Who is responsible for a belief that reached you through a machine?

When someone tells you something, responsibility for your belief is shared. She is answerable for what she said; you are answerable for whether it was reasonable to take her word. That sharing is part of what lets testimony carry knowledge: you can defer, because the warrant runs through her.

Where the load falls with a machine

When the source is a language model, the warrant does not run through it. As I argue in Is an AI chatbot an informant or an instrument?, your standing is that of a reader of an instrument: whatever warrant you have is generated by your own use of it, by its record on questions like yours and your competence in reading it. The load of checking is therefore yours in a way it is not when a competent person tells you something.

This describes where the epistemic work has to be done, if it is to be done at all, and carries no moral accusation against users.

What the makers answer for

Goldberg contrasts relying on an epistemic subject, who can be assessed for rationality and responsibility, with relying on a mechanism, assessable for reliability alone, where the normative burden falls on the user. The attributions of responsibility to designers, deployers and operators, distributed across many hands in Nissenbaum’s sense, should be conceded: they are how we already treat certified instruments, whose makers we blame and whose deployment we regulate.

Those attributions do not, however, turn the system’s outputs into anyone’s testimony. A spokesperson relays a message her principal has authorised and would stand behind, whereas a deployer has not seen the output, does not believe it and disclaims it. Responsibility for putting a system into the world is not authorship of its sentences.

The open part

Our practices of blame and trust were built for the first kind of chain, the kind made of people. Chains that run partly through machines redistribute responsibility in ways those practices do not yet track, and that is why I count the question among the open ones. What is open is not whether responsibility vanishes but how to draw its new distribution accurately enough to act on.