The Gettier Problem

The Puzzle

Edmund Gettier's 1963 paper Is Justified True Belief Knowledge? is three pages long.¹ It presented two cases. Each case satisfies all three conditions of the classical analysis of knowledge as justified true belief (JTB): the agent has a belief, the belief is true, and the belief is justified. Yet in each case, most readers say the agent does not have knowledge.

The simplest version of the puzzle, in modern restatement:

A person looks at a clock that has been broken since exactly 12 hours ago. The clock reads 3:00. The actual time is 3:00. The person forms the belief that it is 3:00, on the basis of the clock reading.

The person has a true belief (it really is 3:00). The belief is justified (a clock reading is normally adequate evidence for the time). All three JTB conditions hold.

But did the person know it was 3:00?

Most readers say no. The clock has been broken for hours; that the time happens to match the broken clock's frozen reading is luck, not a route to knowledge. If the actual time were 3:01, the person would still believe 3:00, with exactly the same justification. The match between belief and truth is accidental. Knowledge, whatever else it is, does not work this way.

The classical JTB analysis, traceable to Plato and codified in modern form well before 1963, says that knowledge equals justified true belief. The 3:00 case satisfies JTB but seems not to be knowledge. Therefore JTB is insufficient: justification, truth, and belief together do not guarantee knowledge.

This is the Gettier problem. It is now standard equipment in any serious epistemology course. Half a century of work has not produced a consensus replacement.

This page assumes What Is Epistemology? and Validity vs Soundness. The latter is more relevant than it sounds: Gettier's cases all turn on the inferential structure of the justification, and being clear about which steps are doing what work is what makes the analysis tractable.

Gettier's Two Original Cases

Gettier's 1963 paper presents two cases, both involving a character named Smith. Both share the same structural form, which is what gives the paper its bite.

Case 1: The Job Application

Smith and Jones have applied for the same job. Smith has strong evidence for proposition (d):

(d) Jones is the man who will get the job, and Jones has ten coins in his pocket.

The evidence: the company president told Smith Jones would be selected, and Smith counted the coins in Jones's pocket ten minutes ago.

From (d), Smith deduces (e):

(e) The man who will get the job has ten coins in his pocket.

Smith accepts (e) on the strength of (d), to which he has strong evidence.

But unbeknownst to Smith, Smith himself will get the job. And, also unbeknownst to Smith, Smith himself has ten coins in his pocket. So (e) is true.

Now check the JTB conditions for Smith's belief in (e):

• True: yes (Smith will get the job; Smith has ten coins).
• Believed: yes (Smith deduced (e) from (d) and accepts it).
• Justified: yes (Smith's evidence for (d) is strong; deduction preserves justification).

All three conditions hold. But Smith does not know (e). His justification for (e) goes through (d), which is false: Jones is not the man who will get the job. (e) happens to be true for an entirely different reason than the one Smith's reasoning tracks.

Case 2: Smith and Brown in Barcelona

Smith has strong evidence for proposition (f):

(f) Jones owns a Ford.

The evidence: Smith has seen Jones drive a Ford for years, Jones has just driven Smith somewhere in a Ford, and so on.

Smith has another friend, Brown, of whose whereabouts he is wholly ignorant. Smith arbitrarily constructs three propositions:

(g) Either Jones owns a Ford, or Brown is in Boston. (h) Either Jones owns a Ford, or Brown is in Barcelona. (i) Either Jones owns a Ford, or Brown is in Brest-Litovsk.

By the rule of disjunction-introduction in classical logic (from $p$, infer $p \lor q$ for any $q$), each of (g), (h), (i) follows from (f). Smith deduces them and accepts them all.

Now suppose two things turn out to be true. First, Jones does not own a Ford; he had been borrowing one or had recently sold his. Second, by sheer coincidence, Brown happens to be in Barcelona.

Then (h) is true: Brown is in Barcelona makes the disjunction true.

Smith's belief in (h) is:

• True: yes (Brown is in fact in Barcelona).
• Believed: yes (Smith deduced and accepted (h)).
• Justified: yes (Smith's evidence for (f) is strong; valid logical inference preserves justification).

But Smith does not know (h). His justification was for the wrong disjunct. The disjunct Smith was justified in believing (Jones owns a Ford) is false. The disjunct that makes the disjunction true (Brown is in Barcelona) is something Smith has no evidence for whatever.

The Shared Structure

Both Gettier cases share a clear pattern.

The shared structure: the agent has a justified false belief somewhere in the justification chain. Logical inference from the false intermediate produces a conclusion. The conclusion happens to be true, but for reasons unrelated to the justification. The agent has a JTB, but the link between justification and truth is mediated by the false intermediate, and the actual truth-maker bypasses the justification entirely.

This is sometimes called the justified-true-belief-via-falsehood structure. The earliest response (Clark 1963, mentioned below) tried to fix the JTB analysis by simply requiring that no falsehood enter the justification chain. As we will see, this is a good first idea but not enough.

The Major Responses

The fifty years following Gettier 1963 produced an industry of proposed fixes. The major candidates, with the case each is designed to handle and the case that breaks it.

No False Lemmas (Clark 1963)

The first response, published the same year as Gettier's paper. Add a fourth condition to JTB:

(4) $S$'s justification for $p$ does not depend essentially on any false lemma.

The condition rules out both Gettier cases directly: in Case 1, the justification depends on the false lemma "Jones will get the job"; in Case 2, on the false lemma "Jones owns a Ford."

Where it breaks. Cases can be constructed where the justification does not pass through any explicit false intermediate, yet the same Gettier-style luck obtains. The classic such case is Goldman's 1976 fake-barn case (covered below): the agent's belief that "that is a barn" is justified by visual perception, with no intervening lemma, yet the case is plainly Gettiered. No-false-lemmas requires the intervening falsehood to be propositional, and many real Gettier cases lack any propositional falsehood the analysis can pin to.

Defeasibility (Lehrer & Paxson 1969)

A different fix.² Add a fourth condition that no truth, if added to the agent's evidence, would defeat the justification:

(4) There is no true proposition $d$ such that, if $d$ were added to $S$'s evidence, $S$ would no longer be justified in believing $p$.

In Case 1, the truth that "Smith will get the job" is a defeater: adding it to Smith's evidence would destroy his justification for (e) by pulling out the basis. The condition rules out the case correctly.

Where it breaks. Misleading defeaters: a true proposition that, if added to the evidence, would defeat the justification, but only because adding it would mislead the agent. Suppose Sue knows by ordinary perception that her friend Tom is in the room. Suppose, unbeknownst to Sue, an unreliable source has been spreading the rumor "Tom is on a flight to Mars," and adding the proposition "an unreliable source has been spreading rumors" to Sue's evidence would (irrationally) shake her perceptual belief. The proposition is true and defeats the justification, yet we would not deny Sue knows Tom is in the room. Patches to handle misleading defeaters multiply rapidly.

Causal Theory (Goldman 1967)

Earlier than Lehrer-Paxson, Alvin Goldman proposed:³

Knowledge requires that the fact $p$ causally connect, in an appropriate way, to the agent's belief that $p$.

In Case 2, the fact that makes (h) true (Brown's being in Barcelona) has no causal connection to Smith's belief that (h). Goldman's causal theory rules the case out.

Where it breaks. Mathematical and other non-causal knowledge is not handled cleanly. The fact that $\sqrt{2}$ is irrational does not causally produce my belief about it; abstract objects do not stand in causal relations. Yet I know $\sqrt{2}$ is irrational. A pure causal theory fails for the entire abstract domain.

Reliabilism (Goldman 1979)

Goldman's second move replaces the causal requirement with a reliability requirement:⁴

A belief is justified iff it is produced by a reliable belief-forming process.

Reliabilism is externalist: justification depends on facts the agent need not know. The Gettier cases break because Smith's justification for the true conclusion runs through a process (deduction from a false premise) that is not reliable for that conclusion. The reliable process is the one that would track the actual truth-maker, which Smith's reasoning does not engage.

Where it breaks. The generality problem: any belief is produced by many processes (visual perception of barn-shaped object, perception of building in countryside, perception under specific lighting conditions, etc.). Which process is the relevant one for assessing reliability? Different choices yield different verdicts, and there is no principled answer.

The demon-world problem: in a world systematically deceived by a Cartesian demon, the agent's belief-forming processes are systematically unreliable. Reliabilism judges all of the agent's beliefs unjustified. Yet many epistemologists think the agent in such a world is no less rational than we are; rationality should not depend on the agent's environment. This is the new evil demon objection due to Cohen and Lehrer 1983.

Sensitivity (Nozick 1981)

Robert Nozick proposed a counterfactual condition:⁵

$S$'s belief that $p$ is sensitive iff, were $p$ false, $S$ would not believe $p$.

In Case 1, were (e) false (the man getting the job had no coins, say), Smith would still believe (e) for the same reasons. The belief is insensitive to the truth of (e). Sensitivity rules out the case.

Where it breaks. Sensitivity has counterintuitive consequences for knowledge under closure. Suppose I know I have hands. Sensitivity requires that, were I to lack hands, I would not believe I have hands. But the closest possible world where I lack hands might still be one where I (as a brain in a vat) believe I have hands. Sensitivity then requires denying that I know I have hands, which is too skeptical. Nozick himself was willing to bite this bullet (he denies closure of knowledge under known entailment), but most subsequent epistemologists found this cost too high.

Safety (Sosa 1999, Williamson 2000)

Safety is the dual of sensitivity:⁶

$S$'s belief that $p$ is safe iff in nearby possible worlds where $S$ believes $p$, $p$ is true.

In Case 1, in nearby worlds where Smith forms the same belief in (e) on the same justification, (e) is generally false (because the underlying basis was already false in the actual world; small perturbations preserve it). The belief is unsafe.

Where it breaks. Defining "nearby" precisely. The definition smuggles in modal-similarity standards that are themselves contested. Safety also fails some intuitive cases: an agent in a world where most barns happen to be real (but barn-facades exist somewhere on the planet) seems to have a safe belief, even though we would not say she has knowledge if she happens to look at one of the few facades.

Knowledge-First (Williamson 2000)

Timothy Williamson's Knowledge and Its Limits makes the most radical move.⁷ The fifty-year search for the missing fourth condition has failed; perhaps it has failed because there is no missing fourth condition. Knowledge, Williamson argues, is conceptually primitive: it cannot be analyzed into more basic conditions because it is itself the most basic epistemic concept.

On Williamson's view, justification, evidence, and even belief should be analyzed in terms of knowledge, not the other way around. Evidence is knowledge. Belief is the disposition to act as if one knows. Justification is what one has when one is in a position to know. The analytic project of giving necessary and sufficient conditions for knowledge is misguided from the start.

Cost. Knowledge-first abandons a centuries-old methodological project. Many philosophers find this unsatisfying: an analysis is supposed to illuminate a concept, and refusing analysis looks like giving up. Williamson's response is that the alternative is not analysis but failed analysis; better to acknowledge knowledge as primitive than to keep producing baroque fourth conditions that fail on new cases.

The position has gained ground in the twenty-first century. It is now one of the major options. Whether it is the right option remains contested.

Goldman's Fake Barn Case

Worth a separate mention because it is the cleanest non-Gettier-1963 example of the same problem.

Henry is driving through the countryside. He looks at what appears to be an ordinary red barn and forms the belief "that is a barn." Unbeknownst to Henry, he is in fake-barn country: the locals have erected papier-mache barn-facades all along the highway, indistinguishable from real barns at his viewing angle. By coincidence, the one structure Henry happens to be looking at is a real barn.

Henry has a true belief (the structure is a real barn). His belief is justified by ordinary perception. He has no false intermediate. Yet most readers say Henry does not know "that is a barn" because he could so easily have been looking at a facade and would have believed the same thing.

The fake-barn case is a Gettier case without the false-lemma structure. It refutes Clark's no-false-lemmas patch directly. It also shows the problem is not specifically about deduction from a false premise; it is about the modal fragility of the connection between belief and truth. The agent's belief just happens to land on truth, and minor variations in the world would have produced the same belief about a falsehood.

This is the case sensitivity, safety, and reliabilism are designed to handle. None handles it perfectly.

What the Failure to Solve Gettier Tells Us

Half a century of failed analyses suggests something stronger than "we have not yet found the right fourth condition."

Two diagnostic possibilities.

First diagnosis: knowledge is genuinely unanalyzable. There is no clean reduction of knowledge to more basic concepts because knowledge is the most basic. Every proposed analysis has been broken because the project itself is incoherent.

Second diagnosis: knowledge is cluster-concept. There is no single set of necessary and sufficient conditions because knowledge attributions track several different things in different contexts: in some cases reliability, in some cases sensitivity, in some cases safety, in some cases the absence of misleading defeaters. The conditions interact and shift with context. The classical analysis tried to compress this into a single tidy formula and could not succeed.

Both diagnoses are live. Williamson defends the first; many virtue-epistemologists and contextualists endorse versions of the second. What is settled: the simple JTB-plus-fourth-condition project, fifty years on, has not converged. The Gettier problem is therefore not a curiosity. It is a substantive constraint on any theory of knowledge, and it is the reason modern epistemology looks radically different from its 1962 predecessor.

Connection to AI

The Gettier problem matters for current AI in concrete ways.

When a language model produces a true assertion, the assertion may be Gettiered in the same structural sense. The model's "justification" (the training-time pattern matching) may not track the truth-maker for the assertion. The assertion is true; the model produces it; the model's basis is structurally weak in a way that, on a slightly different input, would have produced a confident falsehood. This is the standard worry about hallucination with surface plausibility.

Reliabilism gives one diagnosis: a model is justified iff its belief-forming process is reliable for the relevant class of questions. For population statistics of major cities, the model is reliable; for population statistics of small towns, the model is not. Both classes can produce surface-plausible answers. The reliabilist judges them differently.

Safety gives another: a model's assertion is safe iff in nearby cases (similar prompts, similar contexts) it does not produce false answers. Models that are highly peaked in their training distribution produce safe answers within that distribution and unsafe answers outside it.

Williamson's knowledge-first frame applied to AI: ask not "does the model know" but "is the model in a position to assert without epistemic violation." The latter is the practical question and the one product teams already care about under the names calibration and uncertainty quantification. The Gettier-style worry is built into the empirical question of when calibration holds.

This is the synthesis taken up at Knowledge, Justification, and LLMs.

Common Confusions

Confusion 1: Gettier cases show JTB is wrong. Not quite. Gettier cases show JTB is insufficient. The three classical conditions (truth, belief, justification) are still necessary on the standard analyses. What Gettier showed is that they are not jointly sufficient: a fourth condition is needed, and finding it has been hard.

Confusion 2: the cases are contrived and so do not matter. They are constructed, not contrived. The structure they exploit (justification reaching truth via accident) appears in real cases all the time. Goldman's fake-barn case is constructed too, but the structure shows up whenever ordinary perceptual evidence happens to be unreliable in a way the agent cannot detect from the inside. Constructed cases are common in philosophy because they isolate the relevant variable.

Confusion 3: Gettier problems are about luck. Sort of. They are about a specific kind of luck: the kind where the connection between justification and truth is accidental. Some kinds of luck are compatible with knowledge (a lucky catch of a fly ball does not prevent the player from knowing he caught it). The Gettier-relevant kind is epistemic luck: luck in the relation between belief and truth, mediated by justification.

Confusion 4: the problem is solvable; we just need a sharper fourth condition. Maybe. The fact that fifty years of professional epistemology has produced no consensus suggests something stronger may be wrong with the analytic project itself. Williamson's knowledge-first move is the most prominent diagnosis. Whether his diagnosis is correct is an open question, but the empirical track record of analytic responses is sobering.

Two Exercises

Exercise 1. For each of the following cases, decide whether (a) the agent has a justified true belief in the relevant proposition, (b) the agent intuitively counts as having knowledge, and (c) which post-Gettier framework best explains the verdict in (b). Give brief reasoning.

(i) The reliable thermometer in a faulty house. Anna's thermometer reads 72°F. The thermometer is well-made and accurate. The house's air-conditioning is broken; the indoor temperature varies wildly from minute to minute. By coincidence, the indoor temperature happens to be exactly 72°F at the moment Anna looks at the thermometer. Anna forms the belief "the indoor temperature is 72°F."

(ii) The newspaper that misprints by coincidence. Bob reads in his newspaper that the home team won 4-3. The newspaper has typos in this issue: the actual score was 5-2. By coincidence, the home team did win, but with score 5-2, not 4-3. Bob's belief: "the home team won" (not the score, just the winner).

(iii) The trustworthy testimony with a fake context. Carol asks her friend Dan, a competent botanist, to identify a plant. Dan says it is a Quercus alba (white oak). The plant is in fact Quercus alba. However, Dan is in a sleep-deprived state and was guessing without proper care; he just happens to have guessed right. Carol forms the belief that the plant is Quercus alba.

Exercise 2. Construct your own original Gettier case. Specifically, design a scenario in which:

(a) The agent's belief in $p$ is true. (b) The agent has good evidence and reaches $p$ by valid reasoning. (c) The truth-maker for $p$ is not what the agent's reasoning tracks. (d) Most readers will have the intuition that the agent does not have knowledge.

Then identify which of the post-Gettier frameworks (no false lemmas, reliabilism, safety, knowledge-first) most cleanly explains why your case is not knowledge.

Sketch of answers

Answer 1.

(i) Anna and the thermometer. (a) Yes: justified (good thermometer), true (it really is 72°F), believed. (b) No: the thermometer is reliable but the system (broken AC plus accurate thermometer) is not, and the actual temperature could easily have differed from the reading. The reading and the truth coincide by accident. (c) Safety / reliabilism: in nearby cases, the thermometer would still read 72°F but the actual temperature would be different. The belief is unsafe.

(ii) Bob and the newspaper. (a) Yes: justified (newspapers are normally reliable), true (the home team did win), believed. (b) Most readers say no knowledge: Bob's basis (the misprinted score) is false, and the truth of the disjunctive proposition "the home team won" is mediated by the actual score, which Bob's evidence does not track. (c) No false lemmas (Clark) handles this directly: the inferential chain passes through the false intermediate "the score was 4-3."

(iii) Carol and the botanist. (a) Yes: testimony from a competent expert is normally good evidence; the plant really is Quercus alba; Carol believes it. (b) Most readers say no knowledge: Dan's actual cognitive process was unreliable (sleep-deprived guessing). The match is accidental relative to Dan's cognitive state. (c) Reliabilism: the belief-forming process Carol depends on (Dan's testimony based on competent identification) is not the process that actually produced Dan's testimony (sleep-deprived guessing). The reliable process Carol thought she was relying on was not in fact engaged.

Answer 2. Here is one possible answer. A reader's own answer should be different.

The simulated chess engine. Eve is a chess player. She plays a game with a chess engine and reaches a position. She uses the engine's evaluation, displayed on screen, to form the belief "this position is winning for Black." The engine's evaluation is reliable in 99 percent of cases. Today, the engine has a software bug that makes its evaluation random for this specific kind of position. By coincidence, the engine outputs "Black is winning," and the position is in fact winning for Black (verifiable by tablebases, say).

Eve's belief is true (the position is winning for Black), justified (the engine is reliable), and arrived at without false intermediate. But Eve does not know the position is winning, because the engine's actual basis (random output today) does not track the truth-maker (the underlying tablebase fact).

Best framework explaining the case: reliabilism with generality problem. The relevant process question is "engine evaluation in general" vs "engine evaluation of this specific position type today." Under the first description the process is reliable (so the belief is justified by Goldman 1979), but under the second it is not. The case shows why the generality problem matters: the same belief can be classified as justified or unjustified depending on which process is the reference class. Most readers' intuition that Eve does not know suggests the relevant process is the narrow one (engine evaluation today, this kind of position), and reliabilism's verdict tracks that.

A second valid framework: safety. In nearby cases (similar prompts to the buggy engine today), the engine would output many wrong evaluations. Eve's belief is unsafe.

The exercise illustrates that constructing Gettier cases is not hard once you know the structural form. The hard work is figuring out which framework explains why the case fails to be knowledge in a way that generalizes correctly across many cases.

Where the Gettier Problem Lives in Practice

Three concrete uses.

Statistical evidence and the legal system. A defendant could be convicted on probabilistic evidence alone (a population-level base rate) that produces a justified true belief in the verdict. Most legal systems require something more: evidence that individuates the defendant. The intuition tracks the Gettier worry: probabilistic evidence makes the belief justified and (often) true, but the connection between justification and truth is the wrong kind. Judith Jarvis Thomson 1986 and Enoch, Spectre, and Fisher 2012 develop this connection.

AI calibration. A model can be calibrated in aggregate (90 percent of its 90-percent-confidence assertions are true) without being calibrated per case: the truth-makers for individual confident assertions may bear only loose relation to the model's basis. A Gettier-style worry arises whenever model confidence is treated as a proxy for knowledge but the underlying connection is statistical rather than causal or systematic.

Scientific replication. A scientific finding can be true (the underlying claim about the world holds) and supported by valid statistical analysis (justification) yet still fail to count as scientific knowledge if the supporting evidence does not connect appropriately to the finding. The replication crisis in psychology and biomedical science is in part a Gettier-style concern at scale: too many findings turned out to be "true (?) plus justified (statistically) but not knowledge" because the connection between the original justification and the actual phenomenon was fragile.

Prerequisites and Next Pages

• Prerequisites: What Is Epistemology?, Validity vs Soundness.
• Next: Induction and Hume's Problem, the foundational challenge to inductive justification.
• Next: Bayesian Epistemology, the dominant formal framework for graded belief.
• Forward synthesis: Knowledge, Justification, and LLMs.

References

Primary text:

• Gettier, Edmund L. "Is Justified True Belief Knowledge?" Analysis 23, no. 6 (1963): 121-123. Three pages, openly available, the founding document.

Major responses:

• Clark, Michael. "Knowledge and Grounds: A Comment on Mr. Gettier's Paper." Analysis 24 (1963): 46-48. The earliest no-false-lemmas response, published the same year as Gettier.
• Goldman, Alvin I. "A Causal Theory of Knowing." Journal of Philosophy 64, no. 12 (1967): 357-372.
• Goldman, Alvin I. "Discrimination and Perceptual Knowledge." Journal of Philosophy 73, no. 20 (1976): 771-791. The fake-barn case.
• Lehrer, Keith, and Thomas Paxson Jr. "Knowledge: Undefeated Justified True Belief." Journal of Philosophy 66, no. 8 (1969): 225-237.
• Goldman, Alvin I. "What Is Justified Belief?" In Justification and Knowledge, ed. George Pappas, Reidel, 1979.
• Nozick, Robert. Philosophical Explanations. Harvard, 1981. Chapter 3.
• Sosa, Ernest. "How to Defeat Opposition to Moore." Philosophical Perspectives 13 (1999): 141-153.
• Williamson, Timothy. Knowledge and Its Limits. Oxford, 2000.

Modern reference:

• Audi, Robert. Epistemology: A Contemporary Introduction. Routledge, 3rd ed., 2010. Chapter 9 covers the post-Gettier responses with examples.
• Steup, Matthias, and Ram Neta, eds. The Routledge Companion to Epistemology. Routledge, 2014. Includes long survey articles on each major post-Gettier framework.
• Ichikawa, Jonathan Jenkins, and Matthias Steup. "The Analysis of Knowledge." Stanford Encyclopedia of Philosophy. The cleanest single survey of the post-Gettier landscape.

Stanford Encyclopedia entries (link, do not paraphrase):

• "The Analysis of Knowledge."
• "Reliabilist Epistemology."
• "Knowledge How" (cognate distinction).
• "Williamson, Timothy" (covers knowledge-first epistemology).