Reliabilism

Reliabilism is the view that a belief is justified, or counts as knowledge, when it is produced by a reliable process, one that tends to yield true beliefs. Goldman’s process reliabilism is the canonical version: perception, memory and good reasoning confer justification because they are reliable; wishful thinking and guessing do not. It is the leading externalist theory, since the believer need not know that her process is reliable for her belief to be justified.

Two objections dominate. BonJour’s clairvoyant and the new evil demon problem suggest that reliability is neither sufficient nor necessary for justification. And the generality problem, pressed by Conee and Feldman, asks which type a given process belongs to: every token process instantiates many types of widely varying reliability, and the theory needs a principled way to choose.

Reliabilism matters for the epistemology of AI because language models are often very reliable, and it can seem that this settles whether one can learn from them. Goldberg’s extended-process account of testimony, on which the process producing a hearer’s belief includes the speaker’s own production, is the natural home for that thought. I think the generality problem shows why it does not settle it. Type a broken thermometer’s process as reading-this-thermometer-in-this-thermostatted-room and it comes out reliable, so a reliabilism with no constraint on typing licenses the verdict that posting its readings transmits knowledge.

Goldberg does not accept that verdict, and the line he draws is the one I draw. He requires that the process whose reliability constitutes a hearer’s warrant be the operation of an epistemic subject, assessable for rationality and responsibility and not merely for reliability. That is a demand that the reliability be endogenous: a property of the source’s own operation rather than of a fortunate alignment between its outputs and the world (see the endogeneity requirement).

A reliabilist about processes and a tracking theorist about tokens can therefore be held to the same minimum. They differ over the grain at which the connection to truth must hold, not over whether it must. What reliabilism alone cannot do is turn frequency reliability into the warrant of testimony.