Abstract Can Large Language Models (LLMs) perform speech acts? I show how a norm-based argument against LLM-assertion can be generalized so as to support a negative answer to this question. I then investigate how the resulting no-speech-act-view interacts with several other debates about how to conceptualize the linguistic outputs of LLMs. I argue that it conflicts with the popular conception of LLMs as bullshitters and that it helps decide between two competing externalist accounts of how artificially generated expressions refer. I close by discussing a challenge to the no-speech-act-view and by suggesting a novel strategy for addressing it. A general lesson from this discussion is that our understanding of LLMs' linguistic outputs will profit from complementing extant meta-semantic theorizing with more detailed meta-pragmatic theorizing about discourse between humans and LLMs.
Lukas Skiba (Wed,) studied this question.