Detection of AI-Generated Text in Academic Writing Using Transformer-Based Models

Abstrak

The rapid advancement of generative artificial intelligence (AI), particularly large language models (LLMs), has significantly transformed academic writing by enabling the automatic generation of coherent and contextually relevant text. While these technologies improve writing efficiency and accessibility, they also raise serious concerns regarding academic integrity, originality, and authorship. Conventional plagiarism detection tools are ineffective in identifying AI-generated content because such text is often original rather than copied. This study proposes a transformer-based approach for detecting AI-generated text in academic writing by evaluating the performance of four pre-trained language models: BERT, RoBERTa, DistilBERT, and DeBERTa. The research methodology consists of dataset collection, text preprocessing, dataset splitting, transformer model fine-tuning, and performance evaluation using Accuracy, Precision, Recall, F1-score, and ROC-AUC. Experimental results show that all evaluated transformer models achieved excellent classification performance, with DeBERTa producing the highest accuracy of 98.10%, followed by RoBERTa (97.20%), BERT (95.10%), and DistilBERT (93.40%). These findings demonstrate that transformer-based architectures effectively capture contextual and semantic characteristics that distinguish AI-generated text from human-authored academic writing. The proposed approach provides a reliable solution for supporting academic integrity and assisting educators, publishers, and research institutions in detecting AI-generated content within scholarly documents.