Undead Signs: On the Possibility of a Computational Illusion of Meaning

Abstract The position that the output by computational models of language, especially Large Language Models (LLMs), has meaning is usually criticized on the basis of the argument that the models do not have access to the objects to which the produced linguistic signs refer to. In this paper, I will show that a more principled argument is needed in order to distinguish between the actual language use by human beings and the simulation of it in LLMs. I will defend the position that we should view the latter’s outputs as strings of undead signs, i.e. signs that are not themselves already meaningful, but are generated in a way that makes it seem as if they were. From this results the novel phenomenon of a computational illusion of meaning, which is characterized by the large-scale automated production of undead signs.