Philosophy of Science: What Makes Science Science

One year a headline says a daily coffee protects your heart. The next, a headline says it harms it. Then a third study finds nothing. You can respond in three ways: believe the latest, believe none, or ask a better question, which is how a result should change what you think, and when a failed prediction means the theory is wrong rather than the experiment.

In 1846 the French astronomer Urbain Le Verrier faced a planet, Uranus, that was not where Newton’s theory said it should be. He did not give up Newton. He proposed an unseen planet, calculated where it must be, and Johann Galle found Neptune that September within about a degree of the spot. Thirteen years later Le Verrier saw the same kind of problem in Mercury’s orbit, used the same move, and proposed a planet called Vulcan. It never existed. The problem was real, about 43 arcseconds of extra turn per century by modern measurement, and the culprit was Newton’s theory, replaced by Einstein’s general relativity in 1915. The reasoning was identical both times, and it succeeded once and failed once. No one can tell you in advance which kind of case you are in. That is the territory of philosophy of science.

Induction, the ravens and Bayes

Science generalises: every raven seen so far was black, so all ravens are. David Hume (1739–48) showed that no argument, deductive or inductive, can justify that step without assuming that the future will resemble the past. The same worry was raised in India by the Cārvāka school, whose rivals report them denying that inference can give knowledge because the universal link it needs can never be checked. Nelson Goodman (1955) made it worse with “grue”: an object is grue if observed before some future date and green, or not observed before then and blue. Every emerald seen so far is both green and grue. Same evidence, different predictions after the date. Which predicates are fit for generalising is not given by logic.

Carl Hempel’s ravens paradox (1945) makes the trouble concrete. “All ravens are black” is logically equivalent to “all non-black things are non-ravens”. So a white shoe, which is a non-black non-raven, seems to confirm that all ravens are black. The standard modern reply uses Bayes’ theorem: evidence confirms a hypothesis when it is more likely if the hypothesis is true than if it is false, and how much depends on that ratio. The white shoe does confirm, by a tiny amount, and a black raven confirms far more. Move the sliders to see why.

Without JavaScript: with 100 ravens, 10 of them non-black on the rival hypothesis, and 1,000 non-black non-ravens, a black raven multiplies the odds of “all ravens are black” by 1.11 and a white shoe by 1.01. Both confirm; the raven confirms roughly ten times more strongly.

This is the Bayesian answer to induction’s problem, as far as an answer exists: evidence shifts degrees of belief, and a sampling plan (looking at ravens, or at non-black things) decides how much. Its critics point to the problem of priors. Where does the starting 50% come from? Subjective Bayesians accept any coherent prior, and argue that enough evidence washes differences out; objective Bayesians look for constraints. The next sections show how much work the priors do.

Falsification, and the blame problem

Karl Popper (Logik der Forschung, 1934; English edition 1959) saw that no number of black ravens proves “all ravens are black”, but one white raven refutes it. Science, he argued, proceeds by bold conjectures that make risky predictions, and its mark is that they could fail. This is falsificationism, and it is what many scientists mean when they say a claim must be testable. Popper took relativity (whose 1919 eclipse test could have refuted it) as the model, and Freud’s and Adler’s theories, which seemed able to explain any behaviour, as the foil.

Pierre Duhem (1906) and W. V. O. Quine (1951) raised the obvious problem. A theory never meets the data alone. A prediction comes from the theory plus auxiliary assumptions: that the instrument works, that no unseen planet is pulling, that the sample is clean. When the prediction fails, logic says only that something in the bundle is wrong. Neptune and Vulcan are the same fork: Le Verrier blamed an auxiliary assumption both times. There is no rule of logic that says which to blame. How do scientists manage? Partly by priors: the part we trust least takes the blame. Compute it.

Without JavaScript: with 95% confidence in the theory and 70% in the auxiliary, an anomaly leaves the theory at about 86% and drops the auxiliary to about 14%: the blame falls on the assumption you trusted least.

Popper’s reply to Duhem was a decision, not a proof: scientists should adopt conventions that keep the main theory open to refutation. A second problem is that a strict falsifier can kill a true theory. Real measurements are noisy, and real tests sometimes fail for dull reasons. The tool below compares a strict rule (drop the theory at the first failed test) with a tolerant rule (drop it after more than k failures), using exact binomial arithmetic.

Without JavaScript: if a true theory fails 5% of tests by accident, a strict falsifier keeps it through 20 tests only about 36% of the time; tolerating two failures keeps it about 92% of the time, while a false theory that fails 40% of tests is still dropped almost every time.

Kuhn, Lakatos and Feyerabend: how theories change

Thomas Kuhn’s The Structure of Scientific Revolutions (1962) came from history, not logic. Most science, he said, is normal science: puzzle-solving inside a shared paradigm of theories, methods and exemplary solutions. Anomalies pile up and are mostly set aside. If enough accumulate and a rival appears, a crisis follows and sometimes a revolution, after which the field sees its subject differently. Kuhn’s claim of incommensurability, that rival paradigms cannot be fully compared point by point, is the most disputed part, and he later softened it. His real challenge to Popper is that real scientists do not drop a theory at a single failure, and that that is often rational, as the strict falsifier above shows.

Imre Lakatos (1970) tried to keep Popper’s rigour with Kuhn’s history. A research programme has a hard core defended by a protective belt of adjustable assumptions. It is progressive while its changes predict novel facts that turn out true, and degenerating when it only patches after the fact. That turns Neptune (new planet, new prediction, confirmed) and Vulcan (patch, no confirmation) into different cases. His critics ask how long a programme may degenerate before it is abandoned; no deadline can be given. Paul Feyerabend (Against Method, 1975) pushed the point to “anything goes”: history shows that every rule of method has been broken to good effect, so no fixed method deserves the name. He meant that no single rule could govern all science, not that all ideas are equal.

The history that these arguments draw on is global. Ibn al-Haytham’s Book of Optics (about 1011–21) tested light by designed experiments. Astronomers at the Marāgha observatory and later Ibn al-Shāṭir in Damascus (14th century) built models of planetary motion mathematically equivalent to ones Copernicus later used, and historians still debate how that knowledge travelled. Kuhn’s paradigms and Lakatos’s programmes were meant to describe such work.

Thinker	Core idea	Main objection
Popper	Science is bold, testable conjecture	Duhem: tests are of bundles; strictness kills good theories
Kuhn	Normal science, anomalies, revolutions	Relativism; vague “paradigm”
Lakatos	Programmes are progressive or degenerating	No deadline for judging
Feyerabend	No single method fits all science	Seems to leave no standard

Realism or anti-realism?

Does a successful theory describe the world, including things we cannot see (electrons, quarks, genes), or only help us predict what we can? Scientific realists give the “no miracles” argument (Hilary Putnam, 1975): the success of science would be a miracle if its theories were not at least approximately true. Anti-realists reply with the pessimistic induction (Larry Laudan, 1981): the history of science is full of successful theories, such as caloric, phlogiston and the luminiferous ether, that were later rejected as not referring to anything real. Bas van Fraassen’s constructive empiricism (The Scientific Image, 1980) says science aims only at theories that are empirically adequate, and accepting one does not require believing in unobservables. Structural realists (John Worrall, 1989) look for a compromise: what survives theory change is mathematical structure, as Fresnel’s equations for reflected light survived the fall of the elastic ether into Maxwell’s theory.

In the PhilPapers survey of 2009, about three quarters of professional philosophers (75.1%) accepted or leaned to scientific realism and 11.6% to anti-realism; the 2020 survey asked again, and its figure for this question is not given here. A neat objection to the no-miracles argument is that it ignores base rates. Success is a likely result of truth, but if most theories are false the posterior can still be low, which is the logic of a medical test with many false positives. Try it.

Without JavaScript: if 10% of theories are true, true ones succeed 95% of the time and false ones 10%, a successful theory is true with probability about 51%; if false theories succeed only 1% of the time, it rises to about 91%.

What counts as science? The demarcation problem

Popper’s falsifiability was a criterion of demarcation: a line between science and pseudo-science. Courts have wanted such a line. In McLean v. Arkansas (1982) Judge William Overton listed essential characteristics of science, including being guided by natural law and testable. In Kitzmiller v. Dover Area School District (2005) Judge John E. Jones III ruled that intelligent design is not science and cannot be taught as such in US public-school science classes. Philosophers have been less sure that one criterion works. Laudan (1983) called the problem dead: any proposed mark, falsifiability, testability, or being supported by evidence, either excludes real science or includes junk, since astrology makes testable claims that were tested and failed. Paul Thagard (1978) proposed instead that astrology is unscientific because it has been stagnant for centuries and its practitioners do not try to explain its failures, which is Lakatos’s degeneration under another name.

Most philosophers now treat science as a cluster of features, which family-resemble one another without any one being necessary: public evidence, error-correction, openness to criticism, a track record of novel predictions, and a community that checks its own work. This is less tidy than a single test. It also describes the working lesson better, which is that you judge a field by how it behaves when it is wrong. That leads straight to the live case.

A live case: the replication crisis

Every argument above is now being tested on psychology, medicine and economics. How Do We Know Anything? (05/16) tells the story and gives the main replication table; here are the numbers that bear on the philosophy. In 2015 the Reproducibility Project: Psychology tried to repeat 100 studies: 97 of the originals had significant results, 36% of the repeats did, and the effects were about half as large. In 2018 a project on 28 findings across 60 labs found about half failed. In a 2016 survey by Nature of 1,576 researchers, more than 70% had failed to reproduce another scientist’s results and more than half their own. In medicine, a 2012 report from the company Amgen said only 6 of 53 landmark preclinical cancer studies (11%) could be confirmed, and the Reproducibility Project: Cancer Biology (2021) found effect sizes about 85% smaller than the originals where experiments could be redone.

The philosophy shows in the diagnosis. John Ioannidis’s 2005 paper “Why Most Published Research Findings Are False” is Bayes at work: the chance that a significant result is true depends on how plausible the hypothesis was, on statistical power, and on how many ways the data were analysed. Duhem appears when failed replications are blamed on hidden moderators. Kuhn appears in the incentives: journals rewarded novel, positive results. In 2021 Anne Scheel and colleagues found 96% of hypotheses in a random sample of standard psychology papers came out positive, against 44% in Registered Reports, where the method is accepted before the data exist. Play with the base-rate arithmetic yourself.

Without JavaScript: if 10% of hypotheses are true, power is 35% and α is 0.05, then of 1,000 hypotheses about 35 true and 45 false ones come out significant, so only about 44% of significant results are true; with ten analyses tried per false hypothesis, only about 9% are.

What has changed since: preregistration, Registered Reports, larger samples, open data and code, and a research field, metascience, that studies science itself. The crisis is a reason to ask how a result was obtained, not to give up on science. In that it is Popper’s point in practice: a community that finds and publishes its own failures is behaving as one that can learn.

What a graduate just sees

• Induction cannot be proved, and “which predicate to generalise” (green or grue) is part of the problem.
• A white shoe does confirm “all ravens are black”, by a tiny amount; how much depends on how the evidence was sampled.
• A failed prediction blames a bundle, and the blame falls on whichever part you trusted least.
• A strict rule for dropping theories would have dropped many true ones.
• Anomalies are normal; the question is whether the programme is predicting new things or only patching.
• A theory’s success is evidence of truth only given a base rate of true theories, which is why realists stress novel predictions.
• There is no single mark of science; judge a field by how it handles being wrong.
• A significant result is a probability statement whose meaning depends on priors, power and how many analyses were run.

Where it shows up next

• Research Methods in Psychology (35/19) and Statistics (34/05): significance, power and preregistration are the working side of Bayes and Popper.
• Philosophy of Mind (36/05): the theories of consciousness are being tested by adversarial collaborations, a method straight from this debate.
• Contemporary Philosophy (36/12): social epistemology, feminist philosophy of science and the sociology of knowledge.
• In the world: evidence-based medicine and regulation, science journalism, climate and public-health communication, and court rulings on expert evidence.

If you want to go deeper

• 📘 Textbook. Peter Godfrey-Smith, Theory and Reality (University of Chicago Press, 2003), is the widely used survey; A. F. Chalmers, What Is This Thing Called Science? (Hackett, 4th edition 2013) is a clear classic.
• 📘 Primary texts. Popper, The Logic of Scientific Discovery; Kuhn, The Structure of Scientific Revolutions (50th anniversary edition 2012); Feyerabend, Against Method. Ioannidis (2005) is free to read online.
• 📘 Gentler. Samir Okasha, Philosophy of Science: A Very Short Introduction (Oxford, 2nd edition 2016); Stuart Ritchie, Science Fictions (2020), on the replication crisis.
• 🌐 Free text. The Stanford Encyclopedia of Philosophy entries on The Problem of Induction, Karl Popper, Imre Lakatos, Scientific Realism and Science and Pseudo-Science.
• 🌐 Free course. The University of Edinburgh’s Philosophy and the Sciences: Introduction to the Philosophy of Physical Sciences (Coursera, with Michela Massimi and colleagues) opens with scientific knowledge, method and underdetermination, then applies them to cosmology; check the enrolment options for a free route. MIT OpenCourseWare 24.09 Minds and Machines is a good companion on the mind side (36/05).