Trust

Trust is reliance on someone that accepts vulnerability to their failure and that, unlike mere reliance, can be betrayed. I rely on a bridge; I trust a friend. If the bridge gives way I am harmed, but I am not let down by it. Baier located trust in accepted vulnerability to another’s goodwill; Jones in an attitude of optimism about another’s goodwill and competence; Hawley in the belief that the trusted party has a commitment to do what we rely on her to do.

Faulkner’s distinction between two kinds of trust is the one I use most. Predictive trust is the expectation that something will behave as it has, and one can place it in any regular process. Affective trust presses a normative claim on its object, and it is the trust a speaker invites when she tells you something and so becomes answerable for its truth. On Faulkner’s assurance view, testimonial warrant is grounded in the second.

Artefacts are proper objects of predictive trust. On functional accounts such as Simion and Kelp’s, they can also be trustworthy, in the sense of being disposed to fulfil the obligations their functions generate. These are good accounts of warranted reliance on artefacts, and they vindicate the stance I recommend towards language models, which is the calibrated reliance appropriate to an instrument. The point is not that machines are beneath trust. It is that neither predictive trust nor functional trustworthiness is the relation that grounds testimonial warrant, and substituting them changes what is being explained rather than satisfying it.

There is a further question about whom one trusts when one trusts a system. On reductive accounts, trust in an artefact is trust in the people who build and maintain it. But the makers do not assert the outputs: they have not seen them, do not believe them, and stand ready to disclaim them. The vocabulary of trust, applied to the system, can hide where answerability actually lies.