The Unexpected Hanging Paradox

The positions divide on a prior question: is this a defect in the prisoner's reasoning, or a defect in the notion of a self-referential claim about knowledge?

Self-refutation (D. J. O'Connor, 1948). The original diagnosis. The announcement is pragmatically self-defeating: making it destroys the conditions under which it could be true, like saying "I am not speaking". Its fatal problem is empirical. The hanging happened, on Wednesday, and it was a surprise. The decree was true. A statement that turns out true cannot be self-refuting.

Invalid elimination (Quine, 1953). Set out above. The prisoner's argument requires that he retain knowledge of the decree throughout an argument whose whole purpose is to undermine it. Quine's four-case analysis of the last day shows he has smuggled in the assumption that he will know the outcome in advance, which is the very thing in dispute. Widely regarded as correct about the story as usually told, and widely regarded as insufficient.

Self-reference (Michael Scriven, 1951; R. Shaw, 1958). Shaw's contribution is the observation that the decree quantifies over what the prisoner will deduce from that decree. That is a fixed-point construction, not a prediction, and it places the puzzle in the Liar's family. Everything technical that followed builds on this identification.

Paradox regained (David Kaplan and Richard Montague, 1960). The decisive technical intervention. They construct the Designated Student paradox — a finite, fully formal version in which the elimination argument is valid and the contradiction is real. This kills the hope that the problem is merely the prisoner's sloppiness, and it forces the question of which epistemic principle must be abandoned.

Gödelisation (Frederic Fitch). Fitch builds the announcement as a self-referential sentence in the style of Gödel's diagonal lemma, so the decree literally asserts its own unknowability. His 1963 paper on value concepts proves the general result behind all of this: if every truth is knowable, then every truth is known — because "p, and p is unknown" is a truth that cannot be known. That theorem, credited by Fitch to an anonymous referee later identified as Alonzo Church, is the paradox's permanent residue.

Blindspots (Roy Sorensen, 1988). The reconciling move. The decree is consistent and true; it is simply not knowable by the prisoner. Everyone else in the prison knows it. The prisoner's own epistemic position is what makes it inaccessible to him, and there is no contradiction in a truth that one person cannot rationally accept. This explains how the judge can be right and the prisoner reasonable at the same time.

Two paradoxes, not one (Timothy Chow, 1998). Chow's survey sorts the literature and concludes that the epistemological version — about what a rational agent should believe, where retention of knowledge is the failing assumption — and the logical version — a self-referential sentence generating a genuine antinomy — are different problems that happen to share a story. Most apparent disagreements in the literature are two authors solving different ones.

The game theorists' version. Backward induction misbehaves elsewhere in exactly this way. The finitely repeated Prisoner's Dilemma unravels to universal defection from the last round backwards; Reinhard Selten's chain-store paradox (1978) and Robert Rosenthal's centipede game (1981) produce equally counterintuitive predictions by the same route. In every case the induction requires reasoning about behaviour at nodes that the induction has already declared unreachable, and in every case laboratory subjects decline to cooperate with the theory. The connection is structural, not decorative: the prisoner's Thursday-evening reasoning is a decision node that his own argument says will never be reached.

1943–1944. The paradox appears to have entered general circulation through a Swedish radio broadcast during the war, announcing a civil defence exercise in the coming week whose date would not be known in advance. The mathematician Lennart Ekbom noticed the logical problem and passed it to students in Stockholm, from where it spread by word of mouth. The account is reported by Martin Gardner and repeated by Timothy Chow; the broadcast itself has not been independently verified, so this origin should be read as well-attested tradition rather than documented fact.

1948. D. J. O'Connor publishes "Pragmatic Paradoxes" in Mind — the first appearance in print. His version is a military commander announcing a class A blackout that will occur in the coming week on a day the men cannot deduce in advance. O'Connor concludes the announcement is self-refuting.

1951–1952. Michael Scriven ("Paradoxical Announcements") and Paul Weiss ("The Prediction Paradox") continue the exchange in Mind, insisting the problem is genuine rather than a slip.

1953. Quine's "On a So-called Paradox" appears in Mind. Two pages, no formalism, and it sets the terms for everything that follows. The title carries the verdict.

1958. R. Shaw's "The Paradox of the Unexpected Examination" identifies the self-referential structure explicitly, moving the problem out of the psychology of prediction and into the logic of self-reference.

1960. Kaplan and Montague publish "A Paradox Regained" in the Notre Dame Journal of Formal Logic. The Designated Student construction shows the contradiction survives formalisation, and the title answers Quine directly.

1963–1964. Fitch's "A Logical Analysis of Some Value Concepts" (1963) proves the knowability paradox; his "Goedelized Formulation of the Prediction Paradox" applies the diagonal lemma to the announcement itself. The second paper is dated 1963 in some bibliographies and 1964 in others.

1963, 1969. Martin Gardner devotes a Scientific American column to the paradox and later makes it the title essay of The Unexpected Hanging and Other Mathematical Diversions. This fixes the hanging version as the canonical telling and delivers the problem to a general readership. Gardner's own assessment is that no proposed solution had yet commanded agreement.

1988. Roy Sorensen's Blindspots supplies the framework in which the decree can be true, publicly announced, and unknowable by its subject.

1998. Timothy Chow's survey in the American Mathematical Monthly maps five decades of proposals and argues that the field has been arguing about two different paradoxes. It remains the standard reference and the sensible place to start.

Since then the centre of gravity has moved into dynamic epistemic logic, where announcements that change the truth of what they announce are a studied class of object rather than an embarrassment.

It is the reason epistemic logic distinguishes operators from predicates. Kaplan and Montague's construction shows that treating "knows" as a predicate of sentences in a language capable of arithmetic makes contradiction available by diagonalisation. Modal epistemic logic — knowledge as an operator, with a hierarchy or a restricted language — is the standard response, and it exists in its present form partly because of this puzzle.

It gave us the knowability paradox. Fitch's theorem is not a curiosity: it is a standing obstacle for any anti-realist who wants to identify truth with what can in principle be known. If every truth is knowable, every truth is known, and almost nobody wants that. Verificationist theories of truth have had to be rewritten around it.

It shows that public announcements are not neutral. Telling someone something changes what is true about what they know, and some sentences become false by being announced. This sounds obvious and its consequences are not: a security policy, an audit schedule or an inspection regime whose contents are published can be defeated by the publication. Formal work on public announcement logic divides formulas into those that survive being announced and those that do not, and the judge's decree is the standard example of the second kind.

It is the clearest case of backward induction failing. Any process reasoned out from a terminal condition backwards inherits the prisoner's difficulty: the reasoning at each step is conditional on states the reasoning itself declares impossible. Economists working on the chain-store paradox and the centipede game arrived at the same structure independently, and the persistent gap between backward-induction predictions and experimental behaviour is one of the better documented failures in game theory.

It sets a real constraint on surprise. If you want an event to be genuinely unpredictable to a competent reasoner, you cannot get there by cleverness about scheduling — a sufficiently good reasoner will unravel any deterministic scheme whose rules are known. You need actual randomness. Surprise inspections, random drug testing, audit sampling and treaty verification regimes are all designed on this principle, and the paradox is a proof that the principle is not optional.

Randomised audit and inspection design. Tax authorities, safety inspectorates and arms-control regimes all face the judge's problem: they must announce that inspections will occur and keep the timing unpredictable. The solution adopted everywhere is publicly committed randomisation — announce the distribution, not the draw. That is a direct concession that no deterministic schedule can be both published and surprising.

Dynamic epistemic logic. Public announcement logic studies exactly what the judge did: an update operator $[\phi]$ that revises an agent's information state by the announcement of $\phi$. Sentences that are true before the announcement and false after are called unsuccessful, and Moore sentences of the form $p \wedge \neg Kp$ are the canonical case. The paradox is now a worked example in a textbook chapter rather than an open wound, though the underlying knower paradox is not thereby solved.

Security and unpredictability. Kerckhoffs's principle says a system must be secure even when everything about it except the key is public. The unexpected hanging is the epistemological argument for why that key has to be random: any structure an adversary can reconstruct from public rules is a structure they can run the elimination argument on.

Evaluating AI systems. A model that can infer when it is being tested can behave differently under evaluation, and publishing an evaluation protocol makes that inference easier. The mitigation proposed in practice — unannounced evaluations drawn at random, held-out sets never described in public — is the judge's problem taken seriously, and it inherits the judge's difficulty: the policy of randomising is itself public.

Deadlines and unravelling. Markets with a fixed terminal date unravel backwards — job offers made earlier and earlier, bids submitted at the last possible instant, cooperation collapsing in the final rounds. The remedy adopted in practice is to make the endpoint uncertain, which is the same fix in different clothes: remove the last day and the induction has nowhere to start.

Setting up the decree

Let the candidate days be $d_1, \dots, d_n$. Let $H_i$ mean the hanging occurs on $d_i$, with exactly one $H_i$ true. Let $K_i(\phi)$ mean the prisoner knows $\phi$ on the eve of $d_i$ — that is, after $d_1, \dots, d_{i-1}$ have passed without a hanging.

The judge's decree is

$$D ;\equiv; \bigvee_{i=1}^{n} \Big( H_i \wedge \neg K_i(H_i) \Big)$$

together with the exclusivity of the $H_i$. In words: the hanging falls on some day, and on the eve of that day the prisoner does not know it.

The backward induction

Base step ($i = n$). Suppose $\neg H_1 \wedge \cdots \wedge \neg H_{n-1}$ and suppose $K_n(D)$. From $D$ and the failure of the earlier days, $H_n$ follows; since the prisoner knows both, closure gives $K_n(H_n)$. But $D$ asserts $\neg K_n(H_n)$. Contradiction, so $\neg H_n$.

Inductive step. Assume $\neg H_j$ for every $j > i$. Suppose the prisoner survives to the eve of $d_i$ and $K_i(D)$ holds. Then $H_i$ follows, so $K_i(H_i)$, contradicting $D$. Hence $\neg H_i$.

Descending from $i = n$ to $i = 1$ gives $\neg H_i$ for every $i$, contradicting the first conjunct of $D$. So $D$ is false — and yet the judge carries it out.

The four assumptions the induction needs

The derivation is not principle-free. It requires, at minimum:

1. Retention. $K_i(D)$ for every $i$ — the prisoner still knows the decree on every eve.
2. Closure. $K_i(\phi)$ and $K_i(\phi \rightarrow \psi)$ jointly give $K_i(\psi)$.
3. Factivity. $K_i(\phi) \rightarrow \phi$.
4. Availability. The prisoner knows the conclusions of his own earlier reasoning, so that $\neg H_j$ for $j > i$ is available to him as knowledge at stage $i$, not merely as a truth.

Quine's objection is aimed at the interaction of (1) and (4). The step that eliminates $d_i$ uses $\neg H_{i+1}, \dots, \neg H_n$ as known. But those were derived from $K(D)$, and the argument's own conclusion is that $D$ is false — which by factivity means $D$ was never known. The prisoner is using $K(D)$ to derive $\neg D$, and then continuing to use $K(D)$.

Written out, what the argument actually establishes is not $\neg D$ but the conditional

$$K(D) ;\rightarrow; \neg D$$

By factivity $K(D) \rightarrow D$, so the two together give $\neg K(D)$. That is not a contradiction. It is the statement that the prisoner cannot know the decree — and it is entirely compatible with the decree being true. This is Sorensen's blindspot in one line: $D \wedge \neg K_{\text{prisoner}}(D)$ is consistent, so a proposition may be true, publicly announced, and rationally unbelievable by exactly one person in the room.

Quine's own presentation makes the same point through the last day's four cases:

	Prisoner knows the outcome	Prisoner does not
$H_n$	ruled out by $D$	consistent with $D$
$\neg H_n$	consistent	consistent

The prisoner eliminates the top-left cell and then proceeds as though the bottom row were empty. But the top-right cell is exactly what happens: he is hanged on the last day, unexpectedly, because he has convinced himself he cannot be.

The knower, and why Quine is not enough

Kaplan and Montague's move is to treat knowledge not as an operator on sentences but as a predicate $K(\ulcorner \phi \urcorner)$ applied to Gödel codes. Once knowledge is a predicate of the object language, the diagonal lemma applies: for any formula there is a sentence asserting that the formula holds of its own code.

Take $\kappa$ with

$$\vdash; \kappa ;\leftrightarrow; \neg K(\ulcorner \kappa \urcorner)$$

— "this sentence is not known". Assume the standard epistemic principles:

• (T) $K(\ulcorner \phi \urcorner) \rightarrow \phi$
• (K) $K(\ulcorner \phi \rightarrow \psi \urcorner) \rightarrow (K(\ulcorner \phi \urcorner) \rightarrow K(\ulcorner \psi \urcorner))$
• (N) if $\vdash \phi$ then $\vdash K(\ulcorner \phi \urcorner)$

Suppose $K(\ulcorner \kappa \urcorner)$. By (T), $\kappa$; by the fixed point, $\neg K(\ulcorner \kappa \urcorner)$ — contradiction. So $\neg K(\ulcorner \kappa \urcorner)$, hence $\kappa$, and since we have just proved $\kappa$, (N) gives $K(\ulcorner \kappa \urcorner)$. Contradiction again.

This is the Knower paradox. It is the epistemic twin of Montague's theorem and of Tarski's undefinability of truth, and its structure is Löb's: the system cannot consistently contain a factive predicate that applies to its own theorems. Fitch's Gödelized formulation writes the judge's decree as such a fixed point, so that the decree asserts its own unknowability directly. Kaplan and Montague's Designated Student paradox is the finite, weekday-shaped version of the same construction, and in it the elimination argument is valid.

The consequence is decisive for the classification of this entry. When the announcement is a genuine self-referential fixed point, no amount of care about the prisoner's psychology helps. One of (T), (K), (N) or the object-language treatment of knowledge must be given up — the same menu that Tarski's hierarchy, paracomplete logics and paraconsistent logics offer for the Liar.

Fitch's general theorem states the loss cleanly. Let $\Diamond$ be possibility and suppose every truth is knowable, $\phi \rightarrow \Diamond K\phi$. Apply it to $p \wedge \neg Kp$ for some unknown truth $p$. Then $\Diamond K(p \wedge \neg Kp)$. But distribution gives $Kp \wedge K\neg Kp$, and factivity turns the second conjunct into $\neg Kp$ — a contradiction, so $K(p \wedge \neg Kp)$ is impossible. Hence there is no unknown truth: if all truths are knowable, all truths are known. The judge's decree is one instance of this impossible conjunction, addressed to one prisoner.

The relation to backward induction generally

Strip out the epistemic vocabulary and the skeleton is this: a finite sequence of decision points, solved from the last backwards, where the solution at each point assumes the solutions at all later points are common knowledge, and where reaching a given point is evidence that some earlier assumption failed.

The finitely repeated Prisoner's Dilemma has exactly this shape, and unravels to defection in every round. Selten's chain-store paradox and Rosenthal's centipede game have it too, and produce predictions that experimental subjects contradict immediately and consistently. In every case the induction requires the players to condition on being at a node the induction says cannot be reached, and to keep believing in the rationality that got them there. The prisoner's Thursday evening is that node. His argument tells him he will never see it, so it tells him nothing about what to believe if he does.