No choice

On the stylistics of AI-generated texts

Abstract

This paper explores the stylistics of AI-generated texts from a theoretical perspective. It applies sociolinguistic and pragmatic style theories, defining style as meaningful choice among alternatives, to point out the similarities and differences between human and AI-generated styles. An experiment with ChatGPT, Claude, and Gemini demonstrates LLMs’ remarkable ability to rewrite a narrative in diverse styles (e.g., “conversational”, “formal”, “emotive”) based on simple prompts. The outputs show consistent stylistic features across models for given styles, which is confirmed by stylometric clustering. Despite this flexibility, the paper argues that LLMs do not make choices in the human, meaning-generating sense; they are mechanistic, probabilistic systems. Their style imitation success is explained by reference to their training on human texts, which contain complex patterns of human stylistic choices and metapragmatic categorizations that are represented in the LLMs.