No humans-in-the-loop: The people-less stories generated by GPT

Abstract In the last two years, numerous news outlets and academic articles have claimed that GPT-produced text is indistinguishable from human-written text across several genres. In this essay, we argue that not only does GPT not write fictional narratives like a human, it also often fails to write about humans. We demonstrate how the distribution of stylistic features that we can use as proxies for key elements of narrative, including the frequency of named entities, dialogue, pronouns and the second person, can reliably distinguish between human and GPT-generated text. We argue that these textual markers point to a deeper conceptual difference between human-written and GPT-generated text — GPT texts contain few characters, rarely feature a characterized narrator, and describe inanimate objects more often than people or characters. In sum, the thousands of GPT-generated narratives we analyzed struggled to meet the commonly accepted definition of narrative in contemporary narrative theory, which emphasizes the centrality of the human experience in narrative representation, and encompasses the broader rhetorical communication circuit in which narrative takes place.