Author: Samuel Wheeler III

Many writers discussing Artificial Intelligence argue that what a Large Language Model produces are not sentences with truth-values but rather “stochastic parrotings” that can be interpreted as true or false, but in the way that Daniel Webster interpreted the Old Man in the Mountain as a sculpture by God with a message for humanity. Steffen Koch has argued that names used by LLMs refer in virtue of Kripkean communication-chains, connecting their answers to the intended referents of names by people who made the posts in the training data. I argue that although an LLM’s uses of names are not connected to human communication chains, its outputs can nonetheless have meaning and truth-value by virtue of design-intentions of the programmers. In Millikan’s terms, an LLM has a proper function intended by its designers. It is designed to yield true sentences relevant to particular queries.

Speech and writing actions differ from their products. We can call the actions utterings and inscribings. An uttering’s product is an utterance. An inscribing’s product is an inscription. An uttering or inscribing is “abandoned” because features of the uttering or inscribing are not linguistically available to the audience in the utterance or inscription. Iterations of what is linguistically available, whether by the speaker or others, differ in truth-conditions from the original simply because the present and past tenses are indexicals. The force of an original is not linguistically available, so every utterance or inscription is “abandoned.” Repetitions of utterances with indexicals generate expressions with different truth-conditions but may be copies, and so iterations.