Social Epistemology
Knowledge we hold together, and what a machine link does to a chain of it.
Almost everything any of us knows, we know because someone told us. Social epistemology studies the practices and institutions through which that happens: testimony, trust, expertise, and the chains along which knowledge travels from one person to another. My specialisation is the social epistemology of artificial intelligence — what changes when one of the links in those chains is a model.
Testimonial chains work because their links preserve warrant: A tells B, B tells C, and C can end with warrant that traces back to A. In “Assertion Without a Speaker” I argue that a language model is not a link of that kind. Generation does not preserve particular assertions together with their grounds; it synthesises new text weighted by likelihood. Chains of model-generated text reposted by users do not preserve warrant as chains of human testimony do. They preserve text.
None of this makes reliance on these systems irrational. A user can come to a warranted true belief through a model, but her standing is that of a competent reader of an instrument, not of a hearer deferring to an informant: the epistemic load lies with her, and the appropriate attitude is calibrated reliance. The difficulty is that the instrument’s calibration is not transparent — and, on the best current evidence, in respect of the particular content it emits it is not calibrated to anything.