Meaning by Courtesy: LLM-Generated Texts and the Illusion of Content

Abstract

Recently, Mann et al.¹ proposed the use of “personalized” Large Language Models (LLMs) to create professional-grade academic writing. Their model, AUTOGEN, is first trained on a standard corpus and then “fine-tuned” by further training on the academic writings of a small cohort of authors. The resulting LLMs outperform the GPT-3 base model, producing text that rivals expert-written text in readability and coherence. With judicious prompting, such LLMs have the capacity to generate academic papers. Mann et al. even go so far as to claim that these LLMs can “enhance” academic prose and be useful in “idea generation”¹. I argue that these bold claims cannot be correct. While we can grant that the sample texts appear coherent and may seem to contain “new ideas”, any appearance of coherence or novelty is solely “in the eye of the beholder” (Bender et al.²). Since the generated text is not produced by an agent with communicative intentions (Grice 1957³) our ordinary notions of interpretation – and, derivatively, of such notions as coherence – break down. As readers, we proceed with the default assumption that a text has been produced in good faith, naturally trusting what it says to be true (absent indications to the contrary) and expecting these truths to form a coherent whole. But this default assumption is misplaced in generated texts and, if unchecked, will allow both falsehoods and inconsistencies to pass under our radar. Whatever one thinks of the use of LLMs to help create content for commercial publications, their use in generating articles for publication in scientific journals should raise alarms.