Detecting Academic Misconduct: A BERT-Based Approach to Identifying ChatGPT-Generated Content

The misuse of AI tools like ChatGPT in academia is on the rise, with students increasingly submitting AI-generated content as their own work, posing a significant challenge to academic integrity. This study addresses the urgent need for an effective solution to detect and prevent such practices by introducing a BERT-based model designed to distinguish between ChatGPT-generated and human-written text. Through thorough comparisons, transformers emerged as the most promising approach. We trained and evaluated three BERT variants—RoBERTa, DistilBERT, and BERT—using the HC3 dataset, which includes labeled samples of both AI and human-generated text. The dataset was divided into 70% for training and 30% for testing, with preprocessing steps such as removing duplicates. Training with TensorFlow resulted in all three models achieving 100% accuracy across all evaluation metrics, while the RoBERTa model trained with PyTorch achieved 99.2% accuracy. These findings highlight the effectiveness of transformer-based models in detecting AI-generated content. Our work, focused on academic settings, is limited to detecting text from ChatGPT 3.5. The best-performing model will be deployed using a Flask web application framework to provide a practical tool for upholding academic integrity.