Hybrid deep learning framework for AI-generated text detection

Abstract

Recently, large language models have demonstrated remarkable ability to generate human-like textual content. This advancement has increased the need for reliable methods to detect AI-generated texts, particularly to address concerns related to academic dishonesty and the spread of misinformation on social media. This study introduces a hybrid deep learning model, which integrates Bidirectional Long Short-Term Memory networks (BiDLSTM), Transformer blocks, and one-dimensional Convolutional Neural Networks (1D CNNs) to distinguish between AI-generated and human-written texts. The model was evaluated on two diverse datasets, DAIGT and HC3, containing thousands of samples generated by various large language models alongside human-written texts. The proposed model achieved promising performance, with test accuracies of 99.47% on the DAIGT dataset and 97.09% on the HC3 dataset. These results suggest that the hybrid architecture can effectively capture subtle linguistic and stylistic differences between AI-generated and human-written content. This study contributes to ongoing efforts in AI-generated text detection and supports applications aimed at maintaining content authenticity and academic integrity.