AI-Generated Text Detection

The emergence of ChatGPT, a highly advanced generative artificial intelligence (AI) tool, has posed a great challenge for teaching staff in maintaining academic integrity within the classroom. This work presents effective strategies and solutions for addressing this important issue. We specifically introduce a strong transformer-based language model with fine-tuning strategies and a hybrid convolutional neural network–bidirectional long short-term memory-bidirectional encoder representations from transformers (CNN-BiLSTM-BERT) architecture for AI-generated text detection. Preprocessing involved subjecting text to various normalization steps, including case folding into lowercase, tokenization, stop-word removal, and token, segment, and position embeddings. Spaces and digits were also removed to further enhance data consistency and quality. Our experimental research showed state-of-the-art (SOTA) classification performance for distinguishing between essays written by large language models (LLMs) and students. The baseline Hybrid CNN-BiLSTM model showed a stellar ROC-AUC score of 98.92%, which indicates its superior classification accuracy. Similarly, the BERT-cased model also achieved an ROC-AUC value of 97.98%, showing its efficacy for AI-generated text detection. Furthermore, data augmentation with external datasets from various AI models like generative pretrained transformer (GPT) also significantly boosted detection accuracy. This attests to the important role played by diverse training data for enhancing model robustness and generalization. Our findings emphasize the paramount importance of strategic model choice, comprehensive parameter tuning, and consolidation of external data sources for enhancing the accuracy and dependability of AI-generated text detection models. Our work contributes to ongoing efforts toward preserving academic integrity through providing effective tools for validating student essay originality and ensuring responsible AI utilization for educational purposes.