Scholarlink

<p>The rapid advancement of large language models (LLMs) such as GPT-3 and GPT-4 has raised serious concerns for academic integrity, as AI-generated essays often bypass conventional plagiarism detection systems and threaten the credibility of student assessments. To address this challenge, we propose XLNet-CNN, a hybrid detection framework that combines XLNet (Generalized Autoregressive Pretraining for Language Understanding) with a convolutional neural network (CNN) for local feature extraction. Using the EnglishQA Essays Corpus containing 161,640 essays from both human and AI sources, we benchmark the model against representative machine learning, deep learning, and transformer-based baselines, as well as widely used external detection tools. The framework achieves accuracy 0.98, recall 0.96, F2 score 0.97, precision 1.00, BLEU (Bilingual Evaluation Understudy) 30.76, and perplexity (PPL) 3.99. Notably, it eliminates false negatives, ensuring that no AI-generated essay is misclassified as human-written, a critical safeguard for high-stakes academic contexts. These results demonstrate the effectiveness of hybrid architectures in capturing both semantic coherence and stylistic artifacts, offering a reliable and scalable solution for maintaining fairness in academic evaluation while supporting responsible AI use in education.</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

Detecting AI-generated essays using fine-tuned XLNet-CNN hybrid techniques: a study of the academic integrity challenge

  • 期刊论文

  • 发表日期：

    2026年2月2日

查看全文
  • Manish Prajapati,
  • Santos Kumar Baliarsingh,
  • Prabhu Prasad Dev

摘要

The rapid advancement of large language models (LLMs) such as GPT-3 and GPT-4 has raised serious concerns for academic integrity, as AI-generated essays often bypass conventional plagiarism detection systems and threaten the credibility of student assessments. To address this challenge, we propose XLNet-CNN, a hybrid detection framework that combines XLNet (Generalized Autoregressive Pretraining for Language Understanding) with a convolutional neural network (CNN) for local feature extraction. Using the EnglishQA Essays Corpus containing 161,640 essays from both human and AI sources, we benchmark the model against representative machine learning, deep learning, and transformer-based baselines, as well as widely used external detection tools. The framework achieves accuracy 0.98, recall 0.96, F2 score 0.97, precision 1.00, BLEU (Bilingual Evaluation Understudy) 30.76, and perplexity (PPL) 3.99. Notably, it eliminates false negatives, ensuring that no AI-generated essay is misclassified as human-written, a critical safeguard for high-stakes academic contexts. These results demonstrate the effectiveness of hybrid architectures in capturing both semantic coherence and stylistic artifacts, offering a reliable and scalable solution for maintaining fairness in academic evaluation while supporting responsible AI use in education.