The Turk Is Mechanical, and Yet He Plays: Modeling AIs as Testifiers in Social Epistemology

Abstract

Advances in artificial intelligence (AI) have brought AI systems new salience in our epistemic lives: in many situations, AI systems can act much like human experts, advisors, and epistemic peers. Because AIs function like human beings in some epistemic contexts, there are AI analogs to traditional social epistemological questions in many domains. For instance, just as we face the problem of deciding which human experts to trust among many contenders, we must decide which AI pseudoexperts to rely on.

This situation poses a conundrum for social epistemology because most of the socioepistemological concepts and vocabulary we would use to articulate these questions are anthropocentric. But AI systems do not necessarily fit anthropocentric definitions. Testimony, for example, is usually defined as requiring communicative intentions on the part of the testifier, but present-day AIs probably lack such mental states. How, then, should we proceed with the social epistemology of AI when it implicates anthropocentric concepts?

In this thesis, I examine two possible stances on this question and their consequences for social epistemology. I first examine what I call the Revisionary Stance, the latent stance on this question in existing socio-epistemological work (Freiman 2024; Hauswald 2025b; Shin 2026) that revises existing concepts to suit AI or invents new, AI-specific ones. I argue that the Revisionary Stance does not do much to help us solve important questions in AI social epistemology, and propose the Modeling as a Testifier (MAT) stance as a superior alternative. MAT involves modeling, or treating AIs as if they were human testifiers, to answer important socio-epistemological questions. This solves the problem of anthropocentrism by confining anthropocentric terminology, which is inapplicable to AI, to model-talk. After providing an account of the epistemology of MAT via Mary Hesse's account of analogies and analogical modeling in science (1966), I demonstrate MAT's practicality by showing how we can understand some recent work by Amber Ross on appropriate deference to opaque AI systems as an instance of employing MAT to reason about AI (2024). Finally, I respond to some objections and provide suggestions for the further development of MAT.