Scholarlink

<p>As generative AI tools become ubiquitous in education, the need for reliable, transparent, and efficient methods to ensure academic integrity is paramount. While deep learning models offer high accuracy, their computational cost and lack of interpretability often hinder practical deployment. To address this gap, this study evaluates the practical viability of classical machine learning frameworks for detecting AI-generated text. The framework utilizes TF-IDF features and evaluates six classifiers through a rigorous 10-fold stratified cross-validation. Our results demonstrate that multiple classical models, including Support Vector Machines (SVM) and Naive Bayes, achieve stable and near-perfect mean F1-scores (0.9986 and 0.9984, respectively). Crucially, our analysis of the performance-efficiency trade-off highlights that Naive Bayes and Logistic Regression provide this top-tier accuracy with significantly lower computational cost, making them highly practical solutions for real-world deployment. Explainability was enhanced using SHAP and LIME to ensure transparency. This study contributes a robustly evaluated, lightweight, and interpretable framework for AI text detection suitable for educational applications, emphasizing that efficiency is as critical as accuracy for widespread adoption.</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

Educational integrity in the age of generative AI: an interpretable random forest-based detection framework

  • 期刊论文

  • 发表日期：

    2026年4月22日

查看全文
  • Yunhang Liu,
  • Xiaoning Zhang

摘要

As generative AI tools become ubiquitous in education, the need for reliable, transparent, and efficient methods to ensure academic integrity is paramount. While deep learning models offer high accuracy, their computational cost and lack of interpretability often hinder practical deployment. To address this gap, this study evaluates the practical viability of classical machine learning frameworks for detecting AI-generated text. The framework utilizes TF-IDF features and evaluates six classifiers through a rigorous 10-fold stratified cross-validation. Our results demonstrate that multiple classical models, including Support Vector Machines (SVM) and Naive Bayes, achieve stable and near-perfect mean F1-scores (0.9986 and 0.9984, respectively). Crucially, our analysis of the performance-efficiency trade-off highlights that Naive Bayes and Logistic Regression provide this top-tier accuracy with significantly lower computational cost, making them highly practical solutions for real-world deployment. Explainability was enhanced using SHAP and LIME to ensure transparency. This study contributes a robustly evaluated, lightweight, and interpretable framework for AI text detection suitable for educational applications, emphasizing that efficiency is as critical as accuracy for widespread adoption.