A Comparative Study of Plagiarism Detection Software Against AI-Generated Text

Abstract

The proliferation of large language models (LLMs) such as ChatGPT, Gemini, and Claude has fundamentally destabilised the assumptions on which conventional plagiarism-detection software was built. Traditional similarity-matching engines — Turnitin, Grammarly, Quetext, iThenticate, and Copyscape among them — were engineered to catch verbatim or near-verbatim copying by comparing submitted text against indexed corpora of existing publications. AI-generated text, by contrast, is often statistically novel: it does not copy any single source verbatim, yet it may still represent an act of academic or intellectual dishonesty when submitted as original human work. This paper undertakes a comparative investigation of contemporary plagiarism-detection and AI-detection software, examining their design logic, technical architecture, and empirical performance against AI-generated and AI-paraphrased text. Drawing on published benchmarking studies, vendor-disclosed statistics, and independent audits, the paper compiles comparative tables of accuracy, false-positive, and false-negative rates across more than a dozen widely used detection systems. It further synthesises theoretical and practical literature on plagiarism-prevention pedagogy, situating the discussion within the broader scholarship on academic integrity, authorship, and scholarly publishing ethics.¹⁻³ The findings indicate a persistent and, in several cases, widening gap between the promises made by commercial detection vendors and the tools' demonstrated reliability, particularly once AI-generated text is paraphrased, translated, or blended with human writing. The paper concludes that no single tool can be treated as a definitive arbiter of authorship, and it proposes a multi-layered, human-in-the-loop framework for institutions, publishers, and individual researchers that combines technical detection with process-based verification, aligning with recommendations already advanced in the plagiarism-prevention literature.¹,² The paper contributes an integrative, cross-disciplinary account of where detection technology currently stands, why it struggles against generative AI, and what a more defensible verification ecosystem might look like.