Stylometric detection of AI-generated texts: evidence from human and machine-written essays

Abstract The rise of large language models such as ChatGPT has intensified debates about authorship, originality, and integrity in academic and creative writing. Distinguishing between human- and artificial intelligence (AI)-generated texts is therefore not only a technical task but also a pressing concern for the digital humanities, where style, creativity, and attribution remain central. In this study, we adapt stylometric techniques to this new setting, introducing a dataset of paired human- and AI-authored essays across 110 subject areas. We evaluate three established classifiers, analysing how essay length, training data size, and topical variation influence performance, and whether human writing is better represented as a single category or as distinct authorial voices. Our findings show that while AI-generated texts exhibit striking stylistic uniformity, human writing is marked by variability and individuality. This contrast demonstrates both the continuing effectiveness of stylometry for AI detection and its wider relevance for authorship, originality, and voice in the age of generative AI.