For over two millennia, philosophers operated under what seemed like an airtight definition: knowledge is justified true belief. Then in 1963, Edmund Gettier published a three-page paper that shattered the consensus. His counterexamples showed that a belief could satisfy all three conditions and still fail to constitute knowledge—the justification was, in some sense, lucky.

The philosophical response was a half-century of patches: defeasibility conditions, causal theories, reliabilism, virtue epistemology. Each captured part of the intuition but none achieved consensus. The problem, I argue, is that natural language proved too coarse a tool for the surgery required.

Formal epistemology offers a different approach. By translating Gettier's intuitions into the precise vocabularies of probability theory, modal logic, and possible worlds semantics, we can dissect exactly what goes wrong in these cases. We can distinguish probabilistic justification from epistemic luck, separate truth-tracking from mere reliability, and prove formal theorems about when belief and knowledge come apart. What follows is not yet another patch, but a demonstration of how mathematical rigor transforms a philosophical puzzle into a tractable analytical problem.

Consider the classical Gettier structure formally. Let S believe proposition p on the basis of evidence e, where P(p|e) is sufficiently high—say, above some threshold τ. In a Gettier case, p is true, but the chain of reasoning that licenses P(p|e) > τ proceeds through a false intermediate proposition q.

Formally: S believes p because S believes q, and q entails p. But q is false, and p happens to be true for reasons orthogonal to q. The conditional probability P(p|e) is high, but the actual world realizes p through a probabilistically independent path.

Consider Smith's belief that the man who gets the job has ten coins in his pocket. Smith's evidence supports this via the proposition that Jones gets the job. Bayesian updating on Smith's evidence yields high credence in the conjunction, but the truth-makers diverge: the actual world satisfies the existential claim through Smith himself, an event probabilistically independent of Smith's evidence.

We can formalize the failure precisely: P(p|e) is high, but P(p|e, w)—the probability of p given the evidence and the actual world's truth-makers—is decomposable into a justified component and a lucky component. Knowledge requires that the evidential and metaphysical paths to truth coincide.

This probabilistic decomposition reveals why justification thresholds alone cannot solve Gettier. No matter how high we set τ, we can construct cases where evidence supports p for reasons disconnected from why p is actually true. The problem is structural, not quantitative.

Takeaway

Justification and truth can be probabilistically high yet causally divorced. Knowledge requires that the path your evidence supports is the same path the world actually took.

Robert Nozick's tracking theory replaces justification with two modal conditions. Formalized in possible worlds semantics: S knows p iff (i) p is true, (ii) S believes p, (iii) if p were false, S would not believe p, and (iv) if p were true, S would believe p.

Using counterfactual semantics à la Lewis-Stalnaker, condition (iii) becomes: in the closest possible worlds where ¬p, S does not believe p. Condition (iv) requires: in the closest possible worlds where p, S believes p. These are sensitivity and adherence respectively.

Apply this to a Gettier case. Smith believes the man with ten coins gets the job. In nearby worlds where this is false—where neither Smith nor Jones has ten coins, or where someone else with different pocket contents is hired—Smith would still believe it, because his belief tracks Jones, not the disjunctive truth-condition. Sensitivity fails.

The formalization reveals a deep insight: tracking is not a property of the belief itself but of the function mapping worlds to belief states. We require that this function be appropriately correlated with the truth-value function of p across the modal neighborhood of the actual world.

Critically, the choice of similarity metric on possible worlds becomes philosophically substantive. Different metrics yield different verdicts on identical cases. Formal tracking theory thus exposes a hidden assumption in our pre-theoretic judgments about knowledge: which counterfactual variations we hold fixed.

Takeaway

Knowledge is a modal property, not just an actual one. What matters is not only what you believe here, but what you would believe across nearby possibilities.

The sensitivity condition faces well-known problems with skeptical scenarios and inductive knowledge. Ernest Sosa proposed safety as an alternative: S safely believes p iff in all close possible worlds where S believes p, p is true.

Formally, where B(p) denotes S's believing p: sensitivity is ¬p □→ ¬B(p), while safety is B(p) □→ p. These look superficially similar but are modally distinct. Sensitivity quantifies over close ¬p-worlds; safety quantifies over close B(p)-worlds.

The asymmetry matters. Consider a belief formed via a reliable method that happens to land on a necessary truth. Sensitivity is vacuously satisfied—there are no close ¬p-worlds—but this seems too easy. Safety, by contrast, requires that across the methods nearby world-variations would generate, the belief remains true. It captures method-level reliability.

Yet safety has its own pathologies. Lottery cases famously confound it: your belief that you'll lose is safe across close worlds, yet seems not to constitute knowledge. The formal apparatus exposes that the problem lies in our metric of world-closeness—statistically unlikely outcomes may still occupy modally close worlds.

The contemporary formal literature has developed graded versions: p-safe to degree n iff the proportion of close B(p)-worlds where p holds exceeds n. This shifts the analysis from binary modal facts to measure-theoretic structures over possibility spaces, integrating modal and probabilistic frameworks into a unified formal epistemology.

Takeaway

Sensitivity asks what you would believe if things were different; safety asks whether your belief could easily have been wrong. Both matter, but they answer different questions.

The Gettier problem resisted resolution for fifty years not because epistemologists lacked insight, but because natural language could not articulate the distinctions required. Probability theory separates evidential support from truth-making structure. Possible worlds semantics distinguishes tracking from mere correlation. Modal logic differentiates sensitivity from safety.

None of these formal tools singly solves Gettier. What they accomplish is something more valuable: they make the philosophical disagreements precise. When two epistemologists disagree about a case, formal analysis can locate the disagreement in a specific parameter—a similarity metric, a probability threshold, a modal scope.

This is the methodological promise of formal epistemology. By translating intuitions into mathematics, we transform interminable debate into structured inquiry. The Gettier problem becomes not a puzzle to solve but a phenomenon to model, with the model itself revealing which features of knowledge are doing the philosophical work.