Past exam of the philosophy of science supplementary subject of the University of Oxford / 2022 (A14139W1)

Source: Public Oxford paper, bundle pages 94–95. The cover establishes Hilary Term 2022; the attachment filename uses 2021.

The distinction supplies a coherent strategy for an empiricist realism, but does not by itself justify that realism. It separates an object's relatively stable physical constitution from its effects on perceivers. The strongest defensible version concerns the explanatory roles of properties, rather than a division between wholly real properties and wholly unreal sensations.

For John Locke, primary qualities include extension, shape, solidity, number and motion; secondary qualities are powers of bodies to produce experiences such as colours and tastes through their constitution. The experienced colour need not resemble the physical cause of that experience. Importantly, a power belonging to a body is not thereby imaginary: the primary and secondary qualities distinction need not deny that an unobserved object has dispositions to affect perceivers.^[ref]

This helps empiricism accommodate perceptual variation without abandoning an external world. A surface can produce different colour experiences under different illumination while keeping approximately the same spectral reflectance. A warm and a cool hand can give different sensations in the same water; this need not mean that the water has two incompatible temperatures. The observer's initial state belongs in the explanation. Controlled changes in illumination, viewing conditions and bodily state allow repeatable predictions about experience. An empiricist can therefore treat physical objects as explanatory sources of regularities in experience, rather than demand that every appearance be copied into the object.

However, this explanatory success does not establish that ideas of primary qualities resemble mind-independent properties. We cannot step outside every perceptual and instrumental relation to compare an experience directly with its alleged original. Moreover, the motion and arrangement of unobserved constituents are themselves theoretical commitments. If empiricism permits only what is immediately given in sensation, the distinction introduces precisely the inference beyond sensation that requires defence. If empiricism permits fallible explanatory inference, scientific realism becomes possible, but the justification comes partly from that additional account of inference.

George Berkeley's challenge is that perceived extension and shape vary with viewpoint and perceiver just as colour does, and that extension cannot be experienced wholly apart from sensible qualities.^[ref] This undermines the simple test “variable appearance means merely secondary”. It does not prove that stable physical properties are absent. An obliquely viewed circular disc can produce an elliptical image; geometrical optics explains the relation between the disc and that image. Variation can be evidence for a systematic relation, rather than evidence against an external cause.

Modern physics also discourages treating Locke's list as an unrevisable inventory. Special relativity makes measured length depend on the relation between an object and an inertial frame; motion was relational even in Principle of Galilean relativity. Neither fact makes the measurements arbitrary. Conversely, spectral reflectance is an objective, measurable property even though perceived colour depends on illumination and the perceiver. Objectivity, relationality, perceptual variability and explanatory fundamentality are distinct issues.

The plausible distinction is consequently between an experience and the physical properties and relations that explain it, with the explanation open to correction. An empiricist realism can use this distinction to explain why appearances change while a world persists. It cannot obtain a proof of the external world, or a privileged catalogue of intrinsic properties, simply by naming some qualities primary. The gain is an empirically disciplined model of perception; its limit is that the realist interpretation still needs an argument.

Necessary and sufficient membership conditions are neither a sufficient test of naturalness nor a plausible universal prerequisite for scientific classification. They can characterize some natural kinds very successfully. The mistake is to turn that success into a requirement that every scientifically useful grouping must satisfy a single essentialist pattern.

First distinguish three demands. A classification may have an exact extension, scientists may know an exact test for membership, and membership may be grounded in properties that members necessarily share. These are different claims. Ignorance of an exact test does not show that no common constitution exists. Conversely, an exact definition alone does not show that a category tracks a scientifically significant structure. “An object with mass below one kilogram or temperature above 300 kelvin” specifies necessary and sufficient conditions, yet offers no general reason for a common underlying explanation or reliable extrapolation.

Chemical elements provide a favourable case for a comparatively sharp criterion. In the usual classification of atomic nuclei, carbon has atomic number six, meaning six protons. Different isotopes need not have the same mass, and ionization changes electron number without changing the chemical element. The underlying criterion explains why superficially varying samples can belong together. It also distinguishes a defining physical structure from a readily observed stereotype: ordinary appearance is a poor general test of a chemical element. Even here, the claims must be scoped to the relevant physical objects; an arbitrary mixture containing some carbon is not thereby one carbon chemical atom.

Species make the universal requirement much less attractive. The biological species concept organizes sexually reproducing populations through reproductive relations, but an infertile organism can still belong to a species. Universal actual interbreeding is not required, and asexual organisms demand other classificatory resources. A species is historically extended: descent, gene flow, shared development and environmental interactions help explain similarities without requiring every individual to possess every diagnostic trait. Distinguishing members by a list that merely records all actual individuals would not explain why their similarities support further predictions.

Richard Boyd's homeostatic property cluster theory supplies an alternative. Properties can tend to co-occur because processes maintain their association, while individual members lack some traits and boundaries remain imperfectly sharp.^[ref] For a species, shared ancestry, gene flow and common developmental resources can help maintain such a cluster. “Homeostatic” here need not mean a literal fixed physiological set point. The proposal concerns processes that explain a persistent association of traits.

An objection is that a cluster account merely turns every convenient resemblance into a natural kind. The reply must add a constraint: the grouping should support successful inquiry because it tracks relevant causal or historical structure. Similarity of two laboratory samples due only to the current lighting gives little basis for extrapolating their chemical behaviour. Shared constitution, by contrast, can support predictions under changed illumination and independent testing. This distinguishes useful tracking of nature from a category selected solely to fit a past sample.

A second objection is that vague membership makes disagreement impossible to settle. Sometimes sharper operational criteria are needed, but different research purposes may legitimately attend to different processes. A classification useful for reproductive isolation need not settle every question about descent. That is a reason to state the relevant species concept and domain, rather than to require the scientific world to contain one unique partition for every purpose.

Thus exact membership conditions remain valuable where an explanatory common structure supplies them. Their absence is not a decisive reason to reject a proposed natural kind. The better standard is whether the classification reliably supports explanation and inference through a defensible account of the processes that sustain it.

Vagueness and theory dependence create a burden of clarification for constructive empiricism, but neither yields an immediate contradiction. The harder issue is whether observability deserves the special role in belief that the position assigns to it.

Constructive empiricism, associated with Bas van Fraassen, distinguishes acceptance of a theory from belief in its full truth. Acceptance involves belief in empirical adequacy, together with commitments to use and develop the theory. Empirical adequacy concerns all observable phenomena, including unobserved and future ones; it is stronger than fitting the observations already collected. The position interprets theoretical claims literally but allows suspension of belief about unobservable entities. It is therefore not the claim that theoretical language is meaningless.^[ref]

Observability is a capacity notion. An object can be observable although nobody has observed it: a distant large rock might be seen if a person were suitably situated. Conversely, seeing a detector display does not by itself mean seeing the microscopic entity invoked to explain the display. Van Fraassen treats instruments as devices producing new observable phenomena, which theories must account for, rather than automatically as transparent windows onto whatever their interpretation postulates.^[ref]

Consider first vagueness. Gradually reduce the size or contrast of a visible mark. There need be no uniquely defensible final size at which it ceases to be visible for a human observer. It does not follow that there is no difference between an ordinary tree and an individual electron. Many distinctions have clear central cases and disputed margins. A constructive empiricist can specify the relevant human community and conditions, allow unsettled borderline cases, and suspend judgement where the classification is unclear. This response preserves the core position, although it prevents the epistemic policy from being an exact mechanical rule for every proposition.

Now distinguish kinds of theory-ladenness. Learning to recognize a detector pattern may involve theoretical training. Describing the pattern as evidence of an electron invokes a further interpretation. Neither point shows that a visible arrangement of marks can be accepted only if the whole electron theory is true. Competing accounts may agree on the marks while disagreeing about what produced them. Public calibration, comparison between observers and independent experimental procedures can constrain descriptions without requiring an entirely theory-free vocabulary.

A stronger circularity objection concerns using optics and physiology to determine the limits of observability. If those theories mention unobservable mechanisms, does taking their account of those limits seriously require believing in the mechanisms? One response is to accept the theories' observable consequences about human performance, while remaining agnostic about their full interpretation. Measurements of which marks people discriminate need not establish every microscopic explanation of that performance. The response works best when observable constraints on the capacity claim can be identified independently of the disputed mechanism; simply repeating the theory's own classification would be insufficient.

Nevertheless, the realist retains an important objection. Both ordinary observation and instrument-mediated detection rely on causal processes, calibration and background assumptions. Why should a small change in human perceptual capacity determine the boundary of warranted belief? Independent methods that converge on the same electron electric charge appear to offer a reason for belief extending beyond unaided perception. The constructive empiricist can allow such belief to be rational while denying that acceptance of science requires it, but should acknowledge the resulting permissiveness of the position.

Vagueness therefore limits precision rather than destroying the observable/unobservable distinction. Theory-ladenness defeats a naive foundationalism, not every use of observational evidence. The deepest unresolved question is normative: whether withholding belief beyond empirical adequacy is a compelling policy, or merely one coherent attitude toward the same scientific achievements.

A defensible demarcation evaluates a practice's accountability to evidence, rather than treating one syntactic feature of a claim as a complete test. There are clear cases of pseudoscience, even if there is no exceptionless rule that classifies every immature or disputed research programme immediately.

Karl Popper's falsifiability criterion captures an essential insight: a claim has empirical content when some possible outcomes conflict with it. A prediction that a specified sample will always have the same electric charge would exclude outcomes that a claim allowing every possible result does not. If apparent failures can always be redescribed as successes, accumulated favourable stories do not supply a demanding test. Popper also emphasizes risky predictions and warns that protective modifications can destroy a theory's testability.^[ref]

However, falsifiability alone is not sufficient. A bare numerical prediction chosen without evidence may be perfectly falsifiable. So can a particular astrological claim. Testing such a claim seriously can be scientific work; a movement that advertises the claim while systematically ignoring its failure can still be pseudoscience. Logical vulnerability and actual willingness to learn from tests are different properties. Moreover, a false theory can be scientific, while a true guess can lack a scientific basis. Demarcation is not identical to separating true from false statements.

Nor should every discrepancy lead to immediate rejection. Predictions depend on instruments, initial conditions and auxiliary hypotheses. A detector fault may explain a failed prediction more reasonably than rejection of a well-tested physical law. Imre Lakatos therefore evaluates sequences of theories and their treatment of anomalies: a progressive programme generates new testable successes, whereas a degenerating one primarily adjusts to what is already known.^[ref] This improves on assessing an isolated sentence, although deciding when a programme has had enough time is itself contestable.

A workable standard combines several mutually supporting demands. Claims should specify relevant evidence and conditions of application; tests should expose them to substantial risk of failure; data and methods should permit criticism and appropriate scientific replication; and revisions should improve independently assessable performance. The crucial unit is a sustained practice presenting itself as scientific knowledge. Persistent selective reporting, explanations invented solely to absorb each failure, and refusal to permit competent scrutiny undermine that claim to authority.

For a concrete illustration, suppose a proposed material emits a new kind of signal. Initially poor understanding does not make the proposal pseudoscience. Researchers could specify detector settings, compare blinded signal-present and signal-absent trials, publish failures, and seek independent scientific replication. Contrast a practice that discloses only positive trials, changes the claimed conditions after each negative result, and gives no independent test of the revised conditions. The difference concerns the treatment of evidence, even if both practices use technical language and the same apparatus. The comparison is hypothetical; it does not depend on attributing misconduct to a named group.

This standard has borderline cases. A promising theory can temporarily lack accessible tests, and a programme can rationally survive a period of failure. It should then be represented with the corresponding uncertainty, rather than sold as established. Longstanding prestige, popularity and institutional affiliation are evidence about a community at most; they cannot replace examination of its methods.

Such a criterion is fallible and allows degrees, but that is compatible with useful discrimination. The best boundary identifies practices that systematically evade the evidential responsibilities they claim to meet. It preserves Popper's demand for genuine exposure to failure while incorporating the historical fact that responsible science tests and revises interconnected systems of claims.

The covering-law model explains an event by showing how it falls under general laws in specified circumstances. It captures a major form of scientific understanding, but entailment and predictability do not by themselves capture explanatory relevance or direction.

In the deductive-nomological model, let $L$ be laws, $C$ relevant circumstances, and $E$ the event description. The explanans is $L,C$; the explanandum is $E$. The argument has the formThe deduction must be valid, at least one law must be essential to it, the explanans must have empirical content, and its premises must be true for a sound explanation. Merely writing a valid argument from arbitrary premises is insufficient. Carl Hempel and Paul Oppenheim also identify a structural affinity between explanation and prediction: supplied beforehand, the same premises could in principle support the prediction.^[ref]

For an ideal simple pendulum at small amplitude, Newtonian mechanics and the specified constraints yieldHere $ell$ is the pendulum's length, $g$ the local gravitational acceleration, and $T$ its oscillation period. The deduction connects an observed regularity to more general laws and explains why changing length changes the oscillation period. The approximation must be stated: a finite-amplitude, dissipative simple pendulum does not exactly satisfy these premises. A useful model-based explanation can be approximately correct even when the ideal premises are not literally true.

The model's appeal is that it requires more than an evocative story. It displays assumptions and shows why they constrain the phenomenon. It can also explain laws through wider laws: geometrical optics can recover a refraction law from suitable descriptions of wave propagation. A principle theory can explain by constraints without supplying a detailed constructive theory of every constituent process. This is a strength of the account's generality.

The first difficulty is asymmetry. For a vertical pole of vertical length $h$ on level ground and the Sun's elevation angle $\alpha$, geometrical optics gives the shadow's lengthWith suitable premises, either deduction is valid. Yet the pole and illumination explain the shadow, whereas measuring the shadow ordinarily allows inference to the pole's vertical length without explaining why the pole has that vertical length. The same equation supports both calculations. Its logical form does not encode the relevant direction of dependence. Context can change the question, but the distinction between explanation and evidential reconstruction remains.

The second difficulty is irrelevance. Adding the premise that a laboratory wall is blue to a correct simple pendulum derivation preserves validity and the essential use of physical laws. The extra premise contributes nothing to explaining the oscillation period. More elaborate examples show the same problem when irrelevant factors are hidden inside statistical regularities. The explanatory account needs a way to identify what matters under relevant changes, not just what accompanies a valid derivation.

The inductive-statistical model extends the programme to explanations using statistical laws and high conditional probability, relative to appropriate available information. This is an extension beyond the deductive focus of the 1948 article. High probability is nevertheless not necessary for every explanation. Radioactive decay can explain an infrequent detector event through a physical process and its nuclear decay constant, although that particular event in a short interval was unlikely. Nor does a high conditional probability automatically establish explanatory relevance: a reliable indicator may be an effect of the outcome or share another cause with it.

A better assessment therefore retains law-based derivations as one explanatory resource while adding dependence, relevance and the question being asked. Mechanistic explanations identify productive processes; accounts based on interventions ask which changes would alter the outcome. Neither proposal needs to force every mathematical or constraint-based explanation into a causal story. The covering-law model illuminates how laws make phenomena intelligible, but it is too broad as a sufficient test and too restrictive as a universal necessary form.

The history of science can constrain philosophical claims and support conditional methodological lessons. It cannot, without further premises, turn a record of what scientists did into a rule about what everyone ought to believe. Its strongest contribution is often to reveal that an apparently obvious account of science was too simple.

Consider the claim that scientific progress always adds new facts while preserving every central accepted proposition. The transition from Newtonian mechanics to special relativity supplies a counterexample. Under a Galilean transformation, temporal separation is absolute; a Lorentz transformation mixes time and space coordinates. These are incompatible accounts at unrestricted scope. Yet Newtonian mechanics retains great predictive value at low velocity. Historical change therefore supports a distinction between wholesale accumulation of truths and more limited retention of successful relations. It neither shows that all past knowledge was worthless nor guarantees that current fundamental theories are final.

Thomas Kuhn makes the replacement of central commitments, problems and standards central to his account of scientific revolutions. His introductory examples concern changes that reorganize inquiry, rather than merely adding one item to a fixed inventory.^[ref] This supports a warning against reading present categories unaltered into the past. It does not establish that different scientific communities can never compare evidence or understand one another. Semantic or methodological change is weaker than the claim that rational comparison is impossible.

A second lesson concerns anomalies. If actual successful inquiry sometimes retains a theory while examining instruments or auxiliary hypotheses, a methodology demanding abandonment at the first discrepancy misdescribes that inquiry. This matters because experimental predictions concern a package, not a lone sentence. Lakatos's comparison of progressive and degenerating programmes turns attention to whether retaining a framework produces independently corroborated developments.^[ref] The lesson is conditional: persistence is defensible when it opens productive tests and alternatives, not simply because famous scientists once persisted.

The history of science also puts pressure on scientific realism. If empirically successful theories have posited entities later rejected, success alone cannot guarantee literal truth. Larry Laudan's historical challenge targets a proposed explanatory connection between success, reference and approximate truth.^[ref] But a list of failures cannot establish the fraction of successful theories that fail unless the population and selection process are specified. Cases chosen precisely because they were overturned are an especially poor basis for an unqualified numerical forecast about present theories.

There is also a distinction between a historical cause of belief and a justification for belief. Training, instrumentation and institutions affect which theories scientists can develop and test. This makes science a human activity without showing that its results are merely expressions of local preference. A historically informed evaluation should compare the evidence available at the time, the competing proposals, and the costs of different mistakes. Otherwise, hindsight converts the eventual winner's assumptions into premises that earlier participants supposedly should already have known.

These conclusions are philosophically important because they change how we allocate trust. They favour confidence differentiated by claim and domain, explicit attention to auxiliary hypotheses, and humility about extrapolation. They also favour institutions that preserve criticism and scientific replication, because individual ingenuity does not ensure reliable correction. The further normative premise is explicit: methods deserve support insofar as they help inquiry detect error and achieve its evidential aims.

History consequently offers counterexamples to universal philosophical generalizations and evidence for competing explanations of scientific change. It does not choose a unique epistemology automatically. Its value lies in making accounts of rationality and progress answerable to what careful investigation actually reveals, while keeping the passage from description to evaluation open to argument.

Scientific confirmation is evidential support rather than deductive proof. Two distinct problems concern, first, what counts as relevant positive evidence and, second, which part of a tested theoretical package that evidence supports. A defensible response makes background information and the production of evidence explicit. The following two problems are independent: resolving the logical puzzle about positive instances does not eliminate ambiguity about auxiliary hypotheses.

Problem one: the raven paradox. Write the hypothesis aswhere $R$ means “is a raven” and $B$ “is black”. Its contrapositive isThe hypotheses are logically equivalent. A plausible equivalence principle says that identical information must support them equally. A plausible instance principle says that an observed $F$ that is $G$ confirms “every $F$ is $G$”. Combining the principles makes a nonblack nonraven, such as a white shoe, support the raven hypothesis. That conflicts with the intuition that inspecting shoes tells us nothing about ravens. Hempel develops the puzzle in his analysis of instance-based scientific confirmation.^[ref]

My preferred response retains logical equivalence while rejecting an unconditional instance rule. By Bayes' theorem, for evidence $E$, background $K$ and positive relevant probabilities,An instance's bearing depends on how likely it was under the competing possibilities, including how the object was selected. No rule depending only on the grammatical form of an instance can settle that comparison. Branden Fitelson and James Hawthorne analyse the role of background assumptions and Bayesian confirmation in this puzzle.^[ref]

Here is an explicit illustrative model. There are 100 labelled objects: ten known ravens and ninety known nonravens. Under $H$ all ten ravens are black; under the sole alternative $A$, exactly one of the ten is nonblack, with its identity uniformly distributed. Nonraven colours are already known and identical under both possibilities, and the prior probabilities of $H,A$ are equal. Selecting a known white shoe according to a rule independent of $H,A$ has the same likelihood under each, so the posterior probability of $H$ stays $1/2$. A uniformly selected raven, however, is black with probability $1$ under $H$ and $9/10$ under $A$. After observing it to be black,If sampling is instead from known nonblack objects whose membership is uncertain, discovering a nonraven can rule out a possible counterexample and provide support. There is no context-independent prohibition on support from a nonraven. The response explains both the familiar irrelevance intuition and why it is not universally correct. Its limit is that the prior probabilities and sampling assumptions need independent defence; the calculation is a transparent model, not a theorem that all shoes are irrelevant.

Problem two: the Duhem problem of testing a package. Suppose the target hypothesis $T$ and auxiliary hypotheses $A$ jointly imply an outcome $E$:A failure gives $\neg (T\wedge A)$, equivalently $\neg T\vee \neg A$, rather than $\neg T$ alone. Success also does not establish either premise: that would affirm the consequent. Pierre Duhem's point is especially pressing for physical experiments whose predictions require theories of the apparatus as well as the target system.^[ref]

My preferred response combines independent calibration, deliberately varied tests and explicit comparison of packages. For a simple model, suppose a detector has independently calibrated false-positive probability $f$ and detection efficiency $d$. A target predicts a signal on each trial with probability $s$. Then the positive-output probability isUnder a no-signal rival it is $p_0=f$. If $s=1/2$, $d=9/10$, and $f=1/10$, then $p_T=1/2$. One positive reading supplies a Bayes factor of $5$ for the specified target package against the specified no-signal package. But an uncalibrated faulty detector with false-positive probability $1/2$ fits the same positive-reading rate. Target data alone do not distinguish those packages.

Signal-absent calibration trials test the false-positive assumption directly. If ten independent such trials are all negative, their likelihood ratio in favour of $f=1/10$ against $f=1/2$ isThis does not magically isolate $T$: the calibration relies on its own assumptions, and correlated trials would invalidate the power calculation. It shows concretely how additional evidence can discriminate explanations that the initial data leave unresolved. Repeating the target measurement with apparatus based on different physical processes further reduces dependence on one shared failure mode.

Thus the Duhem problem defeats automatic attribution of support or failure, not rational experimental comparison. Where rivals remain empirically equivalent, suspension of judgement about their difference is appropriate. Together, the two responses defend scientific confirmation as a structured, fallible comparison using background information, rather than either pure instance counting or a mechanical verdict on an isolated hypothesis.

Scientific change is a serious threat to indiscriminate scientific realism, but not a deductive refutation of every realist position. A defensible realism must identify which claims deserve belief, at what precision and within which domain, rather than merely promise that mature theories are “approximately true”.

Scientific realism includes the claim that a mind-independent world exists and the stronger epistemic claim that scientific inquiry gives warranted knowledge of at least some unobservable aspects of it. Theory replacement directly challenges the latter inference, not the bare existence of an external world. Disagreement between successive descriptions does not imply that neither describes anything real.

The historical objection is nevertheless substantial. A theory can make successful predictions while its central interpretation is later rejected. The luminiferous aether illustrates the risk of identifying successful wave optics with the literal truth of a particular mechanical medium. If earlier scientists had impressive evidence and still endorsed dispensable theoretical entities, why suppose that comparable evidence now guarantees our own entire interpretation? Laudan presses this challenge against the connection between success, reference and approximate truth.^[ref]

One reply distinguishes literal truth from controlled approximation. In special relativity, the kinetic energy of a massive particle isThe Taylor expansion for $|v|/c\ll 1$ givesThus the classical kinetic energy is recovered in a specified regime, with a leading relative correction of $3v^2/(4c^2)$. It would be wrong to conclude that Newtonian mechanics is globally true, or that all its physical concepts survive unchanged. It would be equally wrong to infer that its low-speed predictions were accidental. The survival of a quantitatively controlled relation is a more discriminating claim than survival of every piece of ontology.

John Worrall's structural realism makes preservation of mathematical relations central, using the transition from Augustin Fresnel's optics to electromagnetism as a key example.^[ref] This explains why a theory might retain a real achievement while abandoning its proposed mechanism. However, structure cannot mean just any equations with some empirical fit: the realist must identify the relevant physical interpretation, domain and counterfactual content. Otherwise the view risks saying little beyond empirical adequacy.

A related selective response commits to elements that carry a theory's tested successes, while treating speculative additions more cautiously. Its strongest evidence is convergence across independent experiments, robustness under changes in modelling assumptions, and success on predictions not used to tune the theory. For example, support for a stable electron electric charge need not amount to belief in every current claim about the ultimate nature of an electron. But selectivity must be defensible before a later replacement occurs; choosing surviving components only in hindsight cannot vindicate past realist confidence.

The anti-realist can still explain scientific success by demanding empirical fit and retaining successful models, without inferring truth about unobservables. That response blocks any simple argument that realism is logically compulsory. Conversely, pointing to historical failures does not establish that every presently supported claim will fail. Such an inference would need a justified comparison class, a measure of success and an account of how testing has changed.

The appropriate conclusion is discriminating fallibilism. Changes threaten claims of finality and blanket trust in a theory's full ontology. They can also reveal stable, independently tested relations and improve our understanding of their limits. Scientific realism survives as a qualified position only if it makes those distinctions explicit and remains willing to revise the very commitments it currently regards as best supported.