“Like” Literature. A Grammar of Commercial Generative Large Language Model Products

Abstract

Given that recently-developed commercial large language model products have been called the ‘heirs to structuralism,’ this article reports on how two such products, whether specifically designed to ‘tell stories’ (Squibler) or not (ChatGPT) generate text from ‘kernel simple stories,’ ‘minimal stories,’ ‘simple stories,’ and non-narrative fragments from Gerald Prince’s 1973 A Grammar of Stories. Prince’s minimal story requires an original state, an event, and a secondary state, appearing in this temporal order, the secondary state a direct result of the event, the original and secondary states different, yet still maintaining some relation. To this might be added Monika Fludernik’s (1996) requirement of embodied experientiality or an experiencing position. The two products sometimes do – and sometimes don’t – fulfill these requirements when asked ‘to tell a story’ about Prince’s minimal stories or non-stories. Predominant in the products’ responses were insistent elaborations of diegetic spaces, multiple positions of embodied experientiality, and alignments of these positions through emotionalized encounters. The products often elaborated causality where none was suggested. If, as Angus Fletcher suggests, computers are “fundamentally incapable of grasping cause-and-effect” (2021, 3), only able to suggest causality by working backward from correlation (2022, 132), their designers seem aware that narrative requires it, including it even when unasked.