Jessica Pepp on “Can reference be purely linguistic? Lessons from LLMs”

The question of reference determination asks what makes it the case that uses of certain expressions (paradigmatically proper names) refer to the particular things they do. This long-standing question in the philosophy of language has taken on additional interest with the arrival of large language models (LLMs). LLM-driven chatbots appear to use names and other referring expressions, and their uses appear to refer. If this is in fact the case, then our theories of reference determination should accommodate them; if it is not, then we owe an account of why not.

Views about this range along a spectrum. At one end, LLM outputs are held to lack meaning and reference altogether, with any semantic significance being merely projected onto them by human interpreters (Bender & Koller 2020; Ostertag 2023; Mallory 2023; Titus 2024; Hattiangadi & Schoubye ms). At the other end, LLMs are treated as full-blown linguistic and cognitive agents on a par with humans (Cappelen and Dever ms). Both endpoints suggest that LLMs need not place any new constraints on theories of reference. On intermediate views, it is more likely that they would, since they might really refer without having various cognitive capacities often invoked to explain reference.

This talk examines one such middle-ground possibility: that LLMs really have, and do, language, without having any further cognitive apparatus (cp. Nefdt ms, Grindrod 2024). Abstracting from the details of actual systems, I consider what I call “pure language-rs”: linguistic beings lacking any extra-linguistic form of cognition or perception.

The central question is whether a pure language-r could refer. I focus on reference borrowing: the idea, found in the causal-historical pictures of Kripke and Donnellan, that a speaker can inherit reference from pre-existing linguistic practices. I distinguish two roles such practices are taken to play: supplying a rule to which the borrower’s use is subject (exemplified by Evans 1984), and supplying a channel through which the referent reaches the borrower, in something like an extended perceptual sense (exemplified by Devitt 1981, 2015). I argue that neither applies straightforwardly to pure language-rs. I then consider whether reference borrowing might be construed in purely linguistic terms, drawing on Bianchi’s (2015) thought-free, repetition-based account, and argue that here, too, there are challenges.

I do not argue that LLMs are, or are not, pure language-rs. But if they are, and if they refer, they may impose new constraints on theories of reference. Moreover, the mere possibility of such beings is itself worth taking seriously in theorizing about reference.

Jessica Pepp (Uppsala University)

To join the session on Teams, please get in touch with Giulia Terzian (giuliaterzian@fcsh.unl.pt).

This event is part of the ArgLab Research Colloquium organised by Maria Grazia Rossi, Giulia Terzian and Alberto Oya at the Laboratory of Argumentation, Cognition and Language of the NOVA Institute of Philosophy. For any inquiries, please contact Maria Grazia, Giulia or Alberto.