Abstract

In Plato’s Gorgias (Hamilton and Cairns, 1961), Socrates provides criticism of rhetoric as a merely persuasive art subordinate to argument. Among Socrates’ problems with rhetoric is the idea that it is essentially inferior to rational discussion, as with the relation of ignorance to knowledge. For instance, flattery plays a role in rhetoric in which agreement is sought through pandering to an ignorant audience: “…the rhetorician need not know the truth about things; he has only to discover some way of persuading the ignorant that he has more knowledge than those who know…” (Op Cit.). This is unlike truth-driven, knowledgeable discussion where agreement comes through genuine understanding, and through the challenging of reasoning to ensure it stands up to scrutiny. Socrates supposes rhetoric is related to knowledgeable argument in the way that cookery is to medicine: “Cookery simulates the disguise of medicine, and pretends to know what food is the best for the body” or cosmetic adorning of the body that permits “…men [to] affect a spurious beauty to the neglect of the true beauty which is given by gymnastic.” (ibid.) Rhetoric is essentially a parody of reason, and a pale imitation. GPT4 and related language models produce their sometimes impressive results in a way akin to the model of rhetoric of which Socrates disapproved. Not every use of generative AI, of course, is intended to be argument. Outputs from these systems might be poetry, or software code, etc. It isn’t the case, therefore, that the Socratic critique of rhetoric is directly applicable to all instances of generative AI output. Poetry generated from such an AI system as this needn’t abide by Socratic standards for rational argument. Nevertheless, the critique of rhetoric highlights a missing dimension that generative AI outputs appear to exhibit – generation from a reasoned basis. The outputs from generative AI, whether text, poetry, code, or whatever, are produced not from a specific basis responding to a given need, but via a generic means of mimicry. Passages of text, for example, are produced from vast arrays of prior data such that they are inherently derivative in terms of their contents and bound to mimicry of their source material in form. The now familiarly ‘convincing’ outputs flatter to deceive as they contain enough crumbs of recognisability that they can be understood as representing some aspect of the genre they pretend to exemplify. Two points will therefore be pursued following from the Socratic critique of rhetoric. One is on the essentialist theme, and another reflects not upon the nature of language model outputs, but the justifications for their uses. In each case, language models will be seen to be without content yet generative of duties produced for ‘us’ as users or consumers of their outputs. Overall, the Socratic critique of rhetoric gives a useful framework for examining language models like GPT4. On the essentialist point, GPT4 etc. can be shown to have no reasons for anything they do insofar as they re-present existing data via sets of transformations (cf. Bishop, 2021). It is thus essentially different to rational discourse, and worse off for it, in meaninglessly echoing and amplifying past forms. There is no possibility of distinguishing the significance of one generated form over another inherent to the generative AI system, marking it out as distinct from human reasoning. On the duties side, the Socratic critique of rhetoric gives us good reason to resist a sort of pandering inherent in language model outputs. It says as much about the banality of mass media that language models can produce convincing, say, news reports or film scripts, as it does about the ingenuity of the language model. ‘We,’ as users or consumers of language models and their outputs, ought to be very careful about the role played by pandering.