The Human and the Mechanical: logos, Veridicality Judgment, and GPT Models

The paper addresses the question of whether GPT models participate of human-like linguistic ability with a specific focus on semantic and pragmatic capacities. We frame the question in terms of having the ability to judge truth and falsity in reference to the world, and consequently understanding of meaning, and offer a new argument based on the formation of veridicality judgment which is the foundation for assertion and for recognizing meaning. The veridicality judgment is always about the world, and presupposes the ability to recognize intentions; the organism that can form it needs to be able to access the world and understand communicative intent. Veridicality judgments are formed on the basis of two components: (i) evidence that relates to reality (exogenous evidence) and (ii) endogenous preferences and private beliefs. GPT models do not relate to reality and cannot have endogenous preferences. Therefore, GPT models cannot form a veridicality judgment, they can only mimic it.