1 What makes a claim scientific?

The standard answer taught in most science classes is that science uses the scientific method: you observe, you form a hypothesis, you test it, you revise. This is not wrong, but it is not sufficient either. Astrology also makes claims about the world and invites empirical testing. Psychoanalysis also generates hypotheses and accumulates confirming evidence. What, precisely, distinguishes science from these?

1.1 Popper, Adler, and the Vienna Origin of the Demarcation Criterion

In Vienna in 1919, the young Karl Popper worked briefly with Alfred Adler and watched psychoanalytic theory account for a case of child psychology.⁵ Popper found that Adler’s theory could explain any human behaviour whatsoever: aggression confirmed the inferiority complex; generosity confirmed overcompensation for the inferiority complex. There was no observation that could, even in principle, count against it. This was the experience that led Popper to formulate the falsifiability criterion: a genuine scientific theory must be capable of being refuted by observable evidence. Psychoanalysis, he argued, was not science — not because it was false, but because it was structured to be unfalsifiable. The psychoanalytic establishment did not accept this characterisation, and the debate about whether psychoanalytic claims can be empirically tested has continued for decades. Seymour Fisher and Roger Greenberg, in a 1977 meta-analysis, concluded that several specific Freudian claims had testable implications and had received partial empirical support.⁶ The demarcation problem — identifying a principled criterion that separates science from non-science — is harder than Popper’s clean distinction suggests: Thomas Kuhn observed that scientists regularly ignore apparently falsifying evidence during periods of normal science. The 2005 Kitzmiller trial (info box above) is the institutional test of Popper’s criterion against contemporary intelligent-design literature; Judge Jones’s opinion is in effect a 139-page philosophical-scientific application of the criterion to specific empirical claims.

1.2 Popper’s Criterion

Karl Popper’s answer, developed in The Logic of Scientific Discovery (German 1934, English 1959) and Conjectures and Refutations (1963), is that falsifiability distinguishes science from non-science.⁷ A claim is scientific if and only if there is some possible observation that could show it to be false. The mark of science is not certainty but vulnerability to refutation.

Popper arrived at this criterion by noticing the asymmetry between confirmation and falsification. No number of confirming observations can prove a general claim — you cannot verify that all swans are white by observing a million white swans. But a single observation of a black swan proves the claim false. Science progresses by trying to prove its claims false; surviving severe attempts at falsification is the best evidence a theory can have.

Apply the criterion: Is astrology scientific? Astrological claims are highly flexible — an astrologist can interpret almost any outcome as consistent with a horoscope. If a claim can accommodate any possible observation, it predicts nothing, and Popper would say it is not scientific. Is Freudian psychoanalysis scientific? Freud’s theory of the unconscious can explain any patient behaviour — aggression and submission, sexuality and its suppression — because it has concepts that can be applied in both directions. Popper saw this not as evidence of the theory’s power but as evidence of its unfalsifiability. Is string theory scientific? This is a live and serious debate: string theory has been enormously productive mathematically, but its predictions (extra dimensions at the Planck scale, supersymmetric particles) may be permanently beyond the reach of any conceivable experiment.

1.3 The Demarcation Problem

Popper’s criterion is elegant and influential, but philosophers of science have found it increasingly difficult to apply cleanly.

Auxiliary hypotheses. Scientific theories do not confront evidence alone; they confront it together with a package of auxiliary assumptions, instrument calibrations, and background theories. When a prediction fails, you don’t know whether the core theory is wrong or whether one of the auxiliaries is wrong. Uranus was discovered in 1781, and over the following decades its observed orbit drifted from the predictions of Newtonian mechanics. Did this falsify Newton? No — Urbain Le Verrier computed the position of an unseen perturbing planet, and on 23 September 1846 Johann Galle observed Neptune at the Berlin Observatory within a degree of the predicted location.⁹ The anomaly that might have falsified the theory instead supported an extension of it. This is Imre Lakatos’s point: scientists work within “research programmes” with a “hard core” they protect from falsification and a “protective belt” of auxiliary hypotheses they adjust.¹⁰

Kuhn’s Reply. Thomas Kuhn, in The Structure of Scientific Revolutions (1962), argues that Popper’s account bears no resemblance to how science actually works.¹¹ Real scientists do not attempt to falsify their theories. When anomalies arise — observations that don’t fit the theory — scientists typically assume the problem lies with the experiment, the observer, or some peripheral assumption. They do not abandon the theory. This is not irrationality; it is the only way science can function. A theory that was abandoned at the first anomaly would never survive long enough to be tested properly.

1.4 What Distinguishes Science, Then?

If falsifiability is too simple a criterion, we need a more complex answer. Several features working together distinguish science from non-science, without any single one being individually sufficient:

• Empirical content: scientific claims make specific, testable predictions about the observable world.
• Internal coherence: scientific theories must be internally consistent.
• Communal scrutiny: scientific claims are subjected to peer review, replication attempts, and criticism by a community of informed sceptics.
• Self-correction: when evidence against a theory accumulates sufficiently, the theory is revised or abandoned. This happens slowly and imperfectly — but it happens.
• Progressive extension: good scientific theories don’t just explain what they were designed to explain; they successfully extend to new domains (Newtonian mechanics applied to tides, to the Moon, to projectile trajectories simultaneously).

None of these is perfectly sharp, and the boundary between science and non-science is genuinely fuzzy in some cases. This is the demarcation problem. But fuzziness at the boundary doesn’t mean the distinction is meaningless: the centre of the category is clear even if the edges are not.

1.5 Questions to Argue About

• Is string theory science? It makes predictions, but they may be permanently untestable. Does this matter?
• Popper says psychoanalysis is not scientific because it cannot be falsified. Does that mean Freud was wrong? Or can a non-scientific theory be true?
• Kuhn argues that scientists don’t actually try to falsify their theories. Does this undermine Popper’s criterion, or does it just show that scientists don’t always practise what philosophers of science preach?
• If there is no sharp boundary between science and non-science, what are the practical implications? Does it matter whether climate science, economics, or nutritional research count as “really scientific”?

2 What is the scientific method — and does science actually follow it?

Every school science curriculum teaches the scientific method: observe, hypothesize, experiment, conclude. This story is clean, linear, and rational. Every scientist learns it. Almost no scientist actually uses it in this sequence — and several philosophers of science have made distinguished careers from pointing this out.¹² The story is, as a description of how science is actually done, substantially false — or at least radically incomplete.

2.1 Bacon’s Inductivism

Francis Bacon (Novum Organum, 1620) proposed what became the standard picture: knowledge advances by the patient accumulation of observations, from which general laws are induced. Collect enough data, without prejudice, and nature’s laws will emerge. This is inductivism: from particular observations to general conclusions.¹⁴

The logical objection is Hume’s problem of induction: no number of particular observations can logically entail a general law. This is not merely a logical technicality; it means that scientific knowledge is always provisional, however massive the evidence base.

The practical objection is that pure inductivism doesn’t describe how science works. Scientists don’t observe without prejudice; they observe through the lens of existing theory, which tells them what to observe, how to measure it, and what counts as a relevant result. As Hanson put it, observation is “theory-laden.”¹⁵ Bacon’s ideal of the blank-mind observer is impossible and, if it were possible, would produce meaningless data.

2.2 The Hypothetico-Deductive Model

The more sophisticated account is hypothetico-deductive: the scientist starts with a hypothesis (often creatively generated, not induced from observations), deduces what observations the hypothesis would predict, and tests those predictions. This is closer to what Einstein and Darwin actually did — they had theoretical ideas first and looked for evidence to test them.

But this model still produces an idealised picture. The step from “hypothesis” to “test” is not a simple logical deduction. Experiments must be designed, which requires decisions about what variables to control, what instruments to use, and what counts as a significant result. These decisions embed assumptions. The famous example: Millikan’s oil-drop experiments (1909–1913) to measure the electron’s charge involved a process of selection — he discarded data that didn’t fit what he expected. This is neither fraud nor good practice unambiguously; it is the messy reality of how knowledge gets made.

2.3 Feyerabend and “Anything Goes”

Paul Feyerabend, in Against Method (1975), pushed the critique to its extreme. There is no single scientific method; the history of science shows that progress has been achieved by violating any methodological rule you care to name. Galileo’s Dialogue Concerning the Two Chief World Systems (1632) is Feyerabend’s central exhibit: Galileo defended Copernicanism in a literary dialogue between Salviati (the Copernican spokesman), Sagredo (the receptive layman), and Simplicio (whose name punningly marks him as the simpleton); his strongest physical argument for the Earth’s motion — that the tides are caused by the daily and annual motions sloshing the oceans — was wrong, while several apparently strong contemporary objections to Copernicanism (the absence of detectable stellar parallax; the lack of a felt wind from the Earth’s rotation) were on the available evidence reasonable. Galileo, on Feyerabend’s reading, used propaganda and psychology as well as observation; he was sometimes wrong about the data while being right about the theory. The rule “believe what the observations show” would have prevented the Copernican revolution.¹⁷

Feyerabend’s conclusion is deliberately provocative:

“the only principle that does not inhibit progress is: anything goes.” — Paul Feyerabend, Against Method, Introduction.¹⁸

He is not saying science is irrational. He is saying that the rationality of science cannot be captured in any neat set of rules.

2.4 How Kepler Actually Worked

Johannes Kepler’s discovery of elliptical planetary orbits (published in Astronomia Nova, 1609) is a case study in how science is actually done versus how it is described. Kepler spent eight years trying to make Tycho Brahe’s precise observational data fit circular orbits — the traditional Platonic assumption that the heavens required perfect circular motion. When the data consistently showed an 8-arc-minute discrepancy from circular predictions (a discrepancy he knew was too large to be observational error, because Brahe’s instruments were far more accurate than their predecessors), he was forced to abandon the circle and try the ellipse.¹⁹

Kepler’s path was neither pure induction (he had a prior commitment to circle-perfect heavens that he was reluctant to abandon) nor clean hypothetico-deduction (the “hypothesis” of elliptical orbits came from trying things that fit, not from a theoretical framework that predicted ellipses). It was stubbornness, creativity, data-respect, and eventually the willingness to overturn centuries of assumption — a process that looks nothing like the textbook method.

2.5 What Counts as an Explanation? Hempel’s Deductive-Nomological Model

Method gets you to a hypothesis that survives testing. But suppose it does survive — what kind of answer have you given? Ignaz Semmelweis, working in the Vienna General Hospital between 1844 and 1848, was confronted with a puzzle: the First Maternity Division, staffed by physicians and medical students, had a maternal mortality from childbed fever of around 18% in April 1847; the Second Division, staffed by midwives, ran at roughly 2%.²⁰ Semmelweis’s eventual hypothesis was that the physicians, who arrived from morning autopsies, were carrying what he called “cadaverous particles” on their hands and introducing them into wounds during examination. He instituted handwashing in chlorinated lime in mid-May 1847; mortality in the First Division dropped to 1.2% by July. In what sense had he explained childbed fever?

Carl Hempel, writing in 1966, argued that the structure of every genuine scientific explanation is the same. An explanation is a deductive argument: its premises (the explanans) must include at least one general law together with statements of antecedent conditions, and its conclusion (the explanandum) must follow logically. Hempel’s own term for the pattern is canonical:

“Explanatory accounts of this kind will be called explanations by deductive subsumption under general laws, or deductive-nomological explanations. (The root of the term ‘nomological’ is the Greek word ‘nomos’, for law.) The laws invoked in a scientific explanation will also be called covering laws for the explanandum phenomenon.” — Carl G. Hempel, Philosophy of Natural Science (1966), Ch. 5.²¹

Hempel opens the book with Semmelweis precisely because the case fits the pattern so cleanly. Semmelweis’s eventual diagnosis — that childbed fever was caused by “decomposed animal matter” (his “cadaverous particles”) carried from the dissection room into wound surfaces — implicitly invokes the general law that introduction of such matter into the bloodstream produces blood poisoning, plus the particular fact that medical students went directly from cadavers to the maternity ward. Together they entail the high mortality in the First Division.²² The diagnosis was correct in its operative content (handwashing prevents transmission) but its proposed mechanism was pre-bacterial: Semmelweis had no germ theory and could not say what the “particles” actually were. The Vienna medical establishment rejected his findings, his contract was not renewed, and he died in a Viennese asylum in August 1865 — beaten by guards and infected through a wound on his hand — two decades before Pasteur and Lister vindicated the underlying claim by way of bacteriology.²³

The DN model is elegant. It is also too liberal. Samir Okasha’s standard counter-example: a flagpole of known height casts a shadow of calculable length, and from that shadow length (with the laws of optics and the sun’s elevation) we can deduce the flagpole’s height. The deduction is valid; it satisfies Hempel’s schema; and it gets the explanatory direction backwards. The flagpole’s height explains the shadow’s length; the shadow does not explain the flagpole.

“Hempel’s covering law model does not respect this asymmetry. For just as we can deduce the length of the shadow from the height of the flagpole, given the laws and additional facts, so we can deduce the height of the flagpole from the length of the shadow … So Hempel’s model fails to capture fully what it is to be a scientific explanation.” — Samir Okasha, Philosophy of Science: A Very Short Introduction (2002), Ch. 3.²⁴

The flagpole case suggests that scientific explanation involves something the DN model omits — most plausibly causal direction, which deductive entailment cannot capture. The alternative often advanced in its place is inference to the best explanation (IBE): we should believe the hypothesis that, if true, would best explain the evidence. IBE has its own difficulties — “best of which available alternatives?” — but it tracks the asymmetry the flagpole exposes.

The deeper point: even when method has done its work, a separate philosophical question remains about what an explanation is, and the cleanest available answer turns out to be insufficient.

2.6 Questions to Argue About

• Feyerabend says “anything goes” in science. Is this a defence of scientific freedom or an abdication of rational standards?
• If scientific observations are always theory-laden, does that mean we can never get an objective test of a theory? Or does “theory-laden” just mean “informed by background knowledge,” which seems unavoidable and not necessarily distorting?
• Millikan discarded inconvenient data points. In hindsight, his final value for the electron’s charge was right. Does this vindicate his data selection, or does it just show that in this case he got lucky?
• The textbook “scientific method” is acknowledged by philosophers of science to be an oversimplification. Should schools continue to teach it, and if so, why?

3 How do scientific models relate to reality?

A map of London is not London. It does not contain the smell of the Underground, the particular grey of the Thames on a November afternoon, or the experience of being stuck behind a tourist who stops on the escalator. It is a representation — simplified, selective, oriented toward a purpose. A scientific theory is not a direct description of reality either. It is a model — a simplified, formalised representation that captures some aspects of what is being studied while necessarily abstracting away from others. The question is what we mean when we say it is true. The relationship between the model and the reality is one of the deepest questions in the philosophy of science.

3.1 Box’s Aphorism

The statistician George Box wrote in 1976:

“all models are wrong but some are useful.” — George E. P. Box, “Science and Statistics,” Journal of the American Statistical Association (1976).²⁶

This aphorism has become a touchstone of scientific self-understanding, and its simplicity conceals real depth. “Wrong” in Box’s sentence is doing more work than the surface reading allows: Box was a statistician writing about idealised models of complex systems for forecasting, where “wrong” means strictly false as a description of the data-generating process, not “approximately incorrect.” Cartwright (How the Laws of Physics Lie, 1983) extends the same point to the fundamental laws of physics; van Fraassen (The Scientific Image, 1980) directs it at the unobservable entities those laws postulate.²⁷

The operative question is therefore not “is the model wrong?” — they are all wrong in Box’s strict sense — but “useful for what?” Useful for prediction, for explanation, for control, and for pedagogy can come apart.

A weather model is not the weather; it is a mathematical system that represents certain features of atmospheric dynamics (pressure, temperature, humidity, wind) at a certain level of resolution. It is “wrong” in the sense that it omits an enormous amount — every butterfly, every microclimatic variation, every human exhaled breath. But it is enormously useful: it predicts tomorrow’s weather with an accuracy that was impossible before numerical weather prediction. The question is not whether it is true but whether it is accurate enough for the purpose at hand.

The same applies to Newton’s laws of motion: they describe the motion of everyday objects with extraordinary accuracy. They are also wrong — or rather, they are an approximation that breaks down near the speed of light (where special relativity takes over) and at very small scales (where quantum mechanics takes over). For most engineering purposes, however, Newton is not just “good enough”; it is the right level of description for the phenomena being modelled.

3.2 The Periodic Table as Model

The periodic table of elements, organised by Dmitri Mendeleev in 1869, is one of the most successful models in the history of science. It organises 118 elements by atomic number and chemical properties; it reveals periodic patterns in reactivity, electronegativity, and ionisation energy; it predicted the existence of then-undiscovered elements — Mendeleev’s eka-aluminium, eka-boron, and eka-silicon, later identified as gallium (1875), scandium (1879), and germanium (1886) — which were subsequently found with precisely the properties the table predicted.²⁸

What does the periodic table model? The atomic structure of matter — but the table itself contains no explicit statement about atomic structure. It is an organisation of empirical regularities that turned out to reflect underlying physical reality (the filling of electron shells). The model preceded the explanation of why it worked.

3.3 Lavoisier and the Overthrow of Phlogiston

The textbook example of one chemical model giving way to a successor is the late-eighteenth-century overthrow of the phlogiston theory by Antoine Lavoisier’s oxygen chemistry. Phlogiston, on the account systematised by Becher (1669) and Stahl (1703), was a substance released from combustible bodies during burning: wood released phlogiston into the air, leaving a calx (ash) behind; metals released phlogiston when heated, leaving the metal calx; the air absorbed phlogiston until saturated, at which point combustion stopped. The theory was productive: it organised chemistry for nearly a century, made successful predictions about which substances would burn, and unified phenomena that the ancient four-element scheme had treated as unconnected.³⁰

Lavoisier’s Traité élémentaire de chimie (1789) overturned it. Lavoisier had shown, by careful weighing — a methodological move that pre-Lavoisierian chemistry had not made central — that combustion added mass to the burning substance rather than removing it. If phlogiston was released, the residue should weigh less; in fact, combusted metals weighed more. Lavoisier proposed that combustion was a chemical combination with a constituent of the air (which he named oxygène, “acid-forming”); that respiration was the same process; that water was a compound, not an element; and that mass was conserved across all chemical reactions. The 1789 Traité is the founding text of modern chemistry.³¹

The case is philosophically rich. It is the cleanest example of a Kuhnian paradigm shift in chemistry: the same experimental data (calxes weigh more than the metals they came from) had been available for fifty years before Lavoisier; what shifted was the theoretical framework that made the data significant. It is also the case Larry Laudan repeatedly cites against scientific realism in his “pessimistic meta-induction”: phlogiston was empirically successful, predictively useful, and eventually had to be abandoned as referring to nothing — and there is no general reason to think the theoretical posits of contemporary chemistry will fare any better in two centuries. Whether that is a fatal worry for chemical realism or a salutary humility about it is the open question.

3.4 Dark Matter: Model Postulated to Save Theory

Dark matter is one of the most striking examples of a theoretical entity postulated not because it was observed but because existing models fail without it. Observations of galaxy rotation rates in the 1970s — beginning with Vera Rubin and Kent Ford’s 1970 Andromeda paper, and consolidated in their 1978–1980 flat-rotation studies of other spirals — showed that galaxies rotate too fast: the outer stars move as if there is far more mass present than is visible.³² The standard response was to postulate an invisible form of matter — “dark matter” — that does not emit or absorb light but exerts gravitational effects.

This postulation has a good precedent: Neptune was similarly postulated to explain anomalies in Uranus’s orbit before it was observed. Dark matter has accumulated strong indirect evidence (gravitational lensing, cosmic microwave background data). But as of 2025, no particle detector has directly detected a dark matter particle. Candidate particles like WIMPs (Weakly Interacting Massive Particles) have been ruled out across most of their predicted parameter space; alternatives like axions are being tested.³³

The question is which way the situation reads. Defenders of dark matter as a scientific hypothesis emphasise the convergent indirect evidence (rotation curves, lensing, the CMB peaks, the Bullet Cluster), the specificity of the predictions, and the fact that the candidate particles are being actively ruled in or out by experiment. Critics — those favouring modified-gravity alternatives like MOND, or those reading dark matter as a Lakatosian auxiliary protecting the standard cosmological model from falsification — emphasise that dark matter has been the auxiliary postulate for 90 years without direct detection and that the parameter space for the originally favoured candidates has been progressively eliminated. The Forced Fork below puts the question to the reader.

3.5 Scientific Realism vs. Instrumentalism

The deepest question about models is whether they are representations of reality or merely instruments for prediction:

Scientific realism: the entities postulated by successful scientific theories really exist. Electrons, quarks, dark matter, and the curvature of spacetime are real things, not just calculational conveniences. The predictive success of science is best explained by assuming that our theories are approximately true.

Instrumentalism: scientific theories are neither true nor false; they are tools. We should evaluate them solely by their predictive success. Whether electrons “really exist” is a meaningless question; what matters is whether the electron model correctly predicts the results of experiments.

Bas van Fraassen’s constructive empiricism (The Scientific Image, 1980) offers a middle position: we should believe that the observable consequences of successful theories are true, but remain agnostic about unobservable entities like electrons.³⁴ This is epistemically cautious but creates strange consequences: it implies we should believe that atoms in a container exert pressure (observable) but remain agnostic about whether the atoms exist (unobservable).

Nancy Cartwright presses the anti-realist case from the other direction. Where van Fraassen targets the unobservable entities, Cartwright (How the Laws of Physics Lie, 1983) targets the equations themselves: “the fundamental laws of physics do not describe true facts about reality. Rendered as descriptions of facts, they are false; amended to be true, they lose their fundamental, explanatory force.”³⁵ Newton’s law of universal gravitation, on Cartwright’s analysis, is true of objects in a model and not of any actual planet, which is also subject to electromagnetic forces, atmospheric drag, and other effects the law abstracts from. The phenomenological laws used by engineers are closer to reality but less general; the more universal a law looks, the less literally true it is in any concrete situation.

3.6 Questions to Argue About

• If all scientific models are wrong (Box), does that mean science never achieves truth? Or does “wrong” in this context mean something different from “false”?
• Dark matter is postulated to explain anomalies but has never been directly detected. Is this good scientific practice (like the Neptune case) or special pleading to protect a theory?
• Scientific realists say electrons really exist because the electron model predicts observable results so well. But good predictive success is compatible with the model being false. Does predictive success give us reason to believe in the existence of unobservable entities?
• Hacking argues that we know electrons exist because we can manipulate them to produce other effects. Does this “manipulation” argument for scientific realism work?

4 Can science be objective?

The ideal of scientific objectivity is that the method — observation, experiment, analysis — produces results that are independent of the observer’s identity, values, and expectations. It is a powerful ideal, and the institutions built on it have produced reliable knowledge across many domains. The evidence from both philosophy and the history of science complicates it. The question is whether the complications require abandoning the ideal, weakening it, or refining what “objectivity” means in scientific practice — and reasonable people who agree on the evidence disagree about the answer.

4.1 Theory-Ladenness of Observation

Norwood Russell Hanson, in Patterns of Discovery (1958), argues that what we see depends on what we already know.⁴² He asks: what does a naive observer see when they watch the sun rise? What does a Keplerian astronomer (who knows the Earth orbits the sun) see when they watch the same event?

The naive observer sees the sun rising. The Keplerian sees the horizon tilting away from a stationary sun. The same visual input generates different perceptions because the conceptual framework within which observation is interpreted differs. This is not a matter of what the observer believes while looking; it is a matter of what they see.

The implications for scientific objectivity are uncomfortable: if all observation is theory-laden, then no observation can serve as a neutral test of a theory, because the observation is already conditioned by theoretical assumptions. Scientists see what their paradigm trains them to see — and may systematically fail to see what their paradigm says cannot be there.

4.2 Heisenberg’s Uncertainty Principle

At the quantum level, certain pairs of properties cannot both be sharply defined at the same time. The Heisenberg uncertainty principle (1927) is usually written \(\Delta x \cdot \Delta p \geq \hbar/2\): the product of the standard deviations of position and momentum in any quantum state has an irreducible lower bound. This is not a statement that instruments disturb particles (the popular “observer-effect” gloss), and it is not a statement about the accuracy of individual measurements — it is a statement about the intrinsic non-commutativity of the quantum operators themselves.⁴³ It is a fundamental feature of quantum mechanics rather than a limitation of current technology.

This has been taken by some philosophers to imply that the observer is necessarily implicated in what is observed at the quantum level — that there is no quantum world independent of measurement. This interpretation (associated with the Copenhagen interpretation of quantum mechanics, developed by Bohr and Heisenberg) is controversial; other interpretations (the Everett “many worlds” interpretation, for example) deny that measurement is special. But the question of what quantum mechanics implies about the observer-independent character of reality remains genuinely open.

4.3 The Replication Crisis

Starting around 2011, psychology began experiencing what is now called the replication crisis: a systematic failure to reproduce the results of landmark published studies. The Open Science Collaboration (2015) attempted to replicate 100 published psychology experiments; only 39% of the replications were judged by the replicating teams themselves to have reproduced the original result, and on a stricter measure — statistically significant effects in the same direction — the figure was 36%.⁴⁴

The implications for scientific objectivity: if a substantial proportion of published results cannot be reproduced, what does this say about the knowledge claims those studies generate? The problem is partly statistical (underpowered studies, p-hacking, flexible analysis), partly incentive-driven (journals publish novel positive results, not replications), and partly about the gap between the clean story told in published papers and the messy, iterative process that actually produced the data.

4.4 Feminist Critiques

Sandra Harding, in Whose Science? Whose Knowledge? (1991), argues for what she calls “strong objectivity” — a methodological demand that goes beyond conventional objectivity.⁴⁶ Standard objectivity demands that researchers not let their values distort their research. Strong objectivity demands that researchers examine the values and social positions that shape which questions are asked, which methods are used, and which phenomena are considered worth studying.

Strong objectivity is not a demand that science be partisan; it is a demand that the social conditions of knowledge production be made part of the analysis. Historical examples: the biological sciences’ long history of providing “natural” explanations for social hierarchies (women’s “naturally” limited intellectual capacity, racial differences in intelligence) turned out to reflect the social assumptions of the scientists, not biological reality.⁴⁷ The objectivity that would have exposed this was precisely the kind that examined the assumptions the science brought to its questions.

Harding’s argument is sometimes assimilated to a stronger constructionist claim — that scientific findings are socially constructed in the sense of being merely contingent products of their social setting. Ian Hacking’s The Social Construction of What? (1999) is sharp on this conflation: a “social construction of X” claim, on Hacking’s analysis, characteristically packs together three theses (X is not inevitable; X is bad as it is; we would be better off without X), and only the first is shared with Harding’s methodological demand.⁴⁸

4.5 Inductive Risk and the Value-Free Ideal

A separate question is whether non-epistemic values — social, ethical, political — have any legitimate role inside scientific reasoning, not merely in selecting topics or interpreting results for policy. Heather Douglas, in Science, Policy, and the Value-Free Ideal (2009), argues that they do, and that the standard ideal of value-free science must be rejected:

“the value-free ideal must be rejected precisely because of the importance of science in policymaking. In place of the value-free ideal, I articulate a new ideal for science, one that accepts a pervasive role for social and ethical values in scientific reasoning, but one that still protects the integrity of science.” — Heather Douglas, Science, Policy, and the Value-Free Ideal, Ch. 1.⁴⁹

Douglas’s central technical move revives the inductive risk argument of Richard Rudner (1953): because no scientific hypothesis is ever conclusively verified, the scientist who decides to accept or reject one must weigh how bad it would be to be wrong in each direction — a weighing that requires non-epistemic values whenever the claim has practical consequences. Setting the p-value threshold at 0.05 rather than 0.01, choosing how strong evidence of carcinogenicity must be before reporting a chemical as carcinogenic, deciding whether to model uncertainty conservatively or optimistically: each of these decisions has, on Douglas’s account, a value-laden dimension that the value-free ideal misdescribes when it treats the decision as purely epistemic. Whether her diagnosis applies to all cases or only to the regulatory-science cases she chiefly draws on is itself contested in the literature.⁵⁰

4.6 Questions to Argue About

• If observation is theory-laden, can any observation provide a genuinely independent test of a theory? Or is this a problem that can be partially solved by using theories from different domains to constrain each other?
• Does Heisenberg’s uncertainty principle show that the observer is necessarily part of quantum reality — or is it just a limitation on measurement?
• The replication crisis shows that a substantial proportion of published psychological findings cannot be reproduced. Does this mean psychology is not a science? Or does it just mean science has a quality control problem?
• Harding argues for “strong objectivity” that examines the social assumptions of science. Is this a strengthening of objectivity, or is it a political intervention that threatens objectivity?

5 How does science change?

Science is often presented as accumulating knowledge steadily, building toward a more complete and accurate picture of the world. Thomas Kuhn showed that this picture, while not entirely wrong, conceals the more dramatic and disruptive process by which science actually changes.

5.1 Kuhn’s Paradigms

In The Structure of Scientific Revolutions (1962), Kuhn introduces the concept of a paradigm — a framework of shared assumptions, methods, and problems that defines what “normal science” is doing at any time.⁵⁴ Normal science is puzzle-solving within the paradigm: it does not question fundamental assumptions but fills in details, resolves inconsistencies, and extends the theory’s scope.

In the 1969 Postscript Kuhn admitted that “paradigm” had been doing two distinct jobs in the original book and offered a sharper vocabulary: a disciplinary matrix (the whole bundle of values, methods, symbolic generalisations, and shared exemplars binding a scientific community) and a concrete exemplar (a worked-out problem-solution that trains scientists how to handle similar problems).⁵⁵ Most lay use of “paradigm shift” elides this distinction.

Plate tectonics is a clean example of a disciplinary-matrix shift: an entire community’s organising framework changed in roughly a decade. The Copernican revolution is messier — a multi-stage process more easily described in terms of successive exemplars (Copernicus, Kepler, Galileo, Newton) than as a single matrix-replacement.

Anomalies — observations that cannot be accommodated within the paradigm — accumulate. At first, they are ignored or explained away. When they become impossible to ignore, a scientific revolution occurs: the old paradigm is replaced by a new one. The Copernican revolution replaced the Ptolemaic geocentric model. The germ theory of disease replaced miasma theory. The plate tectonics revolution replaced the “permanentist” view of stable continents.

Crucially, Kuhn argues that different paradigms are incommensurable: scientists working in different paradigms are not just disagreeing about facts, they are working in different conceptual worlds, asking different questions, seeing different phenomena. This is why paradigm shifts are so difficult and so slow: the scientists defending the old paradigm are not being irrational — they are operating within a framework that, for them, still works.

5.2 Continental Drift

Alfred Wegener, a German meteorologist, proposed in The Origin of Continents and Oceans (1915) that the continents had once been joined and had drifted apart.⁵⁶ His evidence was substantial: the shapes of Africa and South America fit together like torn pieces of a map; identical fossil species appeared on both sides of the Atlantic; geological strata matched across coastlines thousands of miles apart. The theory was rejected, sometimes contemptuously, by the geological establishment for more than four decades. At a 1926 symposium of the American Association of Petroleum Geologists convened specifically to evaluate Wegener’s hypothesis, the geologist Rollin T. Chamberlin argued that if Wegener’s theory were accepted, geologists would have to “forget everything which has been learned in the last 70 years and start all over again.”⁵⁷

The rejection was not entirely irrational. Wegener proposed no plausible mechanism for how continents could move through oceanic crust — his own tentative suggestions (centrifugal force from the Earth’s rotation; tidal drag from the Sun and Moon) were quantitatively far too weak to do the work, and his critics knew this. By the standards of physics available at the time, the idea seemed physically impossible. It was only in the 1950s and 1960s, when the discovery of mid-ocean ridges and paleomagnetism provided both evidence and a different mechanism (seafloor spreading driven by mantle convection — not Wegener’s continents-through-crust picture), that plate tectonics was accepted — and rapidly became the organising paradigm of the earth sciences.⁵⁸

A more sceptical reading of this story is worth noting. Naomi Oreskes (The Rejection of Continental Drift, 1999) argues that the standard “Wegener was vindicated” narrative is partly a retrospective reconstruction: the modern theory of plate tectonics differs in important respects from what Wegener actually proposed (the mechanism, the timescale, several specific claims about geological correlations).⁵⁹ On Oreskes’s reading the narrower lesson is that the geological community resisted a theoretical framework with substantial evidence in its favour for longer than the evidence alone warranted, and that the resistance was loosened only when an independent line of evidence (paleomagnetism) and a workable mechanism (seafloor spreading) became available together. On the older reading the narrative survives intact: Wegener was substantively right; the community took fifty years to admit it. Which reading is right matters for what the case shows about scientific resistance to outsider theories.

5.3 Maxwell, Higgs, and the 48-Year Gap

The Higgs case (info box above) is the contemporary version of a phenomenon Eugene Wigner identified in 1960 as “the unreasonable effectiveness of mathematics in the natural sciences.” Its canonical historical version is Maxwell. In 1865, the Scottish physicist James Clerk Maxwell published a unified mathematical theory of electricity and magnetism, in which he derived as a consequence of the equations themselves that electromagnetic disturbances should propagate through space at approximately the speed of light — already known to be roughly 300,000 km/s.⁶⁰ Maxwell concluded that light was an electromagnetic wave, and his equations predicted the existence of electromagnetic waves at other frequencies. Twenty-two years later, in 1887, Heinrich Hertz confirmed the existence of what are now called radio waves, working entirely from Maxwell’s mathematical prediction.⁶¹ The radio, television, mobile telephone, and wireless internet are the practical descendants of a set of mathematical symbols manipulated by Maxwell at King’s College London during his professorship there from 1860 to 1865.

The Higgs interval (48 years) and the Maxwell interval (22 years) raise a question about what counts as knowledge during the gap. Did physicists know the Higgs boson existed before 2012, on the strength of the theoretical prediction? They had strong indirect grounds: electroweak precision fits at LEP and the Tevatron narrowed the allowed mass window, and the Standard Model without some such mechanism was internally inconsistent at high energies. This is not nothing. But it is also not the same epistemic state as a 5σ direct detection. The question of what to call the intermediate state — between a successful mathematical prediction and its experimental confirmation — is a permanent feature of physics on Wigner’s reading and a temporary inconvenience of measurement on the deflationary alternative. The philosophical question — whether the fit between mathematics and physics is a deep fact about the universe or a selection effect in which we remember the mathematics that worked and forget what did not — remains genuinely open.

5.4 Why Scientists Resist New Ideas

If paradigm shifts are how science advances, why do scientists resist them? Several mechanisms:

Professional investment. A career spent on one theoretical framework is threatened by its replacement. This is not cynical self-interest; it is a rational response to epistemic uncertainty. You believe your framework; you’ve built your understanding of the field within it.

The rationality of conservatism. New theories typically explain less, at first, than the established ones. The new paradigm often works well in the narrow domain that motivated it and badly everywhere else. It takes time for a new theory to prove itself adequate to the full range of phenomena.

Incommensurability. Scientists trained in one paradigm may genuinely not understand what the new paradigm is claiming, because the concepts are different. Communication between paradigms requires translation, which is always approximate.

5.5 Questions to Argue About

• Kuhn’s incommensurability thesis implies that scientists in different paradigms can’t really communicate with each other. Is this too strong? Are there historical cases of genuine communication across paradigm shifts?
• The Wegener case suggests that correct theories can be rationally rejected. Does this mean the sociology of scientific communities is as important as the logic of scientific evidence in determining what science accepts?
• Scientists resisted the germ theory of disease and continental drift, both of which turned out to be correct. How should this history affect our attitude toward current scientific consensus? Does it give grounds for scepticism — or just caution?
• The Higgs boson was predicted in 1964 and confirmed in 2012. During those 48 years, did physicists “know” it existed? What would it mean to say they did?

6 What can science explain — and what can’t it?

Science is the most powerful method for producing knowledge about the natural world. That claim would have struck most educated people in the 17th century as hubristic. It strikes most educated people today as obvious. That change — from hubristic to obvious — is itself one of the most consequential intellectual events in human history. But even standing at the summit of that achievement, there is a serious question about what “the natural world” includes, and whether there are genuine domains of inquiry that scientific methods cannot reach.

6.1 Reductionism and Its Limits

Reductionism is the view that complex phenomena can be fully explained by their component parts. Chemistry is explained by physics (atoms and their interactions); biology is explained by chemistry; psychology is explained by neuroscience; sociology is explained by psychology. In principle, if you knew enough physics, you could explain everything.

The reductionist programme has had extraordinary successes. The discovery that DNA encodes genetic information in sequences of four bases — and that this code can be read, edited, and engineered — is a triumph of biochemical reductionism that has transformed medicine.

Emergence is the counter-claim: that complex systems have properties that are not predictable from, and cannot be reduced to, the properties of their components. The “wetness” of water is not a property of individual water molecules; it emerges from the interaction of vast numbers of them. The behaviour of an ant colony cannot be predicted from the chemistry of individual ants; it is a collective property of the interaction. Ernst Mayr argued that biology is irreducibly autonomous: biological concepts like fitness, adaptation, and selection operate at a level that cannot be reduced without loss of explanatory power.⁶⁶

The reductionist response distinguishes explanatory irreducibility (we cannot in practice predict colony behaviour from ant chemistry) from ontological irreducibility (there is something present in the colony beyond the ants and their interactions). On this reading the first is uncontroversial — a reflection of computational limits and the usefulness of higher-level concepts — and the second would require evidence of physical effects unaccounted for by lower-level laws. Wetness, in this reading, is just the macroscopic behaviour of large numbers of water molecules under standard conditions; no ontologically extra “wetness” is doing causal work. Strong-emergence defenders (Jaegwon Kim’s later work on downward causation, certain biologists working on developmental systems, some philosophers of mind) reply that the reductionist begs the question by demanding lower-level evidence for irreducibility — and that the difficulty of providing such evidence is precisely what one would expect if the higher-level kinds were genuinely real. The dispute is genuinely live within philosophy of mind and biology, less so within physics; how widely either side of the dispute extends across the sciences depends on the science.

6.2 The Hard Problem of Consciousness

David Chalmers, in The Conscious Mind (1996), distinguishes the “easy problems” of consciousness (explaining how the brain processes information, controls behaviour, reports its states — “easy” because they are tractable by standard scientific methods, even if technically difficult) from the “hard problem”: why is there subjective experience at all?⁶⁷

The hard problem asks: why, when information is processed in the brain, is there something it is like to be the brain processing that information? Why doesn’t neural processing occur “in the dark,” without any accompanying inner experience? This question resists scientific explanation not because we lack data but because the standard scientific vocabulary — causes, mechanisms, functions — seems inadequate to it. Even a complete functional account of how the brain produces colour perception would leave open the question: why does red look like this?

The most mathematically developed response is Giulio Tononi’s Integrated Information Theory (IIT). Tononi proposes that consciousness is identical to integrated information (Φ): a measure of how much a system’s cause-effect structure is irreducible to the cause-effect structure of its parts. The technical move: cut the system across its weakest link (the partition that destroys the least information) and see how much information is lost; a system with high Φ is one whose causal powers cannot be recovered from any decomposition.⁶⁸ IIT is mathematically precise and makes testable predictions, but it implies that simple networks of logic gates would be conscious — which most working scientists find counterintuitive enough to count against the theory. Two other frameworks are worth knowing: Global Workspace Theory (Baars, Dehaene) proposes that consciousness arises when information is broadcast to a workspace accessible to multiple cognitive systems — it explains access but says nothing about why the broadcast is experienced.⁶⁹Predictive processing (Clark, Friston) treats the brain as a prediction machine constantly minimising error against its model of the world.⁷⁰ None has solved the hard problem.

Thomas Nagel’s famous formulation, in “What Is It Like to Be a Bat?” (Philosophical Review, October 1974), is the canonical statement of the limit. Bats navigate their environment through echolocation — emitting ultrasonic pulses and processing the returning echoes to construct a spatial map of striking precision. We know, in extraordinary scientific detail, how the system works: the frequency modulation of the calls, the neural processing, the cochlear mechanics. What Nagel argued is that this scientific knowledge leaves something entirely untouched: the subjective character of the bat’s experience.⁷¹ What is it like, from the inside, to perceive a moth through sound? There is, Nagel argued, something it is like to be a bat — there is a subjective experience — and no amount of neuroscientific information about its echolocation system will tell us what that experience is. The bat is the cleanest illustration of the limit because the bat is real and because its perceptual world is alien enough to make the gap between physical description and subjective experience intuitively clear. The Owen fMRI findings (treated in the info-box above) sharpen the point. Owen’s protocol can establish, against negative behavioural evidence, that a vegetative-state patient is conscious. It cannot tell us what is is like to be locked inside that body. The third-person science and the first-person experience separate cleanly.

6.3 What Science Cannot Explain

The scope question has several layers:

Value questions. Science can tell you that a particular pesticide causes cancer in laboratory rats at a certain dose. The standard view — going back to Hume’s distinction between “is” and “ought” — is that science cannot tell you whether the risk is acceptable given the agricultural benefits, because that is a question about values, not facts. This view is contested: empirical findings about welfare, suffering, and human flourishing arguably bear on normative questions even if they do not settle them, and Heather Douglas has argued that non-epistemic values already enter scientific reasoning at the point where evidence is weighed. The minimal claim — that science alone, without value premises, does not deliver normative conclusions — is the textbook position. The stronger claim, that science is wholly silent on what we should do, is what evolutionary ethicists, neo-Aristotelians on naturalised goodness, and Sam Harris’s moral landscape programme each, in different ways, deny.

Questions about existence. Science explains the structure of the universe and the processes by which complex structures emerged. It does not explain why there is a universe at all rather than nothing. This may not be a scientific question.

The problem of induction. Science assumes that the future will resemble the past — that the laws governing the universe are stable across time and space. This assumption cannot itself be justified scientifically (it would be circular). It is a presupposition of science, not a result of it.

Historical singularities. Science works best on repeatable phenomena. The origin of life, the origin of the universe, the particular sequence of evolutionary steps that led to consciousness — these are historical singularities that can be studied but not experimentally replicated.

6.4 Questions to Argue About

• Is emergence a genuine scientific phenomenon, or does “emergent properties” just mean “properties we don’t yet know how to reduce”? What would evidence for genuine irreducibility look like?
• Chalmers’ “hard problem” suggests that consciousness cannot be explained scientifically. Does this imply consciousness is supernatural, or just that science needs new concepts?
• “Science can tell us what is but not what ought to be.” Is this a satisfying account of science’s limits? Or can scientific findings (about welfare, suffering, flourishing) bear on normative questions?
• If the assumption that the laws of physics are stable across time cannot be scientifically justified, does that undermine science — or is it an acceptable foundational assumption?

7 What are the ethics of science?

“The scientist is not responsible for the laws of nature,” goes the defence, “but only for discovering them.” It is a clean argument with a long history in scientific self-understanding. It is also disputed: critics argue that scientific methods involve subjects whose consent and welfare matter, that scientific findings have foreseeable consequences whose risks must be weighed, and that scientific applications reshape the world in ways that require governance. The cases below — Tuskegee, He Jiankui, Oppenheimer, Haber — are the standard exhibits in that dispute. Some illustrate clear ethical failure (Tuskegee’s denial of effective treatment); others (Oppenheimer’s role in a wartime weapons programme) admit defensible arguments on more than one side. Whether the “science is neutral; responsibility lies with users” defence holds up across the spectrum, or breaks down at some specific point on the spectrum, is for the reader to argue.

7.1 The Tuskegee Study

The US Public Health Service Syphilis Study at Tuskegee (1932–1972) recruited 399 Black American men with syphilis and 201 uninfected controls in rural Alabama.⁷⁴ The men were told they were being treated for “bad blood”; they were not. When penicillin was established as an effective treatment for syphilis in the mid-1940s, the men were not offered treatment. The study continued for 40 years so that researchers could observe the “natural progression” of untreated syphilis. It ended only when a whistleblower, Peter Buxtun, went to the press in 1972.

The Tuskegee study was not a rogue operation; it was funded by the federal government, published in peer-reviewed journals, and known to public health officials for four decades. Its exposure led directly to the National Research Act (1974) and the Belmont Report (1979), which established the ethical principles — respect for persons, beneficence, justice — that govern human subjects research in the US.⁷⁵

The Nuremberg Code (1947), produced in the wake of Nazi medical experiments, had already established the requirement for informed voluntary consent.⁷⁶ Tuskegee showed that these principles had not been internalised or enforced.

7.2 CRISPR and He Jiankui

In November 2018, Chinese biophysicist He Jiankui announced the birth of the world’s first gene-edited human babies — twin girls, given the pseudonyms Lulu and Nana, whose embryos had been edited to disable the CCR5 gene in an attempt to mimic the naturally-occurring CCR5-Δ32 deletion that confers partial resistance to HIV infection. He timed the announcement to coincide with the Second International Summit on Human Genome Editing in Hong Kong (27–29 November 2018), where he was due to present; the work was not preceded by peer review or regulatory approval and had been conducted in secret. The Shenzhen court that later sentenced him confirmed the existence of a third edited child born to a second couple in 2019.⁷⁷

The scientific community’s response was immediate condemnation. The reasons are multiple: He’s own sequencing data showed that the editing had not in fact reproduced the natural CCR5-Δ32 deletion — Lulu carried at least one off-target intergenic mutation, and both twins showed mosaicism (different cell lines carrying different edits in the same embryo); the editing was performed on healthy embryos (the babies’ father was HIV-positive, but standard sperm-washing IVF protocols already prevent paternal-to-embryo HIV transmission); the long-term safety for the children and their potential offspring is unknown; and the consent process was ethically compromised. He was sentenced to three years in prison by a Shenzhen court on 30 December 2019, and released in April 2022.⁷⁹

7.3 Oppenheimer and Responsibility

J. Robert Oppenheimer directed the Manhattan Project at Los Alamos, which developed the atomic bombs dropped on Hiroshima and Nagasaki in August 1945. After watching the first test explosion at Trinity, New Mexico, on 16 July 1945, he later recalled:

“We knew the world would not be the same. A few people laughed, a few people cried. Most people were silent. I remembered the line from the Hindu scripture, the Bhagavad Gita; Vishnu is trying to persuade the Prince that he should do his duty, and to impress him, takes on his multi-armed form and says: ‘Now I am become death, the destroyer of worlds.’ I suppose we all thought that, one way or another.” — J. Robert Oppenheimer, in the NBC News documentary The Decision to Drop the Bomb (1965), directed by Fred Freed.⁸⁰

The question of scientific responsibility is sharpest here. The scientists who built the bomb knew its purpose. Many had been motivated by the fear that Nazi Germany would build one first. When Germany surrendered in May 1945, the project continued toward its use against Japan.

The Szilard petition. Leo Szilard — the physicist who had originally conceived the nuclear chain reaction, and who in 1939 had drafted with Einstein the letter to President Roosevelt that initiated the American bomb programme⁸¹ — spent the summer of 1945 trying to stop the bomb’s use. In July 1945, Szilard drafted a petition urging President Truman not to use atomic weapons against Japan without a public demonstration and an explicit surrender ultimatum first. Approximately seventy Manhattan Project scientists signed it, primarily at the Metallurgical Laboratory in Chicago.⁸² General Leslie Groves, the military head of the project, blocked the petition from reaching the President before the bomb was dropped on Hiroshima on 6 August 1945; it was subsequently classified, and its signatories were, in several cases, punished in the security-clearance reviews of the following decade. The petition is the ethical fulcrum of the whole Manhattan Project history: the scientists closest to the work were not silent about what they feared it would be used for, and the channels through which scientific conscience could reach political decision-makers were closed by the very institutional arrangements that had made the work possible in the first place.

The ethical question, then, is not only about what scientists know when they produce a weapon. It is also about what they can do with that knowledge inside an institutional architecture designed to use their work while silencing their judgement. The claim that science is morally neutral — that knowledge is just knowledge, and responsibility lies with those who apply it — is a claim that the Szilard petition already refutes: the scientists attempted responsibility and were structurally denied it.

7.4 Questions to Argue About

• The Tuskegee researchers argued (implicitly) that the scientific knowledge gained justified the harm to the participants. Can any scientific knowledge justify research that violates consent? Or is consent an absolute constraint?
• He Jiankui conducted gene editing on human embryos without adequate oversight. But if the editing was safe and the children were not harmed, was there a genuine ethical violation? Or only a procedural one?
• Oppenheimer helped build the atomic bomb, then spent the rest of his life with ambivalence about having done so. Was he morally responsible for Hiroshima and Nagasaki? What is the right account of scientists’ responsibility for applications of their work?
• Is there scientific research that should simply not be done — regardless of whether it could be done safely — because the knowledge it would produce is too dangerous? What would determine this?

8 What is the relationship between science and truth?

Science is widely taken to be the most reliable method for producing knowledge about the natural world. That assessment is widely shared — and, interestingly, one that science itself cannot fully justify without circularity, since any argument for science’s reliability must ultimately use the methods of science. (Hume identified the structurally similar problem about induction; the parallel is exact.) Some philosophers have argued that this circularity is benign and unavoidable for any knowledge-producing practice; others (Feyerabend, some constructivists) have argued that it should make us more modest about science’s claim to special epistemic status compared with other practices. What “most reliable” means — reliable for what purposes, by what measure, in comparison with what alternatives — is itself part of the philosophical question, as is what the relationship between scientific knowledge and truth turns out to be.

8.1 Scientific Consensus

Climate science provides the most politically charged example. The Intergovernmental Panel on Climate Change (IPCC) represents a global consensus among climate scientists: Earth’s average temperature has increased approximately 1.1°C since pre-industrial times, the cause is primarily the emission of greenhouse gases from human activity, and continued emissions will cause further, potentially catastrophic changes.⁸⁶

This consensus is about as solid as scientific consensus gets: thousands of independent researchers, multiple independent lines of evidence (temperature records, sea level rise, ice cores, ocean acidification, phenological changes), multiple independent methodologies.

The “97%” figure that sticks in public discourse comes from one specific study — Cook et al. (2013) — which classified the abstracts of 11,944 climate papers (using AI plus human raters) as endorsing, rejecting, or being neutral on anthropogenic climate change; of those that took a position, 97.1% endorsed it.⁸⁷ Subsequent surveys with different methodologies — Powell (2017) examining 11,602 peer-reviewed articles, Lynas et al. (2021) sampling more recent literature — report consensus levels above 99%. “97%” sticks in public memory because the Cook paper landed there first; what the underlying literature actually offers is one methodologically explicit estimate among several converging estimates, not a single magic number.

Yet public scepticism persists — particularly in certain political communities. Why? The reasons are not primarily epistemological: they are political, economic, and motivated. The fossil fuel industry has funded deliberate campaigns to cast doubt on climate science, importing the tobacco industry’s strategy of manufacturing uncertainty where the scientific picture is clear. This is a case where the sociology of knowledge — who is producing doubt, for what purposes, with what funding — is directly relevant to epistemology.

8.2 Does Science Approach Truth?

Three positions:

Scientific realism: science converges toward truth. The success of science in predicting novel phenomena, manipulating the world, and extending to new domains is best explained by assuming our theories are approximately true. As theories improve, they become more accurate descriptions of reality.

Anti-realism (instrumentalism): science produces models that work, not descriptions that are true. “Truth” doesn’t apply to scientific theories; “empirical adequacy” is the appropriate criterion. Science doesn’t converge toward truth; it converges toward better prediction.

Structural realism (John Worrall): science correctly describes the structure of reality (the mathematical relationships), even if we can’t know the nature of the underlying entities. What persists through scientific revolutions is the mathematical structure (Fresnel’s equations for light, for example, were preserved in Maxwell’s electromagnetism even as the ontology changed from elastic ether to electromagnetic field).⁸⁹

The historical record gives some support to anti-realism: the history of science is littered with theories that were empirically successful for a time and then abandoned — the caloric theory of heat, phlogiston, the luminiferous ether. If past empirically successful theories were false, why should we expect current ones to be true? This is Larry Laudan’s pessimistic meta-induction: the history of science is a history of theories that were empirically successful and subsequently judged false.⁹⁰

8.3 COVID-19 and Scientific Updating in Real Time

The COVID-19 pandemic offered an unusually compressed view of scientific consensus forming, failing, and reforming in public.

Two specific cases. On 14 January 2020 the WHO publicly reported that “preliminary investigations conducted by the Chinese authorities have found no clear evidence of human-to-human transmission” of the novel coronavirus identified in Wuhan; within two weeks the same agency had reversed the claim and was preparing to declare a Public Health Emergency of International Concern (30 January 2020).⁹¹ More substantively, WHO’s early transmission guidance treated SARS-CoV-2 as a droplet-borne pathogen — hence the 1–2 metre / 6 ft distancing rule — and discounted airborne aerosol transmission. In July 2020 Lidia Morawska and Donald Milton, with 239 co-signing scientists, published an open letter in Clinical Infectious Diseases arguing that the existing evidence already established aerosol transmission and that ventilation, not just distance, was the operative variable.⁹² WHO partially acknowledged airborne transmission on 30 April 2021, more fully in December 2021, and in April 2024 redefined “airborne” in a way that absorbed the aerosol-scientists’ position.⁹³

This is the question of how scientific claims are revised as evidence accumulates, played out at unusual speed and in public view. On a Kuhnian reading the COVID episode runs the familiar pattern — anomaly, contested expert dissent, gradual incorporation — at compressed speed; on a Bayesian-updating reading it is just rapid revision under a strong evidential signal, with no need for paradigm-talk. Whether what science was doing in those eighteen months counts as converging on the truth about how the virus actually spreads, or merely as replacing one useful model with a better one, is what Position A and Position B below disagree about.

8.4 AI and the Production of Scientific Knowledge

Large language models can now produce plausible-sounding scientific abstracts, generate hypotheses, and write up experimental results — in some cases passing peer review.⁹⁴ This raises a version of the epistemological questions about shared knowledge: when an AI system produces a scientific claim, what is the epistemic status of that claim? Is it knowledge? Whose?

The scientific community’s current response — developing detection tools, requiring data and code sharing, tightening peer review — is essentially an extension of existing epistemic practices: provenance, transparency, independent replication. What AI-generated scientific content reveals is that these practices were always the epistemologically relevant ones; they matter more, not less, as the cost of generating plausible content falls.

The more interesting question is the upstream one: AI systems trained on the existing scientific literature may systematically amplify the biases, errors, and publication distortions in that literature. If the replication crisis means that the published literature contains many false positives, then AI systems trained on that literature will “know” things that are false — and will express that false knowledge with the confident, technically correct language of scientific writing.

8.5 Questions to Argue About

• Laudan’s pessimistic meta-induction argues that since past scientific theories were false, current ones probably are too. Does this undermine confidence in science? Or is there an asymmetry between past failures and current success?
• Scientific consensus is how science communicates its collective judgement about well-established findings. But scientific consensus has been wrong before (continental drift was rejected by consensus for decades). When, if ever, is it rational to dissent from scientific consensus?
• The WHO’s transmission guidance for SARS-CoV-2 changed substantially between January 2020 and April 2024 — from “no clear evidence of human-to-human transmission” through droplet-only to a redefined notion of airborne. Does this look more like science approaching the truth about how the virus spreads, or more like science replacing one useful model with a better one without either being “true”?
• If science doesn’t converge on truth but only on better models, does it matter? Should our practical attitude toward science depend on our answer to this philosophical question?

9 Media

• Richard Feynman, The Pleasure of Finding Things Out (1981 lecture, published 1999) — Feynman on what science feels like from the inside: the pleasure of uncertainty, the discipline of not knowing, the joy of discovery. One of the most honest accounts of scientific practice by a great scientist.
• Morten Tyldum, The Imitation Game (2014) — Dramatisation of Alan Turing’s work at Bletchley Park; more relevant to mathematics and logic than to the philosophy of science, but raises the ethics of secrecy, sexuality, and the treatment of scientific genius by the state.
• Carl Sagan, Contact (1985, novel; 1997 film) — A meditation on the relationship between scientific evidence and belief; unusually rigorous about epistemological questions for popular science fiction. The film’s ending handles the tension between evidence and faith more carefully than most accounts of the science-religion conflict.
• Primo Levi, The Periodic Table (1975) — Memoirs structured around the elements of the periodic table; a chemist’s account of his own experience, including Auschwitz. Not primarily about the philosophy of science, but about what science can mean to a person and what it cannot protect against.
• Christopher Nolan, Oppenheimer (2023) — Three hours on the Manhattan Project, the Trinity test, and Oppenheimer’s security hearing. The film is primarily about political power and conscience, not physics; but it forces the question of scientific responsibility in a form that is difficult to evade.
• Stanley Kramer, Inherit the Wind (1960) — A dramatisation of the 1925 Scopes “Monkey Trial,” in which a Tennessee schoolteacher was prosecuted for teaching evolution. The film is more useful than the trial transcript for thinking about the rhetorical structure of science-vs.-religion confrontations and the conditions under which scientific evidence is or is not accepted as politically relevant.
• James Watson, The Double Helix (1968) — Watson’s account of the discovery of the structure of DNA; read against the documented record of Rosalind Franklin’s contribution, which Watson minimised. The book is an invaluable case study in the social dimensions of scientific discovery, credit, and gender bias in science.

10 Bibliography