Transformer-Based Approach for Detecting LLM-Generated Scientific Text

Large Language Models (LLMs), such as Chat-GPT and Gemini, have become pivotal in generating human-like textual content. However, their ability to replicate human-authored scientific writing raises concerns regarding research integrity, particularly in mitigating the risks of fabrication and the dissemination of falsified data. Ensuring the credibility of scientific publications necessitates effective detection mechanisms. This study introduces Transformer Catch LLM (TcLLM), a transformer-based deep learning approach for detecting LLM-authored scientific text. TcLLM is designed to distinguish human-authored from LLM-generated scientific content. The AIGTxt dataset is expanded with Gemini-generated text to support this task, enhancing diversity in writing styles. Extensive experiments were conducted using TcLLM-BERT and TcLLM-RoBERTa to evaluate classification performance. The results demonstrate that TcLLM-RoBERTa outperforms TcLLM-BERT, achieving higher accuracy and lower misclassification rates in identifying LLM-generated scientific text.