Evaluating the Misuse of Generative Language Models in Academic Writing

Abstract

The emergence of large language models (LLMs) such as ChatGPT has introduced a distinct category of academic misconduct that traditional plagiarism frameworks are ill-equipped to address: AI-assisted plagiarism, sometimes labeled "AI-giarism" in the literature [1,2]. Unlike classical plagiarism, which involves copying an identifiable source, AI-assisted plagiarism involves the generation of superficially original but unearned text, raising questions about authorship, originality, and epistemic responsibility that predate but are magnified by generative AI. This paper synthesizes recent empirical literature (2023–2026) on the prevalence of generative AI use in student and scholarly writing, the technical performance and fairness of AI-text detection systems, institutional and publisher policy responses, and the divergent perspectives of students, faculty, and editors. Drawing on survey data from over a dozen independent studies collectively covering more than 100,000 respondents [19,21,22,25,27,29,30], detection-accuracy benchmarks reporting false-positive rates ranging from below 1% to above 60% depending on population and tool [13,17,19,20,45], and publisher policy documents from COPE, Elsevier, Wiley, and other bodies [31–40], the paper argues that AI-assisted plagiarism is best understood not as a single behavior but as a spectrum of practices whose ethical status depends on disclosure, task type, and degree of cognitive offloading. The paper further shows that current detection technology is structurally unreliable and systematically biased against non-native English writers [45,47–50], meaning that punitive detection-first policies risk institutional harm disproportionate to the problem they aim to solve. The paper closes with a multi-stakeholder framework for policy design and identifies priority areas for future research.