How words mean: lessons from large language models

Time period:

1 December 2025 – 31 December 2029

Project leader:

Jessica Pepp

Project member:

Pär Sundström, Dimitri Coelho Mollo, Bram Vaassen

Funder:

Swedish Research Council

Type of award:

Project grant

Total funding:

4,536,000 SEK

Given the right prompt, ChatGPT will output, “Ibuprofen can increase the risk of gastric ulcers”. The same output may be produced by a human doctor. It is natural to assume that these two outputs have the same meaning. But do they?Earlier research has by and large argued in one of two ways about this question. One argument has been that outputs of ChatGPT and similar large language models are meaningless since ChatGPT lacks the mental states that according to established theories are required to ground meanings. Another argument has been that the doctor’s and ChatGPT’s utterances have the same meaning on the ground that established theories can, after all, deliver that result.This project explores two paths not taken in earlier research. It employs an “adversarial methodology” in that the two paths run counter to one another. It has two aims.The first aim is to test the hypothesis that ChatGPT’s output is meaningful but has a meaning that differs from the doctor’s utterance.The second aim is to work from the assumption that the two outputs have the same meaning and develop a novel account of what makes this the case. Specifically, we will explore the hypothesis that ChatGPT expresses the same meanings that we express, though not because of any underlying mental states, but because ChatGPT has “quasi-mental states” that ground its meanings.