The problem of induction

The problem of induction is the philosophical question of whether and how inductive reasoning — the inference from observed instances to unobserved ones — can be rationally justified.⁷ First given systematic formulation by David Hume in the eighteenth century, the problem strikes at the foundations of empirical science, which relies pervasively on the assumption that patterns observed in the past will continue into the future. If this assumption cannot be justified without circularity, then the entire edifice of scientific knowledge rests on a foundation that resists rational defense. The problem has generated an enormous philosophical literature and remains one of the central questions in the philosophy of science, with implications for the role of prediction in science, the demarcation problem, and the logic of confirmation more broadly.^{7, 3}

Hume’s original formulation

David Hume articulated the problem of induction in both A Treatise of Human Nature (1739–1740) and the later An Enquiry Concerning Human Understanding (1748). Hume began by distinguishing two types of knowledge: “relations of ideas,” which are known a priori and whose denial involves a contradiction (as in mathematics and logic), and “matters of fact,” which are known through experience and whose denial is always conceivable without contradiction.¹ All reasoning about matters of fact, Hume argued, is founded on the relation of cause and effect, and our knowledge of causal relations is derived entirely from experience. We observe that fire has always been accompanied by heat and that bread has always nourished, and on this basis we expect fire to produce heat and bread to nourish in future cases.²

Hume then posed the critical question: what is the rational foundation for this inference from past experience to future expectation? The inference appears to depend on an assumption that Hume called the “uniformity of nature” — the principle that the future will resemble the past, or more precisely, that unobserved cases will resemble observed ones in the relevant respects.¹ But how is this principle itself to be justified? It cannot be justified a priori, since there is no logical contradiction in supposing that nature’s course might change. It cannot be justified by appeal to past experience (the fact that nature has been uniform so far), since any such appeal would itself presuppose the very principle it aims to establish — it would assume that because uniformity has held in the past, it will hold in the future, which is precisely the inference in question.^{1, 7}

Hume concluded that inductive inference cannot be given a rational justification. It is founded not on reason but on custom or habit: the repeated experience of constant conjunctions between events produces in us a psychological expectation that the conjunction will continue, but this expectation is a matter of human nature, not of logic.² Hume did not regard this conclusion as grounds for skepticism about the practical reliability of induction — he acknowledged that we cannot and should not cease to form expectations based on experience — but he insisted that the philosophical question of justification admits of no satisfactory answer. The gap between what experience has shown and what we expect it to show cannot be bridged by reasoning alone.¹

Goodman’s new riddle of induction

Nelson Goodman, in Fact, Fiction, and Forecast (1955), transformed the problem of induction by revealing a deeper difficulty than Hume had identified. Even if the general principle that the future will resemble the past could somehow be justified, Goodman showed that this principle alone is insufficient to determine which inductive inferences are warranted. The problem is not merely that induction lacks a deductive justification, but that the very notion of “resembling the past” is radically ambiguous.⁵

Goodman introduced the predicate “grue,” defined as follows: an object is grue if and only if it is first examined before some future time t and is green, or is not first examined before t and is blue. All emeralds observed to date are green, but they are equally grue — every green emerald examined before time t satisfies both predicates. Thus the evidence that all examined emeralds are green confirms both the hypothesis “all emeralds are green” and the hypothesis “all emeralds are grue,” yet the two hypotheses make incompatible predictions about emeralds first examined after time t: the first predicts they will be green, the second predicts they will be blue.⁵

Goodman’s point was not that we should seriously entertain the possibility that emeralds will turn blue. Rather, he demonstrated that the logical structure of inductive arguments, considered alone, cannot distinguish between legitimate and illegitimate generalizations. Something beyond the form of the argument — some account of which predicates are “projectible” (suitable for use in inductive generalizations) and which are not — is needed to make induction work. Goodman’s own proposal was that projectibility is determined by “entrenchment”: predicates that have a long history of successful use in inductive inferences (like “green”) are more projectible than novel constructions (like “grue”).⁵ W. V. O. Quine later argued that projectibility tracks natural kinds — genuine categories of the world as identified by successful science — though this appeal to natural kinds raises its own circularity concerns, since the identification of natural kinds itself depends on inductive inference.¹⁰

The new riddle has generated an extensive literature and remains a live problem in philosophy of science.¹² It demonstrates that any solution to the problem of induction must do more than justify the general principle that experience is a guide to future expectation; it must also explain why certain generalizations, and not others, are supported by the evidence.

Popper’s falsificationist response

Karl Popper accepted Hume’s conclusion that induction cannot be rationally justified and proposed to dissolve the problem by arguing that science does not, and need not, rely on induction at all. In The Logic of Scientific Discovery (1934), Popper maintained that the logic of science is not inductive but deductive: scientists propose bold conjectures and then attempt to refute them through severe empirical tests. A theory that survives repeated attempts at refutation is “corroborated” — it has proved its mettle — but it is never confirmed or made probable by the evidence. The asymmetry between verification and falsification is the key: no finite number of confirming instances can establish a universal law, but a single genuine counterexample can refute one.³

Popper argued that the growth of scientific knowledge proceeds not by the accumulation of confirmed generalizations but by the elimination of error. Science advances by conjecture and refutation: imaginative hypothesis-generation followed by rigorous attempts to falsify. The method is deductive throughout, and no inductive step is required at any point.⁴ This approach connects directly to the demarcation problem: for Popper, what distinguishes science from non-science is precisely the willingness to expose theories to falsification, not the accumulation of confirming instances.

Critics have argued that Popper’s solution does not fully escape induction. The decision to rely on a corroborated theory — to use it for technological applications, policy decisions, or further research — seems to require the inductive assumption that a theory that has survived past tests will continue to survive future ones. Popper’s notion of “corroboration” is supposed to be purely backward-looking (a report of past test results), but its practical use appears to involve an implicit expectation about the future.⁷ Popper addressed this objection by distinguishing between the logical content of corroboration (which is deductive) and practical decision-making (which he conceded involves a non-rational element), but many philosophers have found this distinction unsatisfying.¹³ Thomas Kuhn further argued that Popper’s account misrepresents actual scientific practice, in which scientists typically work within an established paradigm rather than constantly attempting to refute their own theories.¹⁴

Bayesian approaches

Bayesian confirmation theory offers what many philosophers regard as the most promising framework for understanding inductive reasoning. On the Bayesian account, rational agents hold degrees of belief (credences) in propositions and update those credences in response to new evidence using Bayes’s theorem, first articulated by Thomas Bayes in 1763.⁹ The theorem states that the posterior probability of a hypothesis given the evidence is proportional to the prior probability of the hypothesis multiplied by the likelihood of the evidence given the hypothesis. As evidence accumulates, agents who begin with different prior probabilities will, under fairly general conditions, converge toward the same posterior probability — a result that provides a formal model of how experience can rationally shape belief.¹⁵

The Bayesian approach does not claim to solve Hume’s problem in Hume’s terms. It does not provide a non-circular justification for the principle that the future will resemble the past. What it does provide is a formal framework within which the degree to which evidence supports a hypothesis can be precisely quantified, and within which the conditions under which rational agents should converge in their beliefs can be specified. The prior probabilities function as the subjective element in the framework: they encode an agent’s background assumptions before the evidence is considered. Bayesian convergence theorems show that, given sufficient evidence, the influence of differing priors washes out and agents converge on the same posterior, regardless of where they started.^{8, 15}

Several objections have been raised against the Bayesian solution. The problem of priors remains significant: in many cases, the choice of prior probability is not well-constrained by the evidence, and different priors can lead to wildly different conclusions from the same data. J. M. Keynes explored this difficulty in A Treatise on Probability (1921), arguing that logical probability relations between evidence and hypothesis are objective but often indeterminate.¹¹ Additionally, some critics note that Bayesian reasoning presupposes that the true hypothesis is among those being considered — an assumption that itself cannot be justified without something resembling inductive reasoning. Goodman’s new riddle also creates difficulties for Bayesianism: assigning prior probabilities requires deciding which hypotheses to consider, and this decision reintroduces the question of projectibility.^{5, 12} The Bayesian framework is nonetheless widely employed in the evaluation of arguments across many domains, including science, philosophy of religion, and decision theory.¹⁵

Pragmatic justifications

Hans Reichenbach proposed an influential pragmatic justification for induction in Experience and Prediction (1938) and The Theory of Probability (1949). Reichenbach conceded that induction cannot be shown to be reliable — we cannot prove that it will succeed. But he argued that if any method of inference from experience will work, induction will work. More precisely, if the observed relative frequency of an event converges to a stable limit, then the inductive method of using observed frequencies to estimate that limit will eventually converge to the correct value. If the frequency does not converge — if nature is not uniform in the relevant respect — then no method of inference will succeed. Induction is therefore the best bet: it will succeed if anything will, and we lose nothing by employing it even in the case where nothing works.^{6, 16}

Reichenbach’s argument does not claim to establish that induction is reliable, only that it is optimal among available strategies. The argument has the structure of a dominance reasoning: in every possible scenario, induction does at least as well as any alternative, and in some scenarios it does strictly better. Critics have noted that the argument works only for the long run — it guarantees convergence to the correct limit eventually but says nothing about how quickly convergence occurs, and in the short run vastly different inductive rules (which weight recent evidence more or less heavily, or which start from different initial estimates) are all equally guaranteed to converge eventually. The argument therefore does not uniquely justify the specific inductive practices we employ.⁷

Other pragmatic approaches include the ordinary language dissolution associated with P. F. Strawson, who argued that asking for a justification of induction is confused in the same way as asking for a justification of deduction: inductive standards simply are what we mean by “good evidence” and “rational expectation,” and it makes no sense to demand a further justification of these standards from outside the practice.⁷ This approach has been criticized as question-begging — the skeptic can reply that the question is precisely whether our ordinary standards are correct — but it has influenced subsequent accounts that treat induction as a basic epistemic practice not reducible to deduction.

Significance and open questions

The problem of induction occupies a peculiar position in philosophy: it is widely acknowledged to be genuine and deep, yet its practical consequences are negligible. No scientist has abandoned empirical research because Hume showed that induction lacks a non-circular justification, and no engineer has declined to rely on well-tested regularities because of Goodman’s grue predicate. The problem’s significance lies not in its practical upshot but in what it reveals about the structure of empirical knowledge: that the connection between evidence and theory is never one of logical entailment, that the gap between observation and generalization must be crossed by something other than pure reason, and that any adequate epistemology of science must account for this crossing.⁷

The problem also carries consequences for debates beyond the philosophy of science. In the philosophy of religion, the logical structure of arguments from evidence — the degree to which observations of order, complexity, or suffering confirm or disconfirm theistic hypotheses — depends directly on the logic of induction and confirmation. The fine-tuning argument, the argument from evil, and probabilistic arguments for and against God’s existence all presuppose a framework for assessing evidential support, and the problem of induction reveals that no such framework can be taken for granted.¹⁵

The relationship between the problem of induction and the role of prediction in science is particularly close. The predictive success of scientific theories is frequently cited as the best evidence that science tracks truth, but this inference is itself inductive: it extrapolates from past predictive success to the likely reliability of current theories. If induction cannot be independently justified, then the claim that predictive success is evidence of truth is itself in need of justification — a point that connects the problem of induction to broader questions about scientific realism and the aims of inquiry.^{14, 7}

No consensus has emerged on a definitive solution. Popper’s falsificationism, Bayesian confirmation theory, pragmatic justifications, and ordinary language dissolutions each illuminate aspects of the problem but none has secured universal assent. What remains clear is that the problem of induction, far from being a mere historical curiosity, continues to define the boundaries of what can be known through experience and the conditions under which such knowledge is rational.⁷