Given the unprecedent advancements in the field of simulation of human linguistic behavior by generative AI models, one might expect – notwithstanding the obvious limitations of such models – that this development will be theoretically advantageous to the deflationist and the reductive theories of meaning. To test this hypothesis, we first explicate what we take to be the Use Theory of Meaning. Roughly, it is neither the language users’ representational states (or the neural correlates thereof) nor the reference relation to extramental reality (“fixation of reference”), but the regular use of linguistic expressions in various contexts that accounts for the phenomenon of meaning. We first narrow down this rather vague claim that has its roots in in Wittgenstein’s Philosophical Investigations. We do that (1) by explicating two crucial notions from Sellars’ early conception of language games, and (2) by showing how these notions are complemented and further developed by Horwich in his non-normative, deflationist and reductive account of meaning. We then examine the potential of generative AI models to generate, in a conversational manner, longer portions of text which not only conform to the syntactic rules of natural language, but which also seem to satisfy, and to a surprising degree, its semantic and pragmatic constraints. We consider the similarities and differences between an artificial and a human language user, with special regard to the issues of rule following and normativity.
Janović et al. (Thu,) studied this question.