Detection of AI-written and Human-written Text using Deep Recurrent Neural Networks

With the development of Artificial Intelligence (AI)-based language models, it is becoming pertinent and will become even more pertinent in the future to be able to distinguish between AI-based generated text and human-based generated text. The implications of having humans present work generated by AI language models and claiming it their own have serious implications at all levels with the basic being the ethical implication. In this paper, we propose the use of modified deep learning models using the Deep Recurrent Neural Network (DRNN) for the classification of text to be either AI-generated or human-written. Two modified architectures are proposed DRNN-1 and DRNN-2. This led to the second contribution of this work which is the development of a dataset containing short answers to simple questions in Information Technology (IT), Cybersecurity, and Cryptography given to junior and senior students in Computer Engineering & Science, and IT to produce a total of 450 answers. The same questions were given to ChatGPT for a total of 450 answers. The combined dataset consisted of 900 answers in the three domains. Though both proposed architectures produced good results, the DRNN-2 achieved better results with a test accuracy of 83.78% using the cybersecurity questions alone and 88.52% using the combined total dataset. This is considered one of the very excellent results achieved in this new emerging field of research.