The Elusive Nature of Knowing: Revisiting Reliabilism and Epistemic Luck — Epoche B2
Reliabilism's Real Problem Is Not Luck The claim I want to correct The usual complaint about reliabilism is that it underestimates luck. Reliable processes, the complaint runs, can still deliver true beliefs by accident, so reliability cannot be what turns true belief into knowledge. The fake-barn case is produced as the proof, and the essay ends. The complaint is not wrong so much as shallow. It suggests that reliabilism gives a determinate but mistaken verdict about such cases, which a further condition could correct. What actually happens is more troubling for the theory. Reliabilism as standardly stated does not give a determinate verdict about the fake-barn case at all. Depending on how the belief-forming process is described, the theory says the belief is justified or says it is not, and nothing in the theory settles which description is the right one. The luck is a symptom. The disease is that a central term of the theory has never been fixed. Where the problem starts Edmund Gettier's 1963 paper showed that justified true belief is not sufficient for knowledge [1] . In his first case, Smith has excellent evidence for a conjunction: that Jones will get the job, since the company president has told him so, and that Jones has ten coins in his pocket, since Smith has just counted them. From this he validly infers that the man who will get the job has ten coins in his pocket. In fact Smith himself gets the job, and — unknown to him — he has ten coins in his own pocket. His belief is true, justified, and not knowledge. The diagnosis that has stuck is luck. Smith's belief is true, but its truth has nothing to do with what made it justified. Knowledge requires that the connection between belief and truth not be accidental. Alvin Goldman's reliabilism [2] , set out in 'What Is Justified Belief?' (1979), locates the connection in the process. A belief is justified if it is produced by a process that tends to yield truths. This is externalist: what matters is the process's actual reliability, not the believer's opinion of it. Two kinds of luck Duncan Pritchard draws a distinction that clarifies what reliabilism can and cannot do [3] . It grows out of the taxonomy of veritic luck in Epistemic Luck (2005 [4] ), and the labels under which it now travels are those of 'Anti-Luck Virtue Epistemology' (2012). In some cases luck intervenes between the believer and the fact: something goes wrong in the chain, and something else independently makes the belief true. Gettier's own cases are of this kind, and a process-based theory handles them well, because the process that produced the belief is not one that tends to produce truths about who will get the job. In other cases the luck is environmental . Nothing goes wrong in the process; the believer is simply in a place where the process could easily have misfired. Goldman's example in 'Discrimination and Perceptual Knowledge' (1976 [5] ), which he credits to Carl Ginet, is the standard illustration. Henry looks at a barn and forms the belief that it is a barn. His vision is good; the light is good; the object really is a barn. But he is in a district where nearly every other barn-shaped structure is a façade, and he cannot tell the difference from the road. He does not know. The reply that looks like a rescue The reliabilist has an obvious move. Describe the process narrowly enough and it is not reliable after all: forming beliefs about barns by roadside vision in a district full of façades yields truths perhaps one time in twenty. So Henry's belief is unjustified, and the theory delivers the intuitive verdict without any new condition. This move works too well. Apply the same narrowing to ordinary cases and everything falls over. Consider your belief that there is a cup on the table. The process might be described as vision; or as vision in good light; or as vision in good light by a person of your age with your prescription at this angle at this hour. It might equally be described as vision by a person who happens to be in a country where three per cent of tables display trompe-l'œil cups, if such a country exists. Broad descriptions make almost every perceptual process reliable. Narrow enough descriptions make each process a one-off, reliable if the belief is true and unreliable if it is false — which makes the theory circular. And the reliabilist has no rule telling us which level to use. The generality problem This is the generality problem, pressed in its now-standard form by Earl Conee and Richard Feldman in 'The Generality Problem for Reliabilism' (1998 [6] ). Every particular act of belief-formation is a token of indefinitely many types. Reliability is a property of types, not of tokens: a single event has no frequency of success. So the theory's central claim — that a belief is justified when produced by a reliable process — has no content until a rule is given for selecting the relevant type. Conee and Feldman argue that no adequate rule has been produced, and their argument is not that the candidate rules are unattractive. It is that the constraints pull against each other. The selected type must be narrow enough to be explanatory, broad enough to have a frequency, uniform enough to apply across cases, and it must deliver the verdicts the theory was designed to deliver. Proposals that satisfy any three tend to fail the fourth. Now return to the fake barns. What the case shows is not that reliabilism gives the wrong answer. It is that the theory gives whichever answer the description of the process supplies, and the description is being chosen after the fact to fit the verdict we already have. That is not a theory being refuted by a counterexample; it is a theory that has not yet said enough to be tested. What safety adds, and what it does not The best-known repairs work at a different level. Robert Nozick [7] , in Philosophical Explanations (1981), proposed sensitivity : S knows that p only if, were p false, S would not believe it. Ernest S