As a Large Language Model: Ontological-Category Cue Effects on Agent and Message Evaluations

Abstract

The tendency for AI to refer to itself—for instance using first-person pronouns and referring to (in)capabilities—raises questions about the interpretation and effects of machines’ self-referential language in human-machine communication. AI vary in their tendencies to identify themselves as machines or to mask that ontological category in the course of interactions. To examine how self-referential ontological-category cues (i.e., “As a large language model …”) influence judgments of contextualized agents and their responses, a 2×2×2 experiment was conducted. Participants (N = 800) evaluated an exchange between an inconspicuous user and ChatGPT, manipulated to represent three variables: Machine cue present/absent × natural/technical topic × creative/logical framing. Experimental findings point to a weak interaction effect of the cue and the topic suggesting a mild “stay in your lane” effect. Findings have implications for whether and in what context machines may be more or less favorably evaluated when their machine status is cued.