Semantics circle: Theories of AI

Abstract

The rapid development of generative AI has given scientists, including linguists, a lot of work to do. We find ourselves dealing with a new technology, which apart from its machine learning foundations, we have little understanding of. There is an awful lot of work to do in trying to decide which capabilities large language models do and don't have and how such capabilities can be utilised and improved upon in a responsible and sustainable way. As a result, most of the linguistic attention to AI has been pretty one-sided. The focus is on grasping and modifying the linguistic behaviour of AI systems. There has been remarkably little interest so far in trying to understand the linguistic behaviour of humans in the context of human-AI communication. We know that, generally, humans tend to transfer social behaviour to non-humans in situations of human-to-non-human interaction. Little is known, however, about the extent to which such transfer also applies to fine-grained aspects of linguistic pragmatics. Understanding this is important since we know that the pragmatic choices made by human interlocutors are - at least to some extent - based on theorising about conversational partners. In the case of human-AI interaction, not much is in place to ground such theorising. How do humans form theories of AI and how does that influence their pragmatic behaviour? In this talk, I will discuss the beginnings of a project targeting these questions.