2025/08/05 by Gabi Kirilloff, C Jimmy Carroll, Zeina Daboul +4 · 1 voice
Computer Science · Medicine · #Text Readability and Simplification #Artificial Intelligence in Healthcare and Education #Topic Modeling
paper · pdf · doi:10.1162/99608f92.6d5fb5ef
We compare a corpus of 6,000 paragraph–length synthetic texts generated by GPT4 “in the style of” ten canonical, nineteenth-century, English language authors to authentic texts written by the same authors. We selected these authors based on existing humanities scholarship and metrics of cultural significance, including number of Goodreads reviews. The contemporary popularity of nineteenth-century writers has been largely determined by their place in the literary canon. Selection of nineteenth-century texts based on contemporary popularity allows us to investigate GPT’s biases in respect to literary texts that are both culturally valued and widely read. Key questions that motivated our research include: How do the stylistic markers of synthetic “literary” texts differ from those of their authentic counterparts? Which authors or demographics (such as male authors) is GPT best at mimicking? What can the varying effectiveness of GPT’s ability to replicate the style of certain authors tell us about how LLM training data has downstream effects on the text it generates in a wide variety of domains?