1 What is an argument?

In everyday speech, “argument” means a quarrel. In logic, it means something precise and technical: a set of propositions in which some (the premises) are offered as reasons to believe another (the conclusion). The confusion between these two senses is itself philosophically interesting — and tells us something about what distinguishes the logical enterprise from mere persuasion.

1.1 John Snow and the Broad Street Pump

In August 1854, cholera killed more than five hundred people within ten days in the Soho district of London. The prevailing medical theory — miasma, the idea that disease spread through foul-smelling air — had the support of the General Board of Health and most of the medical establishment. The physician John Snow was convinced the disease was waterborne but had no microscope and no germ theory. What he had was an argument. He mapped every death against its street address and found they clustered around a single water pump on Broad Street. He then examined the exceptions systematically: workers at a local brewery who drank only beer had not died; a widow in Hampstead who received Broad Street water by cart had died. Each exception, rather than refuting his hypothesis, reinforced it. Snow presented his evidence to the local Board of Guardians, who removed the pump handle on 8 September 1854 — famously regarded as one of the first acts of evidence-based public health.⁴ Snow’s case is the textbook illustration of the difference between a rhetorical performance and a genuine argument: his premises were independently verifiable, his conclusion followed from them, and the argument was structured so that a single clear counterexample could have destroyed it.

1.2 The Logical Skeleton of a Claim

Every genuine argument has a structure. Strip away the rhetoric, the examples, the tone, and what remains is a skeleton of premises and a conclusion. Logic is the study of what makes that skeleton strong or weak.

The simplest structure: modus ponens (Latin: “the way that affirms”):

1. If P, then Q.
2. P.
3. Therefore, Q.

Example:

1. If the bacteria is sensitive to penicillin, the patient will recover.
2. The bacteria is sensitive to penicillin.
3. Therefore, the patient will recover.

This is valid. The conclusion must follow from the premises — if the premises are true, the conclusion cannot be false.

A small caveat for what follows: “valid” here means classically valid — valid in the system formalised by Frege, Russell, and Whitehead. Modus ponens is its central inference rule. Alternative systems (relevance, intuitionistic, paraconsistent, dialetheic) modify or weaken classical inference; they are working logics used in mathematics, computer science, and the philosophy of vagueness, and are taken up in §8.⁵

This structure appears everywhere, often hidden. “We should tax carbon because it damages the atmosphere and damaging the atmosphere is wrong” conceals:

1. (Implied) We should stop things that are wrong.
2. Damaging the atmosphere is wrong.
3. Carbon emissions damage the atmosphere.
4. Therefore, we should stop carbon emissions [i.e., tax them].

1.3 Deductive vs. Inductive Arguments

There are two fundamentally different types of inference:

Deductive: The conclusion is guaranteed by the premises. If the premises are true and the argument is valid, the conclusion cannot be false. Mathematics and formal logic are deductive. The price: the conclusion can only contain what the premises already implicitly contained.

Inductive: The premises support the conclusion, but do not guarantee it. All observed ravens have been black; therefore (probably) all ravens are black. The conclusion goes beyond the premises. This is how science works — but it is also where Hume’s problem of induction bites.

The difference is important for TOK. When someone says “studies show that X leads to Y, therefore policy Z is correct,” this involves both an inductive step (from studies to generalisation) and often a deductive step that conceals a value premise.

1.4 Distinction from Persuasion

Aristotle distinguished three modes of appeal in rhetoric (Rhetoric, 4th c. BCE):⁶

• Logos: appeal to reason and logic
• Ethos: appeal to the character and credibility of the speaker
• Pathos: appeal to the emotions of the audience

Logos is what logic studies. Ethos and Pathos are effective — often more effective than Logos — but they are not arguments in the logical sense. They can produce belief without justification.

flowchart LR
  P["<b>Premises</b><br/>claims offered<br/>as reasons"]
  P -- "an inference rule<br/>licenses the step" --> C["<b>Conclusion</b><br/>the claim they<br/>are reasons for"]
  C --> DE["<b>Deductive</b><br/>the premises <i>guarantee</i> it.<br/>The price: nothing in the<br/>conclusion that was not<br/>already in the premises"]
  C --> IN["<b>Inductive</b><br/>the premises <i>support</i> it.<br/>The conclusion goes beyond<br/>them — which is why<br/>Hume is waiting in Lesson 4"]

1.5 The Socratic Method as Argument

Socrates (as depicted in Plato’s early dialogues) did not deliver lectures. He asked questions. The elenchus (Socratic refutation) is a method of argument by cross-examination: take the interlocutor’s claim, show that it implies something else they also believe, show that the two beliefs contradict each other. The interlocutor must revise.

Example (Euthyphro, c. 380 BCE): Euthyphro claims that piety is what the gods love. Socrates asks: Do the gods love something because it is pious, or is it pious because the gods love it? (The Euthyphro Dilemma.)⁷ Either answer creates problems for Euthyphro’s definition. He cannot hold all his beliefs at once.

The elenchus is a tool of negative dialectic: it doesn’t tell you what the right answer is, but it eliminates bad answers by showing their internal contradictions.

The popular tag often attached to this method — “I know that I know nothing” — is not a sentence Socrates ever utters in the dialogues. What Plato actually has him say at Apology 21d is more careful: he is wiser than the man he has just questioned only in that he does not think he knows what he does not know. The crisper Latin formula (scio me nihil scire) is a much later distillation.⁸

Watson and the 200+ co-signatories of the open letter to The Lancet in May 2020 worked exactly this method: they accepted the form of Mehra et al.’s argument and demanded the authors confront a premise (data integrity) the authors could not consistently supply.

1.6 The Regress Problem: Carroll’s Tortoise

Lewis Carroll (“What the Tortoise Said to Achilles,” Mind, 1895) posed a problem about the justification of logical inference that has never been fully resolved.⁹

To see the paradox, hold one distinction firmly in mind: a premise is a claim inside the argument (it can be true or false, and it is what the argument rests on); an inference rule is the pattern that licenses getting from the premises to the conclusion (it is what the argument moves by). “All humans are mortal” is a premise; “from ‘all A are B’ and ‘c is A’ conclude ‘c is B’” is an inference rule. Carroll’s Tortoise accepts every premise Achilles offers — and still refuses to move.

Achilles and the Tortoise discuss the following argument:

• 1. Things that are equal to the same are equal to each other.
• (B) The two sides of this Triangle are things that are equal to the same.
• (Z) Therefore the two sides of this Triangle are equal to each other.

The Tortoise accepts (A) and (B) but refuses to accept (Z). Achilles says: “But if you accept A and B, you must accept Z.” The Tortoise agrees — and asks Achilles to add that as a new premise: call it (C). Now the Tortoise accepts A, B, and C — but still refuses to accept Z without a further premise explaining why A, B, and C together entail Z. The regress continues without end.

Carroll’s point: the rule of inference (modus ponens — if the premises are true, the conclusion follows) cannot itself be justified by adding it as a premise. Any justification of an inferential rule already uses inferential rules.

This is not a puzzle to be solved. It is a structural feature of formal systems. Inference rules are not conclusions of arguments — they are the machinery that produces conclusions. They must be accepted as basic, not derived. (What does ground inference rules, if not further inferences, is taken up in §8 — Quine vs BonJour.¹⁰¹¹)

1.7 Questions

Questions 2 and 4 are retrieval checks — answer them from memory before you argue the rest.

1. Can you be persuaded of something true by an invalid argument? And if so, is that a problem?
2. “You can’t get from is to ought” (Hume). State the rule: if an argument contains only factual premises, what kind of conclusion can they not yield — and what kind of premise has to be smuggled in to get there?
3. The Socratic method is negative — it destroys positions rather than building them. Is destruction a valid form of philosophical progress? Or does philosophy need to construct as well as demolish?
4. Every argument has premises, and premises need justification — but Carroll’s Tortoise runs the regress on something else. State what the Tortoise demands of Achilles at each step, and what the parable shows about whether an inference rule can be justified by adding it as a premise.

2 What makes an argument valid — and what makes it sound?

Validity and soundness are not synonyms. They are technical terms, precisely defined, and the distinction between them is one of the most important tools in all of philosophy. Confusing them is a mark of philosophical inexperience; understanding them clearly unlocks enormous analytical power.

2.1 Validity

An argument is valid if the conclusion cannot be false when all the premises are true. Validity is about the logical relationship between premises and conclusion, not about whether the premises are actually true.

This argument is valid:

1. All fish can fly.
2. Salmon are fish.
3. Therefore, salmon can fly.

Premise 1 is false. The conclusion is false. But the argument is valid — because if premise 1 were true, and if premise 2 were true, the conclusion would have to be true.

This argument is invalid:

1. All cats are mammals.
2. Dogs are mammals.
3. Therefore, dogs are cats.

This is the fallacy of the undistributed middle: from “All A are B” and “All C are B,” it does not follow that “C is A,” because the middle term (“mammals”) is not distributed — it does not refer to all mammals in either premise, so the two premises do not pin the cats and the dogs to the same subset of mammals. (The related but distinct fallacy of affirming the consequent has propositional, not categorical, form: “If P then Q; Q; therefore P” — for example, “If it rained, the street is wet; the street is wet; therefore it rained.”) The premises here are true; the conclusion is false; therefore the argument is invalid.

Key insight: a valid argument guarantees the truth of the conclusion given the truth of the premises.

2.2 Soundness

An argument is sound if it is both (1) valid and (2) has all true premises. Soundness is what we actually want in practice: a sound argument guarantees a true conclusion.

In philosophy, we often work with valid but unsound arguments — thought experiments and hypotheticals where premises are stipulated rather than established.

flowchart TB
  A["An argument"] --> Q1["Could the conclusion be false<br/>while every premise is true?"]
  Q1 -- "yes — and here<br/>is the case" --> INV["<b>Invalid</b><br/>the counterexample you<br/>just built is the verdict"]
  Q1 -- "no" --> VAL["<b>Valid</b><br/>the form carries truth from<br/>premises to conclusion"]
  VAL --> Q2["Are the premises<br/>in fact true?"]
  Q2 -- "no" --> UNS["Valid but unsound —<br/>where thought experiments<br/>and hypotheticals live"]
  Q2 -- "yes" --> SOU["<b>Sound</b><br/>the conclusion is true"]

2.3 Modus Ponens and Modus Tollens

The two most important valid argument forms:

Modus ponens (“affirming the antecedent”):

• If P, then Q.
• P.
• Therefore, Q.

Modus tollens (“denying the consequent”):

• If P, then Q.
• Not Q.
• Therefore, not P.

Modus tollens is how falsification works in science. The general theory of relativity predicts that light bends around massive objects (P → Q). Eddington’s 1919 observation found that light does bend (Q — prediction confirmed). But if the light had not bent (¬Q), that would have falsified the theory (¬P), by modus tollens. This is the logical structure of Popper’s falsificationism.

A wrinkle worth knowing: the tidy symmetry between a conditional and its contrapositive does not survive in every logic. Modus ponens and modus tollens are both valid in constructive (intuitionistic) mathematics, but the classical equivalence behind them is not: contraposition runs in only one direction (from P → Q to ¬Q → ¬P), and recovering a conditional from its contrapositive requires double-negation elimination, a move the intuitionist refuses. (This is taken up in §8.)¹⁷

Modus tollens is also how a defendant would falsify the state’s argument — given access to the premises. The Loomis case is exactly the institutional shape where that access is denied: when premises are produced by a system whose internal workings are a trade secret, the modus-tollens move the defence would otherwise make is unavailable.

flowchart LR
  C["The conditional<br/>If P, then Q"]
  C --> MP["<b>Modus ponens</b><br/>affirming the antecedent<br/>P — therefore Q"]
  C --> MT["<b>Modus tollens</b><br/>denying the consequent<br/>not-Q — therefore not-P"]
  MP --> MP2["Relativity says the light<br/>will bend; 1919 found<br/>that it does"]
  MT --> MT2["Had the light <i>not</i> bent,<br/>the theory would have<br/>gone — this is what<br/>falsification is"]

2.4 The “True Premise, False Conclusion” Test

A powerful diagnostic: if you can construct a scenario where all the premises are true and the conclusion is false, the argument is invalid. This is the method of counterexample. It is one of the core techniques of philosophical analysis.

“All politicians are untrustworthy; the local librarian is untrustworthy; therefore, the local librarian is a politician.” Is this valid? Construct a counterexample of the same form: “All cats are mammals; all dogs are mammals; therefore all dogs are cats.” Premises true, conclusion false.

2.5 Questions

Questions 3 and 4 are retrieval checks — answer them from memory before you argue the rest.

1. Why do people often find invalid arguments compelling? (Connect to base-rate neglect and the prosecutor’s fallacy.)
2. A valid argument with a false conclusion must have a false premise. So: if you find yourself rejecting a conclusion, you are committed to rejecting at least one premise. What are the consequences of this for political and ethical debate?
3. State the two conditions an argument must meet to be sound, and say whether arguments from analogy or inference to the best explanation are valid in the formal sense. Can an argument be good — worth believing — without being either?
4. Mathematics consists entirely of valid deductive arguments from axioms. State exactly what a valid deduction guarantees about its conclusion — and what it leaves unestablished. Does that make mathematical conclusions certain?

3 How should I update what I believe when new evidence comes in?

Most reasoning in real life is not deductive. The premises are not certain, the conclusion is not entailed, and what we want to know is not “does this follow?” but “does this evidence make this hypothesis more likely?” The mathematics of that question — how a prior belief should be revised in the light of new data — is Bayes’s theorem. The most common way to misuse it is to ignore the base rate: the prior probability of the hypothesis before the evidence arrives. Smart people, including expert witnesses, get this catastrophically wrong.

3.1 Bayes’s Theorem (the toy version)

The theorem, due to the eighteenth-century English clergyman Thomas Bayes, describes how a belief should change in response to evidence.²¹ The intuition is one line:

posterior is proportional to prior × likelihood.

In words: the probability of a hypothesis given some evidence is proportional to the probability of the hypothesis before the evidence (the prior) multiplied by the probability of seeing this evidence if the hypothesis were true (the likelihood). Graham Priest puts the failure mode bluntly: many bad arguments are “seductive because people often confuse probabilities with their inverses, and so slide over a crucial part of the argument.”²² The probability of evidence given a hypothesis (\(P(E \mid H)\)) is not the probability of the hypothesis given the evidence (\(P(H \mid E)\)).

3.2 Base-rate Neglect: This Time You Compute

A disease affects 1 person in 1,000 (the base rate). A test for the disease is 99% accurate — 99% of sufferers test positive (sensitivity), and 99% of non-sufferers test negative (specificity). You test positive. What is the probability you have the disease?

The intuitive answer, and the one most doctors give when polled, is “about 99%.” It is wrong, and this is the one calculation in the unit that will not be handed to you worked. The three inputs above are everything you need.

3.3 Representativeness and the Base-rate Failure

Daniel Kahneman’s chapter “Tom W’s Specialty” in Thinking, Fast and Slow — reporting work done with Amos Tversky — shows the same error in plain prose. Asked to guess whether “Tom W” — described as quiet, neat, and orderly — is studying computer science or humanities, both lay subjects and trained graduate students in psychology pick computer science, even when explicitly told that the humanities cohort is many times larger.²³ We judge by representativeness — how well the description matches a stereotype — and ignore the prior probability of the category. Kahneman’s chapter title for the cab-problem analysis is exact: Causes Trump Statistics. We reach for the causal story (the witness saw a blue cab, so the cab was blue) and discard the statistical fact (most cabs in the city are green).

3.4 The Prosecutor’s Fallacy, Named

The Sally Clark case has a name in the statistical literature: the prosecutor’s fallacy. It consists in asserting \(P(\text{evidence} \mid \text{innocent})\) — the probability of seeing this evidence if the defendant were innocent — and inviting the jury to hear it as \(P(\text{innocent} \mid \text{evidence})\). The two quantities can differ by orders of magnitude. The fallacy recurred in the conviction of the Dutch nurse Lucia de Berk (2003, quashed 2010) and in DNA-match cases where “the chance of a random match is 1 in 10 million” is presented without reference to the size of the database searched.

flowchart LR
  Q["Two questions<br/>that sound alike"]
  Q --> A["P(evidence | innocent)<br/><br/>How likely is this evidence<br/>if the defendant is innocent?<br/><br/><i>Meadow's 1 in 73 million</i>"]
  Q --> B["P(innocent | evidence)<br/><br/>How likely is innocence,<br/>given this evidence?<br/><br/><i>what the jury needed</i>"]
  A --> F["<b>The prosecutor's fallacy</b><br/>assert the first,<br/>invite the jury to<br/>hear the second"]
  B --> F

3.5 Bayesianism as a Theory of Rational Belief

Bayes’s theorem is a calculation. Bayesianism is the broader epistemological claim that rational belief is graded — that beliefs come in degrees, that those degrees obey the probability axioms, and that the rational response to evidence is to update one’s degrees of belief in line with the theorem. This is a substantive claim, not an obvious one. Classical logic treats belief as binary: a proposition is accepted or it is not. The Bayesian thinks this is a coarsening of what rationality actually requires. The classicist replies that probabilistic belief leaves no place for the classical notions of validity and proof — that a proposition either is or is not entailed, and degrees of belief are about psychology, not logic.

The Sally Clark case sits inside this dispute. The jury was operating in a classical-acceptance frame: presented with a number — “1 in 73 million” — they were asked to render a binary verdict (guilty or not guilty). The Bayesian frame would have asked them instead to compute a posterior, weighing the prior probability that a mother kills two of her infants against the prior probability that two SIDS deaths occur in the same family, before any single-figure statistical claim was allowed to dominate the deliberation. Neither frame is logically forced; the question is which the institution should run on, and at what cost.

3.6 Reading an Error Bar: Eddington’s Two Intervals

Updating on evidence requires knowing how much the evidence is worth, and the experimental sciences print that worth in the margin of the result itself: the error bar. A reported value of \(1.98 \pm 0.12\) is not one number but an interval — a claim about where the true value plausibly lies — and the width of the interval is the measurement’s confession of how far it could still be wrong. A narrow interval is strong evidence; a wide one should move a rational believer only a little. The most consequential pair of error bars in twentieth-century physics makes the point with real numbers. In May 1919, two British expeditions photographed stars near the eclipsed sun to test general relativity, which predicted that starlight grazing the sun’s edge would bend by 1.75 arcseconds; a Newtonian corpuscular calculation predicted half that, 0.87; no deflection at all was a third live option. The published results: Sobral, in Brazil, \(1.98 \pm 0.12\) arcseconds; Príncipe, off West Africa, where cloud spoiled most of the plates, \(1.61 \pm 0.30\).²⁴

3.7 Questions

Questions 1 and 3 are retrieval checks — answer them from memory before you argue the rest.

1. Your circled number from the screening exercise: state it, and the arithmetic behind it — of 1,000 people screened, how many test positive, and how many of those positives are true? Why does the rational answer sit an order of magnitude below the doctor’s “about 99%” — and what goes wrong when the doctor’s intuition sets screening policy?
2. Sally Clark’s jury was told a number — 1 in 73 million — that was both technically wrong and rhetorically devastating. Is the deeper failure innumeracy (jurors and judges who could not interrogate the figure) or epistemic theatre (an expert witness deploying a figure designed to overpower interrogation)?
3. A defendant’s DNA matches the crime scene with a “1 in a million” random-match probability, found by searching a police database of five million profiles. Compute the expected number of chance matches in that search — and name the fallacy committed when the jury hears “1 in a million” as the probability of innocence.
4. If rational belief is a matter of degree (Bayesianism), what becomes of the classical logical notion of acceptance — the binary yes/no judgement that a proof produces? Are these two pictures compatible, or does one have to give way?

4 Can we be certain through induction?

Almost everything we believe about the world — that aspirin cures headaches, that the sun will rise tomorrow, that jumping from heights causes injury — is based on induction. We generalise from what we have observed to what we have not yet observed. And yet induction, as Hume showed in the 18th century, cannot be rationally justified without circularity.

4.1 Hume’s Problem

David Hume (A Treatise of Human Nature, 1739; An Enquiry Concerning Human Understanding, 1748) posed the problem of induction with a precision that has never been answered satisfactorily:²⁷

The observation: past regularities give us no logical guarantee about future regularities. We have observed the sun rising every day. But: does the past rising of the sun provide any logical reason to expect it to rise tomorrow?

The would-be justification: of course past regularities are a guide to the future — nature is uniform, things that have always happened will continue to happen.

Hume’s response: How do you know nature is uniform? By observing that it has been uniform in the past. But you are using inductive reasoning to justify inductive reasoning.

“There can be no demonstrative arguments to prove that those instances of which we have had no experience resemble those of which we have had experience.” — David Hume, A Treatise of Human Nature, Book I, Part III, Section VI²⁸

Behind the circularity stands Hume’s Fork: all propositions are either relations of ideas (like mathematics — necessarily true, but empty of empirical content) or matters of fact (empirically known through experience, but not necessarily true). The uniformity of nature is not a relation of ideas, so no demonstration can secure it; it is a matter of fact, so only experience — the very thing in question — can support it.

flowchart LR
  P["Every proposition<br/>is one or the other"]
  P --> R["<b>Relations of ideas</b><br/>mathematics, definitions.<br/>Necessarily true — and<br/>empty of any claim<br/>about the world"]
  P --> M["<b>Matters of fact</b><br/>known through experience.<br/>About the world — and<br/>never necessary"]
  R --> G["'Nature is uniform' is<br/>not a relation of ideas,<br/>so no demonstration reaches it;<br/>and as a matter of fact it<br/>can be supported only by<br/>the experience in question"]
  M --> G

4.2 Russell’s Inductivist Chicken

Bertrand Russell (The Problems of Philosophy, 1912) gave a now-famous illustration. A chicken is fed every morning by the farmer. Every morning, without exception, the farmer comes and feeds it. The chicken, by induction, comes to expect this. Then one morning the farmer comes — and wrings its neck instead.²⁹

The chicken’s inductive reasoning was impeccable. Its conclusion was disastrously wrong. The past regularity provided no logical guarantee of the future. Russell’s moral, in Problems of Philosophy Chapter VI, is more careful than the chicken story alone suggests: he holds that induction must be backed by a separate Principle of Induction, which itself cannot be proved from experience and must be accepted as a piece of a priori knowledge if induction is to have any rational warrant.

Russell’s chicken is itself a repurposing of Hume’s earlier example, in Enquiry §IV, of bread that has nourished us in the past and is expected to nourish us again. Hume’s point was that “there is no known connexion between the sensible qualities and the secret powers” — having seen one loaf nourish gives no rational ground to expect the next to do so, even though we cannot help expecting it.³⁰

4.3 The Black Swan

Before Europeans reached Australia, every swan ever observed was white. “All swans are white” was, for Europeans, a very well-confirmed inductive generalisation. Then black swans were found in Australia.

Karl Popper (The Logic of Scientific Discovery, 1934) used this to argue that no number of confirming instances can prove a universal generalisation — but a single disconfirming instance can disprove it. Hence his criterion: scientific claims should be falsifiable. We should try to refute our theories, not accumulate confirmations.³¹ (Taleb’s modern extension of this point is the RBS case in the info-box above.)

4.4 Falsifiability vs. Verificationism

The Logical Positivists of the Vienna Circle (1920s–1930s) proposed verificationism: a statement is meaningful if and only if it is in principle verifiable through observation. Metaphysical claims (“God exists”) and ethical claims (“murder is wrong”) are not verifiable and hence meaningless — not false, but empty. The classic English statement is A. J. Ayer’s Language, Truth and Logic (1936).³²

Popper rejected verificationism and proposed falsificationism instead: what marks science is not that its claims can be verified but that they can be falsified — there is some observation that would count against them.

The asymmetry: No number of white swans proves “all swans are white.” But one black swan disproves it.

flowchart LR
  U["A universal claim<br/>All swans are white"]
  U --> C["One more white swan"]
  U --> F["One black swan"]
  C --> C2["Adds nothing decisive.<br/>No finite number of<br/>confirmations proves<br/>a universal claim"]
  F --> F2["Settles it. A single<br/>counter-instance refutes<br/>the claim outright"]
  C2 --> P["<b>Popper's criterion</b><br/>what marks a claim as<br/>scientific is not that it<br/>could be verified but that<br/>it could be falsified"]
  F2 --> P

VaR was, on Popper’s criterion, falsifiable: any single trading day on which loss exceeded the 99% threshold would have refuted the claim that 99% of trading days lay within it. The question the RBS case raises is not whether VaR was falsifiable but whether falsifiability inside the observation window is enough — and Popper’s criterion has nothing to say to a system whose specifications were correct on every day before 2008.

4.5 Responses to Hume

No one has solved Hume’s problem. But responses range:

• Pragmatic vindication (Hans Reichenbach): induction is the best strategy we have, even if it can’t be justified a priori. If nature is regular, induction will find it; if not, no method will.³³
• Bayesian inductivism (Howson and Urbach): we update probability estimates in response to evidence according to Bayes’ Theorem. Not a solution — Bayes’ Theorem requires prior probabilities, which are themselves inductively based — but a formalisation of the reasoning.
• Statistical-learning theory (Vapnik): induction works under specifiable conditions on the function class and the sampling distribution; outside those conditions, generalisation guarantees fail. Useful machinery, not a refutation of Hume.
• The material theory of induction (John Norton): induction is not a single formal schema but a family of domain-specific inferences licensed by background empirical facts about the subject matter.
• Inference to the best explanation (Peter Lipton): we infer to the hypothesis that, if true, would best explain the evidence — a procedure that runs alongside Bayesian updating rather than replacing it.
• Naturalism (Quine): the question “is induction justified?” is itself answerable only by empirical investigation — there is no standpoint outside our epistemic practices from which to evaluate them.³⁴
• Taleb’s fat-tailed extension: even granting some inductive method, event-rarity scales with consequence — long-run outcomes are dominated by rare, high-impact events from fat-tailed distributions, which mainstream statistical practice systematically underestimates. RBS is the live case.³⁵

4.6 Questions

Questions 2 and 3 are retrieval checks — answer them from memory before you argue the rest.

1. If Hume is right that induction cannot be rationally justified, does that make science irrational? Or just differently rational than we thought?
2. State Russell’s inductivist chicken — the evidence it had, and the morning it got — and what the example shows about what enumerative induction can guarantee. Does that make the problem of induction practically serious, or just a philosophical puzzle without real-world implications?
3. Popper said science advances through falsification, not confirmation. But in practice, scientists protect their theories by adjusting auxiliary hypotheses: state the move precisely — when a prediction fails, which premise gets revised? — and say what the discovery of Neptune in 1846 showed about it, when astronomers rescued Newton from the anomaly in Uranus’s orbit by postulating an unseen planet. Bad science or rational practice?
4. What is the relationship between Hume’s problem of induction and the reliability of memory? If past experience doesn’t logically justify expectations about the future, does past experience of your own identity justify beliefs about who you are?

5 What are the limits of formal systems?

In 1931, a 25-year-old mathematician named Kurt Gödel published a paper with a title so technical it could clear a room: “On Formally Undecidable Propositions of Principia Mathematica and Related Systems.” Inside it were two theorems that shook the foundations of mathematics and, by extension, of formal knowledge.³⁶

The theorems tell us that any sufficiently powerful formal system is either incomplete (there are truths it cannot prove) or inconsistent (it can prove contradictions). You cannot have both completeness and consistency. This is not a technical result that concerns only specialists.

5.1 The Dream of Formalism

By the early 20th century, mathematicians — led by David Hilbert — had an ambitious programme: to place all of mathematics on an absolutely secure axiomatic foundation. Propose a complete set of axioms; show the axioms are consistent (they don’t lead to contradictions); show they are complete (every mathematical truth can be derived from them). Mathematics would then be a perfect, closed system — a realm of certain knowledge.⁴⁰

The crack in that ambition was already visible thirty years before Gödel. In 1901 Bertrand Russell discovered a contradiction inside the axioms of Gottlob Frege’s Grundgesetze der Arithmetik — the set of all sets not containing themselves: if it contains itself, it doesn’t; if it doesn’t, it does — and his letter reached Frege in June 1902.⁴¹ Frege acknowledged the flaw in a despairing appendix to the second volume, already at the printer. Russell and Whitehead’s Principia Mathematica (1910–1913) attempted the repair across nearly 2,000 pages and a “theory of types”; the work takes 379 pages to prove that 1 + 1 = 2.⁴²

5.2 Gödel’s First Incompleteness Theorem

Kurt Gödel, in “On Formally Undecidable Propositions of Principia Mathematica and Related Systems” (1931), proved:

First Incompleteness Theorem: Any consistent formal system capable of expressing basic arithmetic contains true statements that cannot be proved within the system.

The proof uses a piece of self-reference. Gödel devised a method of encoding statements about the formal system within the system itself — Gödel numbering, which assigns a unique whole number to every formula, allowing statements about formulas to be translated back into statements about arithmetic. Using this coding, he constructed a statement G whose informal content is:

G — This statement cannot be proved in this system.

Before the formal argument, a self-descriptive warm-up. “This sentence has five words” is true not because someone proved it but because the sentence correctly describes itself. Look at the sentence; count the words; the content of the claim and the state of the world coincide. G is built to do the same thing, except what it describes is its own unprovability rather than its word count.

The two horns. Suppose the system is consistent — it never proves a false statement. Then:

• If G were provable, the system would have proved a statement that asserts its own unprovability. That is a false claim being proved. The system would be inconsistent.
• If G is not provable, then G’s claim about itself — that it cannot be proved — is accurate.

Why unprovability entails truth. In a consistent system, the first horn is ruled out. So G is not provable. But G asserts precisely that it is not provable — and that assertion, by the argument just given, is correct. A statement whose content is correct is true.

A precision worth registering. Gödel’s 1931 proof in fact required a slightly stronger condition than mere consistency — a property called ω-consistency — to rule out the system proving the negation of G as well. J. B. Rosser strengthened the result in 1936 by constructing a different sentence for which simple consistency suffices. The teaching exposition above silently relies on Rosser’s strengthening; the historical Gödel argument is a hair more delicate.⁴³

5.3 Gödel’s Second Incompleteness Theorem

Even more devastating for Hilbert’s programme:

Second Incompleteness Theorem: A consistent formal system cannot prove its own consistency.

In other words: you cannot use mathematics to prove that mathematics doesn’t contain a hidden contradiction. Any proof of consistency would itself have to be produced within a formal system. MCAS could not recognise from inside its specification that it had left it; Hilbert’s programme could not recognise from inside the system that the system was consistent — the structural shape is the same.

5.4 Consistency and Completeness

Gödel’s results force a tradeoff:

• A consistent system never proves a contradiction (not both P and ¬P).
• A complete system can prove or refute every statement in its language.

These cannot both be achieved in any system powerful enough to encode arithmetic. We must choose. Mathematicians choose consistency, since from a contradiction, anything at all can be derived (the principle of explosion: ex contradictione quodlibet).

What this means: mathematical truth is not identical to mathematical provability — there is a gap, and Hilbert’s programme of securing all mathematics on a finite set of consistent, complete axioms is impossible. What this does not mean: that mathematics is unreliable. Gödel sentences are constructed specifically to be unprovable; they do not arise in ordinary mathematical practice. Whether incompleteness directly implies that the human mind is not a formal system is a separate philosophical question — pressed by J.R. Lucas and Roger Penrose, contested ever since, and the one the question below isolates.⁴⁵

flowchart LR
  H["<b>Hilbert's programme</b><br/>one axiom set for all of<br/>mathematics: complete,<br/>consistent, and provably so"]
  H --> G1["<b>First theorem</b> (1931)<br/>any consistent system able to<br/>express arithmetic contains<br/>truths it cannot prove"]
  H --> G2["<b>Second theorem</b><br/>such a system cannot<br/>prove its own consistency"]
  G1 --> T["<b>The forced trade</b><br/>consistency or completeness<br/>— never both"]
  G2 --> T
  T --> CH["Mathematics takes consistency:<br/>from a contradiction anything<br/>at all follows — <i>ex contradictione<br/>quodlibet</i>"]
  T --> GA["…and truth comes apart<br/>from provability"]

5.5 Questions

Questions 1 and 2 are retrieval checks — answer them from memory before you argue the rest.

1. State Gödel’s two incompleteness theorems, one line each. Which of the two does the slogan “there are mathematical truths that cannot be proved” compress — and does it make mathematics less certain, or just different from what we thought?
2. The Second Incompleteness Theorem: state exactly what it rules out — and therefore where mathematicians’ confidence in their systems’ consistency must come from, if not from a proof inside the system.
3. Some philosophers argue that Gödel’s theorems show human minds transcend formal systems (because we can “see” that G is true, even though the system cannot prove it). Does this argument work?
4. Are there analogues to Gödel’s incompleteness outside mathematics — domains of knowledge where there are truths that the domain’s own methods cannot reach?

6 What is a paradox — and what can we learn from one?

A paradox is not a contradiction. A contradiction is wrong. A paradox is an argument that proceeds from sound-looking premises by valid-looking steps to a conclusion that is absurd or contradictory.

6.1 Zeno’s Paradoxes

Around 450 BCE, Zeno of Elea argued (according to Aristotle’s later account) that motion is impossible. His most famous argument: the race between Achilles and a tortoise.⁵⁰

The tortoise has a head start. Achilles runs faster. To catch the tortoise, Achilles must first reach the point where the tortoise was. But by then, the tortoise has moved forward. Achilles must now reach that point. But again, the tortoise has moved. This sequence has infinitely many steps. Can an infinite number of steps be completed?

Zeno argued: no. Therefore, Achilles never catches the tortoise. Therefore, motion as we perceive it is an illusion.

This sounds absurd — of course Achilles catches the tortoise. But the argument seems valid. Where is the flaw?

The mathematical resolution (Cauchy, early 19th century): an infinite series can have a finite sum. The series \(1/2 + 1/4 + 1/8 + \ldots = 1\). So infinitely many steps can be completed in finite time, if the steps get correspondingly shorter. Modern analysis dissolves the paradox mathematically.

But the philosophical question remains: does the mathematical resolution tell us what actually happens in nature, or does it just tell us how to calculate? Is space and time continuous (as calculus assumes) or discrete (as quantum mechanics might suggest)?

6.2 Russell’s Set Paradox (and the Liar)

The older self-reference paradox is the Liar: “This sentence is false” — true if false, false if true. Known in antiquity through Epimenides the Cretan (“All Cretans are liars”), it motivated Tarski’s hierarchy of languages and Gödel’s incompleteness construction (a sentence that says “I am not provable”).⁵¹

Russell’s paradox is the set-theoretic cousin. In 1901, Bertrand Russell discovered a contradiction at the heart of naive set theory. Consider the set of all sets that do not contain themselves.⁵²

• Does this set contain itself? If yes: it is a set that contains itself — so it should not be in our set (contradiction). If no: it is a set that does not contain itself — so it should be in our set (contradiction).

Russell wrote to Frege, who had just published his Basic Laws of Arithmetic, which relied on naive set theory. Frege’s response is one of the most poignant moments in intellectual history:

“Your discovery of the contradiction has surprised me beyond words and, I should almost like to say, left me thunderstruck, because it has rocked the ground on which I meant to build arithmetic.” — Gottlob Frege, letter to Bertrand Russell, 1902⁵³

The resolution requires restricting set formation — the axioms of Zermelo-Fraenkel set theory prohibit the kind of self-reference that generates Russell’s paradox.

6.3 The Sorites Paradox (The Heap Problem)

While Russell’s paradox attacks the foundations of mathematics, the Sorites paradox attacks something more fundamental still: the vague predicates of ordinary language. Sorites comes from the Greek soros: heap.

1. One grain of sand is not a heap.
2. Adding one grain of sand to a non-heap does not make a heap.
3. Therefore, one million grains of sand is not a heap.

The argument is valid. The premises seem true. The conclusion is false. Where is the error?

The Sorites paradox arises for any vague predicate: tall, bald, rich, young, red. There is no sharp boundary between tall and not-tall. Remove one hair at a time from a full head: when exactly does the person become bald? There is no fact of the matter.

The philosophical responses:

• Epistemicism (Timothy Williamson): there is a precise boundary; we just can’t know where it is.⁵⁴
• Supervaluationism: “John is bald” is neither true nor false in borderline cases — a truth value gap.
• Fuzzy logic: truth comes in degrees. “John is bald” is 0.7 true for borderline cases.
• Contextualism: “heap” is context-sensitive; the boundary shifts with the conversational context.

flowchart LR
  P["A paradox: plausible<br/>premises, valid-looking<br/>steps, an absurd conclusion"]
  P --> Z["<b>Zeno</b><br/>threatens motion, and<br/>the structure of<br/>space and time"]
  P --> R["<b>Russell and the Liar</b><br/>threaten self-reference,<br/>and the foundations<br/>of set theory"]
  P --> S["<b>Sorites</b><br/>threatens the vague<br/>predicates of<br/>ordinary language"]
  Z --> Z2["Cauchy: an infinite series<br/>can have a finite sum.<br/>But is that what happens,<br/>or only how we calculate?"]
  R --> R2["Zermelo–Fraenkel bars the<br/>self-membered set; Tarski<br/>bars a language that holds<br/>its own truth predicate"]
  S --> S2["Epistemicism · supervaluationism<br/>· fuzzy logic · contextualism<br/>— four bills, none cheap"]

The paradox is practically important: law, medicine, and policy deal constantly with vague predicates — when does a fetus become a person? at what point does a company have “significant” market power?

6.4 Questions

Questions 1 and 3 are retrieval checks — answer them from memory before you argue the rest.

1. Zeno’s paradoxes were “solved” by calculus; the Sorites has resisted every such dissolution. Name the four standard responses to the Sorites and say, in one line each, what each does with the boundary — then say which of them the Vasa tribunal would need if it wanted a fact of the matter to discover.
2. The Liar paradox suggests that unrestricted self-reference leads to contradiction. Are there other areas of thought — apart from formal logic — where self-reference is dangerous?
3. State Russell’s paradox — the set it constructs, and the contradiction on both horns — and what the Zermelo-Fraenkel resolution had to give up. What does that tell us about the reliability of intuition in mathematics?
4. The Sorites paradox arises because our concepts are vague. Is vagueness a defect of language that could in principle be eliminated? Or is it essential to how language works?

7 Can logic tell us what ought to be?

Logic tells us what follows from given premises. If you accept certain things, you must accept certain other things. But here is a question logic cannot directly answer: what should you accept in the first place? And, more acutely, can any set of purely factual premises logically entail a moral conclusion? Hume noticed, in a footnote that changed the history of philosophy, that it apparently cannot — and moral philosophers have been arguing with that footnote ever since.

7.1 Hume’s Is-Ought Gap

David Hume noticed something in moral philosophy (A Treatise of Human Nature, Book III, Part I, Section I, 1739):⁵⁶

“In every system of morality, which I have hitherto met with, I have always remarked, that the author proceeds for some time in the ordinary way of reasoning, and establishes the being of a God, or makes observations concerning human affairs; when of a sudden I am surpriz’d to find, that instead of the usual copulations of propositions, is and is not, I meet with no proposition that is not connected with an ought, or an ought not.”

Hume observed that moral philosophers of his time would describe facts about human nature, God, or society — and then, without any logical bridge that he could find, slide into moral conclusions. “People desire happiness. Therefore, we ought to promote happiness.” The word “therefore” conceals a gap. (Natural-law theorists, the targets of Hume’s critique, deny that the gap exists at all: their factual premises about human nature already include teleological content from which obligations follow without any further normative addition. Whether this dissolves the gap or merely smuggles the ought into the description of human nature is the question Hume opened.)

The gap: no set of purely descriptive premises can logically entail a normative conclusion without an additional normative premise. From “people suffer when tortured” you cannot logically derive “you ought not torture people” without some further premise like “you ought not cause suffering.” That further premise is not itself derivable from facts alone.

This is sometimes called Hume’s Guillotine.

7.2 Moore’s Naturalistic Fallacy

G.E. Moore (Principia Ethica, 1903) made a related point. He argued that the property of being “good” is a simple, non-natural property — it cannot be defined in terms of any natural property (pleasure, survival, what God commands, etc.). Any attempt to define good in natural terms commits the naturalistic fallacy.⁵⁷

His test: the Open Question Argument. Suppose you define “good” as “what produces pleasure.” Then “Is pleasure good?” should be a closed question — trivially true by definition. But it isn’t. It remains a genuinely open question whether pleasure is good. Therefore “good” and “productive of pleasure” cannot mean the same thing. Repeat for any natural property: the question remains open.⁵⁸

Moore thought the naturalistic fallacy explained why all previous ethical theories failed.

7.3 The Structure of Ethical Arguments

Every ethical argument must contain at least one ethical premise. You cannot derive an “ought” from a pile of “is” statements. This means ethical arguments always have a normative foundation — and that foundation must itself be defended ethically, not empirically.

Example:

1. The death penalty does not deter crime. [Empirical claim]
2. Punishments that don’t deter crime should not be used. [Normative premise]
3. Therefore, the death penalty should not be used. [Normative conclusion]

Premise 1 can be contested empirically. Premise 2 requires moral argument. Someone might accept 1 and reject 2 on the grounds that punishment has other purposes (retribution, justice).

flowchart TB
  subgraph GAP["What Hume's guillotine forbids"]
    I1["<b>Is</b><br/>The death penalty<br/>does not deter crime"]
    I1 -- "no valid step" --> O1["<b>Ought</b><br/>The death penalty<br/>should not be used"]
  end
  GAP ~~~ REP
  subgraph REP["The only repair: supply the missing premise"]
    I2["<b>Is</b><br/>The death penalty<br/>does not deter crime"] --> O2["<b>Ought</b><br/>The death penalty<br/>should not be used"]
    N["<b>Ought</b>, supplied<br/>Punishments that do not<br/>deter should not be used"] --> O2
  end

When knowledge claims in ethics are at issue, distinguishing the empirical and normative components of an argument is a crucial TOK skill.

7.4 Can Science Legislate Values?

A recurring question in modern thought: can neuroscience, evolutionary biology, or psychology tell us what is good? Sam Harris (The Moral Landscape, 2010) argues yes: if morality is about wellbeing, and science can study wellbeing, then science can study morality.⁵⁹ E.O. Wilson (Sociobiology, 1975) suggested that human moral intuitions are adaptations selected for their fitness benefits.⁶⁰

Hume’s guillotine, on the standard reading, cuts these arguments: even if science can tell us what maximises wellbeing, it cannot — without a further normative step — tell us that we ought to maximise wellbeing. That step requires a moral commitment that science cannot supply.

The standard reading is not unanimous. John Searle (“How to Derive ‘Ought’ from ‘Is’”, 1964) argued that constitutive facts about institutional practices — promising, marrying, signing a contract — generate ought-conclusions from is-premises, because the institutional facts already encode normative content.⁶¹ Philippa Foot (Natural Goodness, 2001) revived an Aristotelian line on which “good” for a living thing is grounded in the natural form of life it has: a sound oak is one that grows tall and bears acorns; a good human is one whose practical reasoning is in order — and these are not normative additions to the natural facts but their proper description.⁶² Whether these responses dissolve Hume’s gap or merely relocate the normative premise inside “constitutive rule” or “natural form of life” is itself contested.

7.5 Questions

Questions 1 and 4 are retrieval checks — answer them from memory before you argue the rest.

1. “Natural selection has produced in us a tendency to care for our children. Therefore, it is natural — and right — to care for our children.” Name the fallacy this argument commits, and state the missing premise that would be needed to make it valid — and where that premise would have to come from.
2. Is Hume’s is-ought gap a logical point or a metaphysical one? Could there be a world in which “is” and “ought” were more directly connected?
3. If all ethical arguments require at least one normative premise, where do those foundational premises come from? Are they intuitions? Conventions? Revealed by God? Constructed?
4. State Moore’s Open Question Argument: the test it runs on any proposed definition of “good,” and what the question’s remaining open is supposed to show. Does the argument work? Can you find a case where it breaks down?

8 Does logic have to be classical?

Classical logic — the logic Aristotle articulated and that Frege and Russell formalised — rests on certain principles that feel self-evident. Among them: the principle of non-contradiction (nothing can be both true and false), the law of excluded middle (every proposition is either true or false), and the principle of explosion (from a contradiction, anything follows). These feel like bedrock — like the very frame within which thought is possible. But that they feel like bedrock is not by itself an argument that they are bedrock. There are well-developed logics that reject each of them, used today in working mathematics, computer science, and the analysis of vagueness.

8.1 The Principle of Non-Contradiction

Aristotle called it the most certain of all principles:

“It is impossible for the same thing to belong and not to belong at the same time to the same thing and in the same respect.” — Aristotle, Metaphysics, Book IV, Chapter 3⁶⁵

Non-contradiction is the logical backbone of rational argument.

But: is the principle a discovered truth about reality, or a constructed rule for our logical practices? Could reality be, in some domains, genuinely contradictory?

8.2 Fuzzy Logic

The lesson on the Sorites paradox above showed how badly classical logic handles vague predicates: one grain is not a heap; adding a single grain to a non-heap produces a non-heap; therefore (by induction) no number of grains is a heap. The argument is valid; the conclusion is absurd; one of the premises must go. Lotfi Zadeh’s 1965 proposal — fuzzy logic — pays for the absurdity by abandoning bivalence: truth values are not just 0 (false) or 1 (true) but any real number in the interval [0, 1]. “John is tall” can be 0.7 true. “These 100 grains are a heap” can be 0.3 true.⁶⁶

This is not just a philosophical move. Fuzzy logic is the working logic of camera autofocus, washing-machine water-level sensors, anti-lock braking systems, and a generation of medical diagnostic decision-support tools. Each is a domain where the world supplies a continuous signal — a focus distance, a soil-moisture reading, a wheel-slip angle, a haemoglobin saturation — and the older binary engineering practice (sharp threshold, with hysteresis around it) misclassified at the boundary often enough to matter. A fuzzy system instead computes a degree of membership in each of several overlapping categories (“dry / damp / wet”), evaluates rules at degrees, and outputs a defuzzified action.

The philosophical move underneath the engineering: classical logic models a world of sharp distinctions, even where the underlying signal is continuous. Whether that produces errors at the boundary or merely registers a useful idealisation depends on what the system is supposed to do. For an anti-lock braking system the binary classical predicate “wheel locked” misses what matters; the engineering moved to fuzzy logic and the system improved.

The cleanest objection is that fuzzy logic just relocates the problem: if “John is tall” is 0.7 true, is that statement (the assignment of degree 0.7) precisely true, or itself only approximately so? The fuzzy logician’s reply — that the assignment is itself a fuzzy proposition, with its own degree — invites a regress that has been a standing topic in the literature; see Williamson, Vagueness (1994), for the case that vague predicates are best handled with classical logic plus epistemicism (there is a sharp boundary; we just cannot know where). In short: fuzzy logic buys a better fit at the boundary, and pays for it with a puzzle about its own degrees.

8.3 Paraconsistent Logic

Paraconsistent logics reject the principle of explosion: they allow contradictions to coexist within a system without licensing arbitrary conclusions from them. This sounds alarming, but it is closer to working practice than the principle of explosion is:

• A legal system may contain inconsistent laws (old law says X; new law says not-X; the contradiction hasn’t been resolved by repeal). Lawyers still reason within it.
• A database may contain contradictory entries (data-entry errors, source conflicts, time-of-update mismatches). Queries still run.
• Quantum mechanics and general relativity are locally inconsistent at certain scales. Working physicists continue to use both, picking which apparatus to invoke per problem.

Graham Priest (In Contradiction, 1987) pushes the claim further into dialetheism — the view that some contradictions are not merely tolerable but true. The Liar sentence (“this sentence is false”) is the canonical case: it appears to be both true and false. Dialetheism is not mainstream, but it is philosophically serious; the alternative (some hierarchy or restriction that blocks the Liar) has its own costs. Whether dialetheism is also the right reading of older non-Western traditions — Buddhist Madhyamaka and the catuṣkoṭi — is a live question, taken up below where it bears on the discovery-vs-construction debate about logic itself.⁶⁷

flowchart LR
  C["<b>Classical logic</b><br/>non-contradiction ·<br/>excluded middle ·<br/>explosion"]
  C --> F["<b>Fuzzy logic</b><br/>gives up two truth values.<br/>Truth comes in degrees:<br/>'John is bald' is 0.7 true"]
  C --> P["<b>Paraconsistent logic</b><br/>gives up explosion.<br/>A contradiction no longer<br/>licenses every conclusion"]
  C --> I["<b>Intuitionistic logic</b><br/>gives up excluded middle.<br/>'P or not-P' must be<br/>constructed, not assumed"]
  P --> D["<b>Priest's dialetheism</b><br/>pushes further: some<br/>contradictions are not<br/>merely tolerable but true"]

8.4 Questions

Questions 1 and 2 are retrieval checks — answer them from memory before you argue the rest.

1. Aristotle called non-contradiction the most certain of principles. State his formulation from Metaphysics IV, and the self-validation argument — what happens to any attempt to doubt the principle? Then say what that argument does not establish: is “cannot be doubted without being used” the same as “true”?
2. Fuzzy logic admits truth values between 0 and 1. State the regress objection precisely — if “John is tall” is 0.7 true, what question re-arises about the assignment itself? — and the fuzzy logician’s standing reply, with the fee it pays.
3. Paraconsistent logic allows for some contradictions without collapse. But if you accept any contradiction, how do you decide which contradictions are “tolerable” and which are disqualifying?
4. Quine said that logic is revisable in principle. If that’s right, what is the relationship between logic and knowledge? Can you even reason about the revision of logic using logic?

8.5 What Grounds the Choice of Logical System?

The Minamata question forced a choice between logical instruments. That choice raises a question logic itself cannot answer: what grounds the choice of logical system? You cannot use logical inference to justify the choice of inferential rules — as Carroll’s Tortoise (§1) showed — because whichever rules you use in the justification are already the ones you are assuming.

Three positions bracket the territory:

• Platonism about logic: logical laws are objective, necessary truths about abstract structures. We discover them.
• Formalism: logic is a formal game with rules we choose. Different games (classical, fuzzy, intuitionistic, paraconsistent) are tools for different purposes.
• Naturalism: logic is continuous with science. Our logical beliefs are revisable in light of experience, just as empirical beliefs are. Quine argued the general revisability thesis in “Two Dogmas” — no statement, including a logical one, is in principle immune from revision in the face of recalcitrant experience.⁷⁰ Hilary Putnam pushed the specific application to quantum mechanics in “Is Logic Empirical?” (1968), arguing that the strange behaviour of quantum systems gives us empirical reason to adopt a non-distributive quantum logic. Both moves are contested; the Quine reading especially has been resisted on the grounds that he never quite committed to revising any particular logical law on empirical grounds.⁷¹

Two more focused positions bracket the dispute over inference rules in particular:

• Quine holds that logical laws are simply the most entrenched nodes in our web of belief — not known a priori in any special sense, just the last things we would revise.
• BonJour (In Defense of Pure Reason, 1998) holds that some knowledge is genuinely a priori — rational insight into necessary truths that no experience could undermine. Logical laws are known a priori because their negation cannot be coherently entertained.⁷²

A non-Western thread sharpens the question. The dialetheist position introduced under Paraconsistent Logic above has a deeper precedent in the Buddhist catuṣkoṭi (Skt. “four corners”), systematically deployed by the 2nd-century Madhyamaka philosopher Nāgārjuna. Mūlamadhyamakakārikā I introduces the four-fold scheme: for any proposition P, the catuṣkoṭi asks whether P, not-P, both P and not-P, or neither P nor not-P. Nāgārjuna’s central use is to argue that the four corners exhaust the available positions on (e.g.) dependent origination — and that ultimate reality (paramārtha) cannot be fixed by any of them.⁷³ Whether the catuṣkoṭi is best read as an early dialetheist logic (Priest’s reading), as a paraconsistent move (Garfield), or as a strictly non-logical device for breaking the reader’s attachment to any conceptual position (the standard Madhyamaka self-understanding), is itself contested — and bears directly on whether “is this a logic?” is a discovery question or a construction question.

8.6 The RBS Verdict, Stress-Tested

At the RBS question you entered a minute in the risk committee’s record: the modellers failed Hume’s standard, or met the only standard there is, or were playing the wrong game altogether. You signed that before the unit’s second half, and the second half has been quietly pressing on it.

Take out what you wrote and run it against three pressures it did not face at the time:

• MCAS. The 737 MAX’s flight-control law was correct inside its specification and had no way of recognising that it had left it — the same shape as a VaR model correct on every trading day before 2008. If you took the practitioner’s seat, does “best available methods” survive a case where the methods’ own specification is what failed? If you took Hume’s or Taleb’s, does your verdict on the modellers now convict every certified system — or can you say where certification stops and misplaced confidence begins?
• The threshold. The 99% in “we will not lose more than £X on 99% of trading days” is a line drawn across a continuous distribution of outcomes — the Vasa and Minamata question wearing a banker’s suit. Does your seat treat that line as discovered or drawn, and does your answer here match the one you gave the certification board?
• The grounding. Whatever standard of rationality your minute invoked, the choice-of-logic lesson asked what grounds a standard at all — Quine’s entrenchment or BonJour’s insight. Is the sense of “rational” you committed to the most entrenched node in our web of belief, or something you claim to see directly? And would your answer to that question survive the seat you took?

Re-sign or revise — in writing, a few sentences either way. A revision is not a defeat; an unexamined signature is. Name what the intervening lessons changed, or say why they changed nothing.

9 Media

Novels, films, and artworks that illuminate the questions above:

• Lewis Carroll, Alice’s Adventures in Wonderland (1865) and Alice Through the Looking-Glass (1871) — Carroll was a logician (Charles Dodgson, lecturer in mathematics at Oxford). The books are saturated with logical puzzles, non-sequiturs, and violations of inference. The Mad Hatter’s riddle (“Why is a raven like a writing desk?”) has no answer — a joke about the expectation that arguments have conclusions.
• Douglas Hofstadter, Gödel, Escher, Bach: An Eternal Golden Braid (1979) — Pulitzer Prize-winning exploration of self-reference, formal systems, consciousness, and Gödel’s theorems, structured around parallels between Bach’s counterpoint and Escher’s impossible drawings. Difficult and brilliant.
• Jorge Luis Borges, “The Library of Babel” (in Ficciones, 1944) — A universe consisting of an infinite library containing all possible books. A meditation on completeness, infinity, and the limits of formal systems.
• Tom Stoppard, Rosencrantz and Guildenstern Are Dead (1967) — A play in which two minor Shakespeare characters discover they are living through a script they cannot control. Questions of logical necessity, free will, and the impossibility of reasoning outside one’s own system.
• Christopher Nolan, Inception (2010) — Nested realities that cast doubt on the reliability of any single level of “truth.” Connects to questions of consistency, self-reference, and what it means for a system to be grounded.
• Edwin Abbott Abbott, Flatland: A Romance of Many Dimensions (1884) — Creatures living in a two-dimensional world encounter beings from three dimensions; the logic of their world is consistent but radically limited. An allegory of the limits of formal systems.

10 Bibliography

11 Notes