The Sorites Paradox

The four surviving families of answer split on which of the three exits they take. What follows is who says what, and what each has to give up.

Epistemicism: the line is real and you cannot find it

Timothy Williamson (Vagueness, 1994) takes the option nearly everyone else rules out on sight. There is an exact number of grains at which the heap stops being a heap. There is a last hair. "Heap" draws a perfectly sharp boundary; we simply do not know where it is, and we cannot know.

The attraction is enormous: classical logic survives untouched, every sentence is true or false, and the paradox is dissolved by rejecting the tolerance premise outright. The cost is a claim that sounds like magic. If nobody has ever fixed the boundary, what fixes it?

Williamson's answer is that meaning is determined by the whole tangle of how a word is used across a community, and that this tangle settles a boundary in the way a huge computation settles an answer nobody has computed. Our ignorance is then a principled consequence of how knowledge works: knowledge requires a margin for error, and there is no margin available at a boundary. Roy Sorensen defends a version of the same position.

The recurring objection is not that epistemicism is inconsistent — it is not — but that it buys logical tidiness with a metaphysical promissory note.

Supervaluationism: true on every way of sharpening it

Kit Fine ("Vagueness, Truth and Logic", 1975) offers the most popular middle road. Take a vague language and consider all the admissible ways of making it precise — every way of stipulating a cut-off that respects the clear cases and the relations between terms. A sentence is true if it comes out true on every sharpening, false if false on every one, and neither if the sharpenings disagree.

This has a lovely consequence. "There is a number of grains at which the heap stops" is true, because it is true on every sharpening. But no particular statement of the form "the heap stops at 1,847 grains" is true, because each is true on only one sharpening. You get the existence of a boundary without any particular boundary. Classical tautologies survive; bivalence does not.

Two costs. First, "admissible" is itself vague, so the vagueness reappears one level up — this is the higher-order vagueness problem, and it is the standard weapon against every non-epistemicist theory. Second, familiar reasoning patterns like reductio and argument by cases stop working once you add an operator for "definitely". Fine himself later moved away from the view.

Degree theory and fuzzy logic: truth by amounts

Joseph Goguen ("The Logic of Inexact Concepts", 1969), building on Lotfi Zadeh's fuzzy sets (1965), replaces true-or-false with a number between 0 and 1. Ten thousand grains: truth value 1. One grain: truth value 0. Five thousand grains: somewhere in between.

Each step of the sorites is then almost true rather than true, and the tiny shortfalls accumulate. Run ten thousand nearly-perfect inferences in a row and you can end up with something worth nothing at all. The argument is not valid after all, and it fails in a way that explains exactly why it looked so convincing.

The complaint is artificial precision. Saying that "this is a heap" is true to degree 0.5 rather than 0.51 replaces one sharp boundary with an infinite continuum of them. Where did those decimals come from? Higher-order vagueness bites here too.

Nihilism: there are no heaps

Peter Unger ("There Are No Ordinary Things", 1979) does the thing that makes people put the paper down. He accepts the tolerance premise, accepts that a single grain is not a heap, and runs the argument backwards: no heaps. Then he notices that the same series can be built for stone, table, cloud and chair, and concludes that ordinary things are incoherent posits. In a companion paper the argument reaches people.

Nobody believes this, including, arguably, Unger, who later disowned it. Its use is as a measuring stick: it shows precisely what the price of keeping both the premises and the logic is, and it makes every other theory look reasonable by comparison.

The rest of the field

Contextualism (Diana Raffman and others) holds that the extension of "heap" shifts as attention moves along the series, so the reasoner never faces two adjacent cases under one fixed standard. Subvaluationism (Dominic Hyde) is supervaluationism's mirror image: borderline sentences are both true and false, which requires a paraconsistent logic. Incoherentism holds that our rules for "heap" are jointly unsatisfiable and that this is a fact about us, not the world.

What no one has produced is a theory that keeps classical logic, keeps bivalence, keeps tolerance, and keeps the heaps. The consensus is only about the shape of the bill, not who should pay it.

Fourth century BCE. Diogenes Laertius (II.108) credits Eubulides of Miletus with seven puzzles, among them the Liar, the Bald Man (phalakros) and the Heap (sōritēs, from sōros, a heap). Eubulides belonged to the Megarian school and was a contemporary and critic of Aristotle. Nothing he wrote survives, so the exact wording he used is unrecoverable; what survives is the name and the shape. The puzzle was almost certainly a weapon in live debate rather than a private worry — the Megarians used it to embarrass opponents who claimed to possess knowledge.

Third century BCE. The Stoics took it seriously enough to develop a doctrine of response. Chrysippus advised the sage to "fall silent" — to stop answering as the series approaches the unclear cases, rather than assert something false. Cicero reports the strategy in the Academica (II.92–94) and treats it as an evasion: the Academic sceptics regarded the sorites as evidence that the Stoic theory of secure comprehension could not be sustained. The debate mattered because the Stoics had staked their epistemology on the existence of self-certifying impressions, and the heap is precisely a case where impressions do not certify anything.

Second century CE. Galen, in On Medical Experience (chapters 16–17), gives the fullest ancient treatment and the most exasperated. His objection is professional: medicine builds knowledge by accumulating cases, and if no single additional case can convert insufficient evidence into sufficient evidence, then no amount of experience ever amounts to knowledge. He describes being asked, one grain at a time, where the heap begins, and refusing to play.

Medieval period. Vagueness lived on in the insolubilia and in scholastic discussions of latitudo formarum — the intension and remission of qualities — but the sorites was mostly treated as a sophism to be classified rather than a problem to be solved.

1923. Bertrand Russell's paper "Vagueness" restarts the modern discussion. He argues that vagueness is a feature of representations rather than of the world, and that logic applies strictly only to precise languages — a position that concedes much of what later theorists try to save.

1937. Max Black's "Vagueness: An Exercise in Logical Analysis" introduces consistency profiles and the systematic study of borderline cases, moving the topic from a curiosity to a technical subject.

1965–1969. Lotfi Zadeh's fuzzy sets and Joseph Goguen's logic of inexact concepts supply degrees of truth. The engineering applications arrive long before philosophical agreement does.

1975. Kit Fine's "Vagueness, Truth and Logic" gives supervaluationism its mature form, including penumbral connection. Michael Dummett's "Wang's Paradox", published the same year, argues that observational predicates are inherently tolerant and therefore that our concepts are genuinely incoherent.

1979. Peter Unger publishes the nihilist reading, and the field acquires its outer boundary.

1994. Timothy Williamson's Vagueness makes epistemicism a serious contender rather than a curiosity, principally by supplying the margin-for-error account of why the boundary is unknowable. Since then the literature has been an argument between four camps rather than a search for a fifth.

The sorites is not a puzzle about sand. It is the discovery that ordinary language and classical logic do not fit together, and it is the oldest unresolved problem in either subject.

It is why logic has alternatives. Multi-valued logics, supervaluation semantics, paraconsistent logics and fuzzy set theory were all developed or sharpened with vagueness as a target. If classical logic had accommodated vagueness, several twentieth-century research programmes would not exist.

It explains why law is written the way it is. Statutes speak of "reasonable care", "undue delay" and "substantial harm" — irreducibly vague, because a precise rule would be gamed and would produce absurd results at the margin. But courts must return a verdict, so the vagueness is resolved case by case rather than in the text. Legal theorists distinguish the core of a rule from its penumbra, borrowing the vocabulary directly from this debate. Every bright-line rule in a statute book — the drinking age, the speed limit, the threshold for a capital gain — is a decision to bolt a sharp edge onto something that has none, and the arbitrariness is not a drafting failure but the price of having enforceable rules at all.

It sets the terms of arguments that will not otherwise end. When a foetus becomes a person, when a patient becomes dead, when a country becomes a democracy, when a model becomes dangerous: each is a sorites series with a decision attached. Recognising the structure does not settle the argument, but it does rule out one bad move — the demand that an opponent name the exact grain. Nobody can name the grain. That is the point of the paradox, not a weakness in the opponent.

It clarifies what tolerance costs. The reason "heap" is useful is exactly the reason it is paradoxical. A word learnable from examples must be insensitive to tiny differences, and a word insensitive to tiny differences cannot have a sharp extension. Usefulness and logical good behaviour are in tension, and natural language chose usefulness.

Machine learning thresholds. A classifier outputs a continuous score and a threshold converts it to a decision. The threshold is a bolted-on sharp edge, and every dispute about false positives versus false negatives is a dispute about where to bolt it. Calibration curves are, formally, degree-theoretic membership functions; the philosophical objection to fuzzy logic — that the exact decimals are unmotivated — is the practical objection to threshold tuning.

Content moderation. "Hate speech", "misinformation" and "harassment" are tolerant predicates enforced at scale by systems that must return a binary verdict millions of times a day. Written policies function as attempted precisifications; appeal processes are the forced march, run in public, by people who did not consent to the experiment.

Fuzzy control. Zadeh's sets found their commercial home in controllers for trains, cameras and washing machines, where rules like "if the load is somewhat heavy, spin somewhat faster" outperform sharp thresholds because real inputs are noisy. The philosophy remains contested; the engineering shipped decades ago.

Clinical cut-offs. Hypertension, obesity, anaemia and diabetes are diagnosed against numerical thresholds in continuous distributions, and moving a threshold by a small amount reclassifies millions of people overnight without any of them changing. The recurring controversies over redefining these boundaries are sorites arguments conducted with public health consequences.

Species and genomes. Biological kinds shade into one another across populations and across evolutionary time; the ring species that interbreed at every adjacent step but not at the ends of the ring is a sorites series with feathers on. Taxonomy handles it the way the law handles reasonableness — by stipulating, and then arguing about the stipulation.

Fix a vague predicate $F$ and a sorites series $a_0, a_1, \dots, a_n$: a finite sequence of cases in which $a_0$ is a clear $F$, $a_n$ is a clear non-$F$, and adjacent members differ only marginally in whatever respect $F$ is sensitive to. For the heap, $a_i$ is a pile of $n - i$ grains.

The tolerance principle

The intuition that one grain never matters is the tolerance principle: for a small enough difference $\delta$ in the relevant dimension,

$$\forall x, \forall y ;\big[ \left| x - y \right| \le \delta ;\rightarrow; (Fx \leftrightarrow Fy) \big]$$

Restricted to the series, this is

$$\text{(T)} \qquad \forall i < n ;(Fa_i \rightarrow Fa_{i+1})$$

Michael Dummett's formulation makes the source of tolerance explicit: observational and communally taught predicates must be tolerant, because competence with them is acquired from clear cases and preserved under indiscriminable change. If $F$ were intolerant, two indistinguishable objects could differ in $F$, and nothing in how we learned $F$ could account for the difference.

Three ways to state the argument

The three standard presentations are classically equivalent but come apart under non-classical logics, which is why the choice of presentation matters.

Conditional form. Premises $Fa_0$ and the $n$ conditionals $Fa_i \rightarrow Fa_{i+1}$; conclusion $Fa_n$ by $n$ applications of modus ponens. Only modus ponens is used.

Inductive form. Premises $Fa_0$ and (T); conclusion $\forall i \le n, Fa_i$ by mathematical induction on $i$. Here (T) is the induction step and $Fa_0$ the base case; the derivation is the ordinary induction schema

$$\big[ Fa_0 \wedge \forall i (Fa_i \rightarrow Fa_{i+1}) \big] \rightarrow \forall i, Fa_i$$

Line-drawing form. The denial that a sharp cut-off exists:

$$\text{(L)} \qquad \neg \exists i ,(Fa_i \wedge \neg Fa_{i+1})$$

In classical logic (T) and (L) are interderivable, since $\neg(p \wedge \neg q)$ is equivalent to $p \rightarrow q$. This equivalence is the hinge of the whole debate: it is what forces anyone who denies tolerance to assert a sharp cut-off, and it is what several non-classical treatments break.

Because $\neg Fa_n$ is as secure as $Fa_0$, the same premises also run backwards by modus tollens, delivering $\neg Fa_0$. The paradox therefore has two conclusions, each contradicting a datum, and no asymmetry to exploit.

How each position formalises its answer

Epistemicism. Rejects (T) and (L); asserts $\exists i (Fa_i \wedge \neg Fa_{i+1})$. Logic and bivalence are untouched. The explanatory burden shifts to why the witness is unknowable, discharged by a margin-for-error principle: knowledge of $Fa_i$ requires $F$ to hold in all cases sufficiently similar to $a_i$, so

$$K(Fa_i) \rightarrow Fa_{i+1}$$

Near the cut-off this is unsatisfiable, so the boundary is knowledge-resistant by construction. Note the contrast: the epistemicist's $K(Fa_i) \rightarrow Fa_{i+1}$ is a claim about knowledge, whereas the sorites premise $Fa_i \rightarrow Fa_{i+1}$ is a claim about truth. Williamson's thesis is that speakers systematically mistake the first for the second.

Supervaluationism. A precisification $s$ maps $F$ to a classical extension, subject to admissibility constraints: clear cases keep their verdicts, and penumbral connections are preserved — if $a_i$ is redder than $a_j$, no admissible $s$ makes $a_j$ red and $a_i$ not. Then

$$\phi \text{ is } \textbf{supertrue} \iff s \models \phi \text{ for every admissible } s$$

Now $\exists i (Fa_i \wedge \neg Fa_{i+1})$ is supertrue, since every precisification supplies a witness. But for each particular $k$, $Fa_k \wedge \neg Fa_{k+1}$ is neither supertrue nor superfalse, because the witness moves between precisifications. A true existential with no true instance is exactly the combination the paradox needs, and it comes at the price of bivalence.

(T) comes out superfalse, so the argument is unsound rather than invalid — the logic is preserved at the level of which sentences are valid. The known casualty is the metarules. Adding a determinacy operator $D$, with $D\phi$ supertrue iff $\phi$ is supertrue, one gets $\phi \models D\phi$ but not $\models \phi \rightarrow D\phi$, so conditional proof fails; contraposition, reductio and argument by cases fail with it. And $D$ reintroduces the problem: "definitely a heap" is itself tolerant, generating a second-order sorites over $DFa_i$, and a third over $DDFa_i$, without end.

Degree theory. Let $v(Fa_i) \in [0,1]$ decrease monotonically with $v(Fa_0) = 1$ and $v(Fa_n) = 0$; take the simplest case $v(Fa_i) = 1 - i/n$. Using the Łukasiewicz conditional

$$v(\phi \rightarrow \psi) = \min!\big(1,; 1 - v(\phi) + v(\psi)\big)$$

each premise has value $v(Fa_i \rightarrow Fa_{i+1}) = 1 - 1/n$: very nearly true, which is why the premises are so persuasive. But validity in this setting means preserving degree, and iterated modus ponens does not. With the Łukasiewicz rule, $n$ chained inferences from premises of value $1 - \epsilon$ bound the conclusion only by

$$v(Fa_n) \ge v(Fa_0) - n\epsilon$$

With $\epsilon = 1/n$ the bound is $0$: the argument is invalid, and invalid by exactly the accumulated margin that made each step look free. Fuzzy logic thereby explains the seductiveness of the sorites, not merely its failure.

The standing objection is that $v$ is a real-valued function. Asking why $v(Fa_{5000})$ is $0.5$ rather than $0.501$ is the sharp-boundary question returning in continuous dress, multiplied.

Nihilism. Keeps classical logic and keeps (T), and takes $\neg Fa_n$ as the fixed point rather than $Fa_0$. From (T) and $\neg Fa_n$, $n$ applications of modus tollens yield $\neg Fa_0$. Generalising over predicates for which sorites series can be constructed — which is nearly all sortal predicates for medium-sized objects — Unger concludes the extensions are empty. The position is formally impeccable and universally rejected, which is a useful demonstration that formal impeccability is not the binding constraint here.

What any solution must do

A solution has to say which of the following it denies, and pay the associated bill:

Denies	Position	Bill
Tolerance (T)	Epistemicism	An unknowable sharp boundary fixed by use
Bivalence	Supervaluationism	Metarule failures; higher-order vagueness
Classical validity	Degree theory	Artificial numerical precision
$Fa_0$	Nihilism	There are no heaps, stones or people
Nothing	—	Contradiction

Higher-order vagueness is the test that keeps failing candidates. Any theory that introduces a new category — borderline, indeterminate, degree $0.5$ — must say where that category begins, and the question is as hard as the one it replaced. Epistemicism's advantage is that it never introduces a new category. That is also the reason people distrust it.