Comparison of algorithms for the recognition of ChatGPT paraphrased texts

Abstract

The rapid development of artificial intelligence, especially AI assistants, is leading to new forms of plagiarism that are difficult to detect using existing methods. Paraphrasing tools make this problem even more complex and challenging especially in minor languages with inadequate resources and tools. This study explores strategies to help detect plagiarism generated by ChatGPT 4.0 and altered by paraphrasing tools. We propose two new datasets consisting of abstracts of doctoral theses in English and Serbian. Both datasets were subjected to ChatGPT paraphrasing, which allowed us to form two classes of texts: human-written and AI-generated, i.e., AI-paraphrased. We then comprehensively compare 19 widely used classification algorithms based on two feature sets: word unigrams and character multigrams. In addition, we compare these to the results of a commercially available pre-trained ChatGPT content detector, ZeroGPT. The results on the English corpus turn out to be very accurate, achieving an accuracy of 95% or more. In contrast, the results on the Serbian corpus were less accurate, achieving an accuracy of just over 85%. Syntax analysis of the training datasets has shown that in Serbian GPT-paraphrased texts, 33.2% of sentences remain the same, and they are found in 63% of documents. GPT-paraphrased English texts showed that 3.2% of sentences remain the same, and they are found in 16% of documents. Syntax analysis of the test datasets has shown that the change of the model temperature influences syntactic features (average number of words and sentences) in English texts and slightly or not in Serbian texts. We attribute all these differences to GPT’s lower paraphrasing ability in minor languages such as Serbian. Presented findings underscore the necessity for making persistent effort in developing tools made for detecting AI-paraphrased texts in academic and professional settings, particularly for minor languages with limited NLP resources, to preserve content integrity and authenticity.