Philosophy of Science: Foundations

What It Is

Philosophy of science is the part of philosophy that asks: what makes a body of inquiry scientific; how do scientific theories change; what do our best theories actually claim about reality; and what kind of inference licenses our scientific knowledge in the first place. The discipline is shaped by four central questions, each of which has a canonical position and a canonical critique.

1. The demarcation problem. What distinguishes science from non-science (pseudo-science, metaphysics, mere description)? Popper's falsifiability is the most-cited answer.
2. The dynamics of scientific change. How do scientific theories actually evolve over time? Kuhn's paradigm shift is the most-cited descriptive answer.
3. The realism question. When a confirmed scientific theory posits unobservable entities (electrons, genes, dark matter), should we believe these entities actually exist? Scientific realists say yes; antirealists say science aims at empirical adequacy, not truth about unobservables.
4. The induction problem. Scientific inference goes from observed data to general laws and unobserved cases. Hume showed this inference cannot be justified deductively. The contemporary reply runs through Bayesian epistemology.

These questions are interconnected. Falsifiability presupposes a view about how observations bear on hypotheses (the induction problem); paradigm shifts complicate the realism question (do the entities of the old paradigm "really" exist?); the induction problem in turn motivates the demarcation discussion (a non-falsifiable theory is one that no inductive evidence could discriminate against). The page walks all four.

This page assumes Induction and Hume's Problem for the technical statement of the induction question, and What Is Philosophy? for general background. The formal credence-based reply to Hume is given in depth in Bayesian Epistemology.

Popper and Falsifiability

Karl Popper, in The Logic of Scientific Discovery (1934, English 1959), proposed that the distinguishing mark of empirical science is falsifiability: a statement is scientific iff it forbids some possible observation. The statement "all swans are white" is scientific because it forbids the observation of a non-white swan; the statement "everything happens for a reason" is non-scientific because no possible observation could refute it.¹

Popper's framing is negative: science does not verify general laws (Hume showed verification by induction is unavailable); science attempts to falsify them. A theory survives by repeatedly resisting falsification attempts; the more risky the predictions a theory survives, the higher its corroboration. Corroboration is not probability of truth on the evidence; it is a measure of the theory's success in resisting refutation.

The strongest motivating contrast for Popper was Freudian psychoanalysis and Marxist historiography: theories that, in his view, could be made compatible with any conceivable observation by post-hoc reinterpretation. The contrast with general relativity (which made the risky prediction that light would bend near the sun, testable in the 1919 Eddington eclipse expedition) was paradigmatic. Falsifiable theories take risks; non-falsifiable theories do not.

Five serious objections to the falsifiability criterion.

The Duhem-Quine thesis. A theory $T$ together with auxiliary hypotheses $A$ entails an observable prediction $O$. If $\neg O$ is observed, this falsifies the conjunction $T \land A$, not $T$ alone. The scientist can always preserve $T$ by modifying $A$ instead. Strict falsifiability is undermined: no single theory faces the tribunal of experience alone.²

Ad hoc rescue. Popper himself acknowledged that theories can be saved from falsification by ad hoc modifications. He attempted to rule these out by stipulating that legitimate modifications must increase rather than decrease falsifiability. But the distinction between ad hoc and legitimate auxiliary-hypothesis modification is itself unclear; sophisticated theorists modify auxiliaries all the time without obviously violating scientific norms.

Existence statements. "Unicorns exist" forbids no observation: no finite search of the world rules it out. By Popper's criterion existence statements are non-scientific, which is implausible. Popper acknowledged the asymmetry and argued the criterion applies to general laws, not existence claims, but the qualification weakens the original thesis.

Probabilistic theories. Quantum mechanics predicts probability distributions, not deterministic outcomes. No finite sequence of outcomes is strictly inconsistent with any probability distribution. Strict falsificationism therefore cannot accommodate the probabilistic theories that dominate modern science.

Confirmation matters too. Scientists clearly do count repeated successful predictions as positive evidence for a theory, not merely as failed falsification attempts. Popper's corroboration concept is widely judged not to deliver the asymmetry he wanted between accumulating corroboration and confirming the theory.

Despite the objections, falsifiability remains the most-cited demarcation criterion in popular discussion of science. The contemporary professional consensus is that falsifiability captures something about science (theories should make risky predictions; theories should not be unfalsifiably reinterpretable) but does not provide a sharp demarcation.

Kuhn and Paradigm Shifts

Thomas Kuhn's The Structure of Scientific Revolutions (1962, second edition 1970)³ gave a descriptive account of how science actually changes over time. Kuhn argued that science alternates between two phases.

Normal science. A community of scientists works within a shared paradigm: a constellation of theoretical assumptions, methodological commitments, exemplary problem-solutions, and trained intuitions. Normal science is puzzle-solving: applying the paradigm to particular problems within its scope. Anomalies (observations that resist the paradigm's tools) accumulate but are usually treated as recalcitrant puzzles rather than as refutations.

Crisis and revolution. Anomalies accumulate; new techniques cease to dissolve them; the paradigm begins to be seen as exhausted. A crisis develops. New candidate paradigms emerge. Eventually one of them is adopted by enough of the community that the paradigm has shifted; normal science resumes within the new paradigm.

The most contested feature is incommensurability: post-revolution and pre-revolution paradigms cannot be straightforwardly compared, because the meaning of central terms shifts. Newtonian mass and Einsteinian mass are not the same concept: the latter is frame-dependent, the former is not. A simple comparison "Newton said mass is $m$, Einstein said mass is $\gamma m$" misrepresents what the two theories actually claim.

Three readings of Kuhn's incommensurability thesis.

• Strong (taxonomic). Old and new paradigms employ kind-terms that partition the world differently; no term-by-term translation is available. Communication across the divide is impossible.
• Weak (translation difficulty). Translation is hard and requires care; some terms shift meaning across the divide and need to be relearned, but theories can still be compared on evidence.
• Methodological. What counts as a good theory (simplicity, predictive accuracy, fit with other theories) is itself paradigm-relative; there is no fully neutral standard for evaluating paradigms against one another.

Kuhn vacillated between these readings; later work clarifies the weak / methodological readings as more defensible than the strong taxonomic reading.

Objections. Kuhn underplays the role of empirical evidence. Even across paradigm shifts, scientists agree on a vast set of observable phenomena (perihelion of Mercury, Brownian motion, the photoelectric effect) and weigh them as evidence for the candidate paradigm. The "irrational gestalt switch" picture of paradigm change is overdrawn. The strong taxonomic incommensurability thesis is incompatible with practice. Working physicists routinely translate between Newtonian and relativistic frameworks for purposes of approximation and pedagogy. The thesis as Kuhn sometimes stated it is too strong to be true.

Kuhn's enduring contribution is the normal-science-as-puzzle-solving picture and the recognition that scientific change is gestalt in character, not purely cumulative; even critics adopt the vocabulary of "paradigm" and "anomaly" and "crisis."

Scientific Realism vs Antirealism

When a successful scientific theory posits unobservable entities (electrons, genes, distant galaxies before we could photograph them), what is the right attitude to take to the entities?

Scientific realism. Our best confirmed theories are approximately true, and the unobservable entities they posit really exist. The success of science (its predictive accuracy, its technological fruitfulness, the unification it achieves) would be miraculous if the theories were not at least approximately tracking the world's structure. This is the no-miracles argument (Putnam, Boyd).

Constructive empiricism (van Fraassen 1980⁴). Science aims at empirical adequacy (saving the observable phenomena), not truth about unobservables. We should accept a theory when it is empirically adequate but withhold belief in its unobservable posits. The position is agnostic about unobservables, not skeptical: we have no positive evidence to believe in them beyond what is needed for prediction.

The pessimistic meta-induction (Laudan 1981⁵). The history of science is a graveyard of confirmed theories: phlogiston, caloric, the ether, Newtonian absolute space. Each was empirically successful in its day; each is now considered wholly false in its claims about unobservables. By induction over the historical record, we should expect our current best theories to be similarly wrong about their unobservable posits. The no-miracles argument is therefore unconvincing.

The realist response splits into two camps.

• Structural realism (Worrall 1989⁶). What is preserved across theory change is structure, not entities. Fresnel's wave theory of light got the structure right (mathematical wave equations); the entity (the ether) was wrong; modern theories preserve the structure and replace the entity. Structural realism limits realism to the mathematical / relational features of theories.
• Selective realism / entity realism (Hacking 1983, Cartwright 1983). We should be realists about entities we can causally manipulate ("if you can spray them, they exist": Hacking on electrons) but agnostic about the high-level theoretical claims that go beyond manipulation. This narrows the realist commitment but defends entities at the manipulation level.

The debate has no resolution. Most working scientists are intuitive realists; most professional philosophers of science recognize antirealism as a serious live option that captures the historical record more honestly.

The Induction Problem and the Bayesian Reply

Hume in the Treatise (1739) and Enquiry (1748) argued that inductive inference cannot be rationally justified. The argument: we infer from observed cases (the sun has risen every day) to unobserved cases (the sun will rise tomorrow) by assuming the uniformity of nature (similar circumstances will produce similar outcomes). But the uniformity assumption can only be justified inductively (we have observed nature being uniform in the past), and using induction to justify the assumption that licenses induction is circular.⁷

For the full statement and contemporary replies see Induction and Hume's Problem. The two main contemporary responses are sketched here.

Bayesian conditionalization. A Bayesian agent has prior credences over hypotheses; observed evidence updates the credences via Bayes's rule. Inductive support is not deductive entailment; it is the rational raising of credence in a hypothesis by evidence that the hypothesis predicts. Hume's circularity does not arise because the framework does not claim deductive justification; it claims rational credal update. See Bayesian Epistemology for the formal treatment.

Inference to the best explanation (Harman 1965). Inductive inference is reconstructed as inference to the best explanation: we infer to the hypothesis that, if true, would best explain the observed evidence. The justification is not deductive but practical: explanatory inference is a constitutive part of rational inquiry, and demanding deductive justification for it presupposes too much.

The induction problem is not solved in the sense that Hume's original challenge is fully met; it is reframed. Both responses give up on the demand for deductive justification of induction and offer instead a different theoretical role for inductive inference. Whether this counts as a solution or as a strategic retreat is contested.

Common Misconceptions

• "Popper proved that no theory can be confirmed." No. Popper argued that strict logical verification of universal generalizations is impossible (Hume's point) and that falsification is logically available where verification is not. He did not deny that theories accumulate corroboration through successful predictions; he gave a particular technical reading of what corroboration amounts to.
• "Kuhn proved that science is irrational." No. Kuhn argued that scientific change is not a purely cumulative process and that paradigm choice involves more than evidence (community judgment, theoretical virtues, training). He explicitly resisted the reading that paradigm choice is irrational or arbitrary; later editions of Structure try to clarify this.
• "Falsifiability and Bayesianism are alternative theories of science." They address different questions. Falsifiability is a demarcation criterion; Bayesianism is a theory of how evidence confirms or disconfirms hypotheses. They are not in direct competition; a Bayesian can accept a (weakened) falsifiability criterion as a heuristic for which hypotheses are scientifically interesting.
• "Scientific antirealists deny that science is reliable." No. Constructive empiricists agree that our best theories are extraordinarily reliable for predicting observable phenomena; they refuse only the further claim that the unobservable entities the theories posit really exist. The position is metaphysically modest, not skeptical about science's track record.
• "The pessimistic meta-induction defeats realism." It is a prima facie challenge, not a defeat. The realist response (structural realism, selective realism) attempts to identify what is preserved across theory change and to limit realist commitment to that. Whether the response succeeds is contested.

Comparisons to Related Views

Go Further

• Stanford Encyclopedia of Philosophy, "Karl Popper" by Stephen Thornton.
• Stanford Encyclopedia of Philosophy, "Thomas Kuhn" by Alexander Bird.
• Stanford Encyclopedia of Philosophy, "Scientific Realism" by Anjan Chakravartty.
• Stanford Encyclopedia of Philosophy, "The Problem of Induction" by Leah Henderson.
• Popper, Karl R. Conjectures and Refutations: The Growth of Scientific Knowledge. Routledge, 1963. The most accessible primary statement.
• Kuhn, Thomas S. The Structure of Scientific Revolutions. 2nd ed. University of Chicago Press, 1970.
• van Fraassen, Bas C. The Scientific Image. Oxford University Press, 1980. The canonical antirealist statement.
• Godfrey-Smith, Peter. Theory and Reality: An Introduction to the Philosophy of Science. University of Chicago Press, 2003. The clearest single textbook.
• Okasha, Samir. Philosophy of Science: A Very Short Introduction. 2nd ed. Oxford University Press, 2016. Excellent short primer.

For the formal credence-update reply to the induction problem, see Bayesian Epistemology. For Hume's original argument, see Induction and Hume's Problem.