Detecting Machine-Generated Text in Academic Writing: Stylometric Fingerprinting of Humans and Large Language Models for Authorship Verification

The proliferation of Large Language Models (LLMs) such as GPT-۴ and ChatGPT has introduced significant challenges to maintaining academic integrity, necessitating robust methods for distinguishing human-written texts from machine-generated content. This paper provides a comprehensive review of contemporary techniques for detecting machine-generated text in academic writing, with a particular focus on stylometric fingerprinting and machine learning approaches. We systematically analyze key methodologies, including stylometric analysis (lexical diversity, syntactic complexity, and punctuation patterns), psycholinguistic mapping, and the trigram-cosine delta metric. Furthermore, we examine advanced machine learning models such as supervised classification, ensemble learning, and Graph Neural Networks (GNNs) integrated with pre-trained language models. The review also explores multilingual and cross-domain detection strategies, benchmark datasets, and performance evaluation metrics. Despite high accuracy rates reported in recent studies (up to ۹۸%), significant challenges remain regarding generalizability across different LLMs and domains, computational efficiency, and ethical considerations related to privacy. The paper concludes that integrating stylometric analysis with advanced machine learning offers a promising pathway for safeguarding academic integrity, while emphasizing the need for continued research to address existing limitations.

کلیدواژه ها:

Machine-Generated Text Detection ، Academic Writing ، Stylometry ، Large Language Models ، Authorship Verification