Justification

Justification is what makes a belief reasonable to hold. Internalists locate it in reasons available to the believer on reflection; externalists locate it at least partly in facts she may have no access to, such as the reliability of the process that formed the belief. BonJour’s clairvoyant, reliably right but with no reason to think so, is the standard case against externalism; Goldman’s reliabilism is the standard externalist theory.

Testimony complicates both pictures. Most of what any of us believes we were told, and we can rarely reconstruct the speaker’s reasons. Burge’s answer is that we have an entitlement, a warrant that need not be accessible to us, to accept what is intelligibly presented as true, absent reasons for doubt. On this picture the hearer’s warrant is not generated by her own resources; it is transmitted from a source whose standing she relies on.

The distinction I need is between warrant that is transmitted and warrant that is generated. A barometer transmits nothing, because it asserts nothing. A user who reads it well still comes to know, and her warrant is generated by her use: by the instrument’s reliability in the conditions of use, her competence in reading it, and the absence of defeaters. Perception has the same structure, so there is no demotion in it. My claim about language models is that a user’s warrant for what a model tells her, when she has warrant at all, is of this second kind (see instrument reading). Nothing is transmitted from the system, and the epistemic load is on her.

The distinction presupposes non-reductionism about testimony, on which ordinary testimonial warrant is not itself inductive evidence about the speaker. A global reductionist would flatten the contrast at the hearer’s end. I do not argue against reductionism; on that view my claims about sources and chains relocate rather than lapse.