Exposing AI Paraphrasers: A Machine Learning Approach to Identifying Machine-Rewritten Plagiarized Content

Abstract

The proliferation of generative language models and automated paraphrasing tools has fundamentally altered the landscape of academic and professional plagiarism. Traditional plagiarism-detection systems, which rely predominantly on lexical string-matching and n-gram overlap, are increasingly circumvented by AI paraphrasers capable of producing semantically equivalent text with negligible surface-level similarity to the source [1,2]. This paper examines the technical, pedagogical, and ethical dimensions of machine-rewritten plagiarism and proposes a hybrid machine learning framework for its identification. The framework fuses lexical, syntactic, and transformer-derived semantic features within an ensemble classification architecture, and is evaluated against illustrative benchmark configurations that synthesize patterns reported across the paraphrase-detection literature. The paper further situates the technical contribution within a broader review of plagiarism theory and detection practice, drawing on institutional and journal-level scholarship on plagiarism-free publishing [1], detection techniques and tools [2], and the role of artificial intelligence in combating plagiarism in scholarly publishing [3]. Findings suggest that hybrid semantic-lexical models substantially outperform single-feature baselines in recovering meaning-preserving rewrites, though performance remains sensitive to paraphrase intensity, domain shift, and adversarial re-paraphrasing. The paper concludes with a multi-perspective discussion covering technical, editorial, and policy implications, and proposes a research agenda for robust, explainable, and ethically deployed paraphrase-plagiarism detection systems.