Abstract:

In this presentation, we address an epistemological question about deep-learning AI systems, namely can anyone justifiably place any sort of epistemic trust in deep-learning AI? By epistemic trust we generally mean the confidence or reliance that someone grants to various sources of knowledge. In the context of deep-learning AI, epistemic trust can be said to involve the belief that the systems we interact with are reliable, fair, trustworthy as sources of knowledge.

Starting from this working definition, we adventure to answer “yes, but defeasibly” to the question formulated at the beginning, and propose to show a way. First, we examine a set of defeasible conditions for epistemic trust in deep-learning AI, in order to make sure that an epistemically normative notion of trust is in place (i.e., a trust that can be justifiably attributed/revoked in different circumstances). We then distinguish between two concepts of epistemic trust, already discussed in the philosophical literature: basic and enhanced epistemic trust, which differ in their defeasible conditions – in other words, one concept of minimal, while the other is maximal. Finally, relying on this analytic framework with defeasible conditions, we argue that deep-learning AI forms a non-homogenous class with respect to how various systems navigate epistemic trust. To illustrate this, we discuss two mini case studies: AlphaFold 2 and ChatGPT (regardless of the model powering it). We conclude that AlphaFold 2 can be endowed with a form of basic epistemic trust, while ChatGPT (regardless of the model powering) is still in a “grey area”, which requires further scrutiny from users.

NEC seminars as well as meetings of the research groups can be attended by the fellows and the alumni of the programs, and their guests.
They are not open to the general public, and attendance is only by invitation, or by permission as a result of a request addressed to NEC by e-mail [nec @ nec.ro].