Scholarlink

<p>Artificial intelligence (AI) tools, such as ChatGPT, possess remarkable text generation capabilities, making them highly valuable in academic writing. However, misusing such advanced technology can lead to unethical outcomes such as academic plagiarism, improper authorship claims, or the propagation of scientifically inaccurate information, each of which can compromise the integrity of scholarly communication. Therefore, institutions and individuals must adopt robust AI models capable of accurately distinguishing between Human-Written (HW), AI-Generated (AIG), and AI-Paraphrased (AIP) texts. In response to this critical ethical challenge, this paper proposes a high-performance classification framework based on the Random Forest algorithm, demonstrating strong efficacy in identifying and categorizing diverse text sources. Beyond technical classification, the proposed model contributes to preserving academic standards, enhancing editorial workflows, and shaping ethical AI policies in research and education. The model leverages four feature groups: Text Similarity Features (TSF), Lexical and Syntactic Features (LSF), Readability and Complexity Features (RCF), and Structural Features (SF), comprising 31 features in total. A dataset of 14,111 human-written abstracts from Scopus-indexed journals (2014–2024) was compiled. Using ChatGPT-4, we generated two additional datasets: 14,111 AIG abstracts based on article titles and 14,111 AIP abstracts from paraphrased originals. The final dataset contains 42,333 labeled samples. Extensive training and testing demonstrated the effectiveness of the Random Forest Classifier, achieving classification accuracies of 98.61% during training and evaluation. The results highlight the critical role of TSF in optimizing performance. Integrating TSF with other feature groups significantly improves accuracy, emphasizing the complementary benefits of combining feature sets. In conclusion, the proposed model effectively distinguishes text origins, reinforcing the value of integrated features in building ethical and reliable AI-based text analysis tools.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A high-accuracy classification model for AI-generated and AI-paraphrased texts

  • 期刊论文

  • 发表日期：

    2025年10月9日

查看全文
  • Ayman Odeh,
  • Haythem Bany Salameh,
  • Ghaleb Elrefae

摘要

Artificial intelligence (AI) tools, such as ChatGPT, possess remarkable text generation capabilities, making them highly valuable in academic writing. However, misusing such advanced technology can lead to unethical outcomes such as academic plagiarism, improper authorship claims, or the propagation of scientifically inaccurate information, each of which can compromise the integrity of scholarly communication. Therefore, institutions and individuals must adopt robust AI models capable of accurately distinguishing between Human-Written (HW), AI-Generated (AIG), and AI-Paraphrased (AIP) texts. In response to this critical ethical challenge, this paper proposes a high-performance classification framework based on the Random Forest algorithm, demonstrating strong efficacy in identifying and categorizing diverse text sources. Beyond technical classification, the proposed model contributes to preserving academic standards, enhancing editorial workflows, and shaping ethical AI policies in research and education. The model leverages four feature groups: Text Similarity Features (TSF), Lexical and Syntactic Features (LSF), Readability and Complexity Features (RCF), and Structural Features (SF), comprising 31 features in total. A dataset of 14,111 human-written abstracts from Scopus-indexed journals (2014–2024) was compiled. Using ChatGPT-4, we generated two additional datasets: 14,111 AIG abstracts based on article titles and 14,111 AIP abstracts from paraphrased originals. The final dataset contains 42,333 labeled samples. Extensive training and testing demonstrated the effectiveness of the Random Forest Classifier, achieving classification accuracies of 98.61% during training and evaluation. The results highlight the critical role of TSF in optimizing performance. Integrating TSF with other feature groups significantly improves accuracy, emphasizing the complementary benefits of combining feature sets. In conclusion, the proposed model effectively distinguishes text origins, reinforcing the value of integrated features in building ethical and reliable AI-based text analysis tools.