Knowledge as Justified True Belief, and the Post-Gettier Replies

What It Is

The classical analysis of propositional knowledge, traceable through Plato's Theaetetus and articulated as a definition by Roderick Chisholm and A.J. Ayer in the mid-twentieth century, holds that a subject $S$ knows a proposition $p$ just in case three conditions are jointly met:

1. $p$ is true.
2. $S$ believes that $p$.
3. $S$ is justified in believing that $p$.

This is the JTB analysis: knowledge as justified true belief. The three clauses encode three intuitions. Truth: knowledge is success-oriented; you cannot know what is false. Belief: knowledge requires a doxastic state; you cannot know what you do not believe. Justification: knowledge is more than a lucky guess; you must have some standing to hold the belief.

For roughly two thousand years the structure of this definition was taken to be uncontroversial; debate centered on what justification amounts to, not on whether the three conditions together suffice for knowledge. In 1963 Edmund Gettier published a three-page paper¹ that constructed two short counterexamples showing the conjunction is not sufficient. The post-Gettier literature consists of a half-century of attempts to repair the analysis or replace it.

This page assumes What Is Epistemology? and gives the post-Gettier reply space at depth. The Gettier problem page covers the cases themselves; for the Bayesian formulation of justification see Bayesian Epistemology.

The JTB Analysis in Detail

The truth condition

The truth condition is uncontroversial in the contemporary literature. If $p$ is false, then no agent knows $p$, regardless of how confident or well-supported the belief is. An agent in 1700 who believed phlogiston theory with all the justification then available did not know phlogiston theory; the theory was false. The truth condition is what makes knowledge a "success term": to know is to get the world right.

The belief condition

The belief condition is mildly controversial. Knowledge clearly requires some doxastic relation to the proposition; you cannot know something you have never entertained. Whether the required state is full belief, high credence, or something weaker is the question. On most formulations the condition is liberal: any state of taking $p$ to be the case suffices. On stronger formulations the condition is full belief; an agent with 0.6 credence in $p$ does not, on this stronger reading, believe $p$ and therefore does not know $p$ even if $p$ is true and the agent's evidence supports it. The Bayesian tradition handles the issue by replacing the belief condition with a credence threshold or by working entirely in credence space (see Bayesian Epistemology).

The justification condition

The justification condition is the load-bearing one. Two large traditions divide on what it requires.

Internalism holds that justification is determined by factors internal to the subject's mental life: evidence the subject possesses, beliefs the subject has access to on reflection, the apparent fit between belief and evidence. The motivating intuition is that two subjects with identical mental states must be equally justified in their beliefs, regardless of how the world outside their head is arranged.

Externalism holds that justification depends on factors external to the subject's mental life: facts about the reliability of the belief-forming process, the actual track record of the agent's faculties, or causal connections to the truth-maker for $p$. The motivating intuition is that justification is a truth-conducive property; whatever justifies a belief should make the belief likely to be true, and likelihood-of-truth is partly a world-side matter.

The internalism / externalism axis cuts across the post-Gettier replies. The no-false-lemmas patch is naturally internalist; reliabilism is paradigm externalist. Contextualism cuts orthogonally.

Gettier's Counterexamples in Structure

Gettier's original cases share a structure that almost every later counterexample preserves. A subject $S$ has a justified false belief $F$. From $F$ the subject competently deduces a further proposition $p$, which happens to be true for reasons unrelated to $F$. The subject thereby has a justified true belief in $p$ that does not count as knowledge, because the route to $p$ ran through a false premise.

Gettier's Smith and Jones case: Smith is justified in believing "Jones owns a Ford" (on the basis of strong evidence about Jones's car-purchasing behavior). Smith deduces "Either Jones owns a Ford or Brown is in Barcelona", a disjunction. Smith does not know where Brown is; he picked Barcelona at random. As it turns out, Jones does not own a Ford (the evidence misled), but Brown happens to be in Barcelona. The disjunction is true, Smith is justified in believing it, and Smith believes it. But Smith does not know the disjunction; the truth of the disjunction has nothing to do with what justified Smith's belief.

The diagnostic phrase: epistemic luck. Gettier cases are cases where the agent's belief is true, but the agent reached truth by luck given how the justification routed through the false lemma. JTB plus a no-luck condition would suffice; the post-Gettier program is the search for that condition.

The Post-Gettier Reply Space

Reply 1: no-false-lemmas (Clark 1963)

Michael Clark's proposal, published months after Gettier's paper, was the simplest patch.² Add a fourth clause: $S$ knows $p$ iff (i)–(iii) hold AND (iv) $S$'s reasoning to $p$ does not essentially depend on any false belief.

The Smith case is blocked: Smith's reasoning depends essentially on the false belief "Jones owns a Ford." Adding clause (iv) removes the case from the extension of knowledge.

The reply works for the original Gettier cases, but a follow-up case set due to Carl Ginet and Alvin Goldman defeats it. The Fake Barn County case: Henry is driving through a region of which he is unaware that the locals have built numerous extremely convincing barn facades; only one structure in the area is a real barn. Henry looks out the window and sees the one real barn, forms the justified true belief "that is a barn," and reasons no false lemmas. He still does not know, because had he looked at any of the other thousand structures he would have formed an equally justified but false belief. The no-false-lemmas clause does not address this.

Reply 2: defeasibility (Lehrer and Paxson 1969)

Keith Lehrer and Thomas Paxson proposed adding a defeasibility clause: $S$ knows $p$ iff JTB holds AND there is no true proposition $d$ such that, if $S$ believed $d$, $S$ would no longer be justified in believing $p$.³

The Gettier case is blocked: the true proposition "Jones does not own a Ford" would defeat Smith's justification.

The defeasibility reply runs into the misleading defeater problem. Some true propositions, if added to the agent's belief set, would mislead the agent into giving up justified beliefs. If "Tom Grabit's twin Tim was on campus today" is true, an agent who saw Tom take a book might lose justification in "Tom took the book" even if there is no actual twin and Tom did take it. The defeasibility reply requires a distinction between genuine and misleading defeaters, and articulating that distinction without circularity is hard.

Reply 3: reliabilism (Goldman 1979)

Alvin Goldman proposed replacing the justification condition with a reliability condition.⁴ $S$ knows $p$ iff (i) $p$ is true, (ii) $S$ believes $p$, and (iii') $S$'s belief in $p$ was produced by a reliable belief-forming process (a process that tends to produce true beliefs).

Reliabilism is the paradigm externalist reply. The justification condition becomes a world-side fact about the actual reliability of the cognitive process, not a fact about the subject's evidence-recognition. The Gettier case is blocked: the process that produced Smith's belief in the disjunction (deduction from a justified false lemma) is not a reliable belief-forming process when applied to disjunctions one of whose disjuncts is randomly guessed. The Fake Barn case is also blocked: vision-in-Fake-Barn-County is an unreliable process.

Reliabilism faces two large objections. The generality problem: any belief is the product of indefinitely many candidate processes at different levels of grain (vision; vision-in-good-light; vision-in-good-light-of-a-barn-shaped-object). Which one is the relevant process for reliability? Without an answer the reliability condition is underdetermined. The new evil demon problem: an agent whose cognitive processes are intrinsically well-functioning but whose environment has been hijacked by a Cartesian demon so that the processes are no longer reliable would, on reliabilism, lack justification despite being subjectively indistinguishable from an unfooled agent. Internalists treat this as a reductio of reliabilism.

Reply 4: contextualism (DeRose 1992; Lewis 1996)

David Lewis and Keith DeRose proposed that the standards for knowledge attributions are context-sensitive: the truth value of "$S$ knows that $p$" depends on features of the conversational context, particularly which error possibilities are currently salient.⁵⁶

On Lewis's Rule of Attention, an error possibility is relevant to a knowledge attribution iff it is not properly ignored in context. In an ordinary low-stakes context, "I know I have hands" is true because skeptical possibilities are properly ignored. In a high-stakes or skeptic-engaging context, the same sentence may be false because the relevant error possibilities widen.

Contextualism explains Gettier cases by the same mechanism: in the Gettier case the actual route to truth makes a particular error possibility (Jones doesn't own a Ford) salient; in contexts where that possibility is properly attended to, the knowledge attribution fails.

The most serious objection is semantic: it is not obvious that ordinary "knows" attributions display the kind of context-sensitivity Lewis posits. Critics have argued the data better fits a sensitive-invariantism (the standards are pragmatically variable but the semantic content is invariant) or a subject-sensitive invariantism (the standards depend on features of the subject's situation, not on the attributor's context).

Reply 5: safety and sensitivity

Two modal conditions have become standard tools in the post-Gettier literature.

Sensitivity (Nozick 1981): $S$'s belief in $p$ is sensitive iff, in the nearest possible world where $p$ is false, $S$ does not believe $p$.⁷ The Gettier case fails sensitivity in the relevant respect; the Fake Barn case fails sensitivity (had it been a fake, Henry would still have believed it was a barn).

Safety (Sosa 1999; Williamson 2000): $S$'s belief in $p$ is safe iff, in nearby possible worlds where $S$ forms the belief on the same basis, the belief is true.⁸ Safety is the modern preferred condition; sensitivity has known problems with inductive knowledge and known-but-not-sensitive beliefs.

Neither condition is offered as a complete analysis; both are necessary conditions on knowledge intended to capture the no-luck intuition.

Worked Example: The Fake Barn Case

Setup: Henry drives through Fake Barn County, where ninety-nine of one hundred structures are barn facades visually indistinguishable from real barns from the road. He looks out the window and sees the one real barn. He forms the belief "that is a barn."

Apply the analyses.

• JTB. Henry's belief is true (it is a barn). His belief is justified (he is using normal vision under normal conditions, with no prior reason to doubt). He believes it. So JTB gives: Henry knows. But this verdict is the wrong one; most epistemologists agree Henry does not know.
• No-false-lemmas. Henry's reasoning does not pass through any false belief. The clause does not block the case. Verdict: Henry knows. Still wrong.
• Defeasibility. Is there a true proposition $d$ such that, if Henry believed it, his justification would be defeated? Yes: "almost all the barn-like structures around here are facades." Adding this would defeat. Verdict: Henry does not know. Correct.
• Reliabilism. Is the belief-forming process (visual recognition of barns in Fake Barn County) reliable? No: in this environment ninety-nine of one hundred visual judgments "that is a barn" would be false. Verdict: Henry does not know. Correct.
• Sensitivity. Had it not been a barn, would Henry still have believed it was one? Yes (the facades are visually indistinguishable). Verdict: belief is not sensitive; Henry does not know. Correct.
• Safety. In nearby worlds where Henry forms the same belief on the same basis, the belief is mostly false (he could easily have looked at any of the facades). Verdict: belief is not safe; Henry does not know. Correct.

The Fake Barn case discriminates: it blocks the simplest patch (no-false-lemmas) and supports the modal-condition family (defeasibility, reliabilism, sensitivity, safety).

Common Misconceptions

• "Justification means having a deductive proof." No. Justification covers any standing that supports the belief: perceptual experience, testimony, inductive evidence, memory. Deductive proof is the strongest form, not the only form.
• "Reliabilism solves all Gettier cases." It solves the original Gettier cases and the Fake Barn case but faces the generality problem (which process counts as the relevant one?) and the new evil demon objection.
• "Contextualism is just relativism." No. Contextualism is a semantic thesis about how the word "knows" works; it is consistent with there being a fact of the matter about whether the agent's epistemic state meets the contextually relevant standard.
• "After Gettier, everyone gave up on the JTB analysis." No. The JTB conditions are still taken as necessary by most epistemologists; the post-Gettier program is the search for a fourth condition that makes the conjunction sufficient. Some externalists (Williamson 2000) argue knowledge should be taken as primitive and the JTB analysis abandoned; this is a minority but influential position.

Comparisons to Related Views

Go Further

• Stanford Encyclopedia of Philosophy, "The Analysis of Knowledge" by Jonathan Jenkins Ichikawa and Matthias Steup. Canonical survey of the JTB analysis and the post-Gettier reply space.
• Stanford Encyclopedia of Philosophy, "Reliabilist Epistemology" by Alvin Goldman and Bob Beddor. Author-of-record survey.
• Stanford Encyclopedia of Philosophy, "Epistemic Contextualism" by Patrick Rysiew.
• Williamson, Timothy. Knowledge and Its Limits. Oxford University Press, 2000. The reference for the "knowledge first" program; argues knowledge resists analysis.
• Sosa, Ernest. A Virtue Epistemology. Oxford University Press, 2007. Influential virtue-theoretic alternative to the JTB program.
• Bonjour, Laurence, and Ernest Sosa. Epistemic Justification: Internalism vs. Externalism, Foundations vs. Virtues. Blackwell, 2003. Excellent dialectical primer on the internalism / externalism axis.

For the formal-credence version of belief and justification, see Bayesian Epistemology. For the original Gettier construction, see The Gettier Problem. For inductive justification specifically, see Induction and Hume's Problem.