however large, logically secures the next case, because the sample can never include what has not yet happened.

Goodman's new riddle: grue

For two centuries the problem was about whether induction can be justified at all. Nelson Goodman (1906 to 1998), in Fact, Fiction, and Forecast (1954), showed there is a deeper problem even if we grant that induction is legitimate: we cannot say which inductions are the valid ones.

Define the predicate grue: something is grue if it is examined before some future time t and found green, or is not so examined and is blue. Every emerald we have ever inspected is green. But every one of those emeralds is also grue. So our evidence supports "all emeralds are green" and "all emeralds are grue" equally well. Yet these predict opposite things about emeralds first seen after t: green predicts green, grue predicts blue. The same past, the same rule of induction, licenses contradictory forecasts.

The natural reply is that green is a real, natural property and grue is a gerrymandered fake that hides a time reference. Goodman's counter is devastating: from the point of view of a speaker whose basic vocabulary is grue and bleen, it is green that looks time-dependent and artificial. There is no vocabulary-neutral standpoint from which green is obviously the "natural" predicate. Which predicates are fit for induction, which Goodman called the "projectible" ones, cannot be read off the world. He suggested they earn their status through a track record of use, what he called "entrenchment," but that only relocates the problem into the history of language.

Distinctions that matter

Three distinctions keep the debate clean. First, Hume's old riddle asks whether induction can be justified; Goodman's new riddle asks which inductions to make, given that we make them. Second, the problem is not skepticism about the external world; Hume grants that the sun probably will rise, he denies only that reason is what tells us so. Third, the problem is not solved by probability. Saying an outcome is "highly probable" on past evidence still assumes the future resembles the past, so probabilistic inference inherits the circularity rather than escaping it.

Lineage

The problem grows out of the empiricist claim that all knowledge of matters of fact comes from experience. If experience is the only source, and experience is always of the past, then knowledge of the future has no source, which is Hume's discovery turned on its own tradition. Immanuel Kant (1724 to 1804) famously wrote that Hume woke him from his "dogmatic slumber," and his response in the Critique of Pure Reason (1781) was to argue that causation and uniformity are not learned from experience but are conditions the mind imposes on any possible experience, so they hold necessarily of the world as we can know it. In the twentieth century the problem became the hinge of philosophy of science: Karl Popper (1902 to 1994) built his falsificationism on the claim that science avoids induction entirely, while Rudolf Carnap (1891 to 1970) and the logical empiricists tried to build a formal logic of confirmation that Goodman's grue example then wrecked. The problem also underlies the practical warnings of The Black Swan.

The strongest case for it

The argument is unusually hard to escape because it is almost purely formal. Grant only two things: that a break in nature's uniformity is conceivable (so no logical proof forbids it), and that circular arguments do not justify their conclusions. From those alone the conclusion follows. Centuries of attempted rebuttals have mostly confirmed the diagnosis rather than dissolved it. Every appeal to the past reliability of induction turns out to smuggle induction back in; every appeal to probability turns out to assume the very uniformity in question. Goodman's riddle then closes off the retreat: even if you help yourself to induction, the world does not hand you the categories to run it, so the shape of your predictions depends on choices no evidence dictates. That two independent arguments, one about justification and one about content, converge on the same gap is strong evidence the gap is real and not a trick of phrasing.

The strongest case against it

Serious philosophers have pushed back, and the best replies do not deny the logic so much as reject the demand.

Peter Strawson (1919 to 2006), in Introduction to Logical Theory (1952), argued that asking whether induction "in general" is justified is a confusion. To call a belief reasonable just means, in ordinary use, that it is well supported by evidence in the inductive way. Demanding a further, external justification of induction is like asking whether the law is legal: the standard of reasonableness is the inductive standard, so the question answers itself.

Hans Reichenbach (1891 to 1953) offered a "pragmatic vindication." We cannot prove induction will work, but we can prove that if any method of prediction works, induction will work too, because it converges on the true frequencies in the long run whenever there are stable frequencies to find. Induction is thus the rational bet even without a guarantee: it wins if winning is possible at all.

Karl Popper cut the knot by denying that science uses induction at all. On his view, scientists never infer laws from instances; they conjecture bold theories and try to refute them. A single counterexample can falsify a universal claim, so knowledge grows by elimination, not accumulation, and Hume's problem never arises because the inductive step is never taken. Critics answer that Popper still needs induction to trust that a theory which has survived tests will keep performing, a difficulty he named "corroboration" without fully resolving.

Bayesians reframe the whole matter as the updating of degrees of belief. Frank Ramsey (1903 to 1930) and others argued that rationality requires only that your beliefs obey the probability calculus and update by conditioning on evidence; the "problem" reduces to a choice of prior probabilities, which is a matter of coherence rather than proof. Detractors reply that this makes induction merely internally consistent while leaving Hume's real question, why those priors track the world, exactly where he left it.

Where it stands now

There is no consensus solution, and most working philosophers treat the problem as genuine and unsolved rather than defused. The dominant view is closer to Hume's own: induction cannot be given a non-circular foundation, and the interesting work lies elsewhere, in specifying which inductive methods are best (a descriptive and normative project about scientific practice) rather than in vindicating induction as such. Goodman's new riddle remains an active problem in the theory of confirmation and now surfaces in machine learning as the "no free lunch" results and the study of inductive bias, where the grue point reappears exactly: an algorithm cannot generalize from data without built-in assumptions about which patterns count, and no such assumption is dictated by the data alone. Bayesian epistemology has become the main formal framework, which sharpens the question of where priors come from without answering Hume's original challenge. The problem endures because it is not a puzzle to be cleared away but a permanent feature of any mind that must act on more than it has seen.

Test yourself

You almost certainly believe the sun will rise tomorrow, and you are right to. Now try to state the reason, without at any point using the fact that it has risen before. If you cannot, you have felt the force of Hume's argument. The next question is the honest one: does that missing justification change anything at all about how you should live and reason, or is acting on a habit you cannot ground simply the human condition?