A Multimodal Approach for Detecting AI Generated Content using BERT and CNN

Abstract

With the advent of Generative AI technologies like LLMs and image generators, there will be an unprecedented rise in synthetic information which requires detection. While deepfake content can be identified by considering biological cues, this article proposes a technique for the detection of AI generated text using vocabulary, syntactic, semantic and stylistic features of the input data and detecting AI generated images through the use of a CNN model. The performance of these models is also evaluated and benchmarked with other comparative models. The ML Olympiad Competition dataset from Kaggle is used in a BERT Model for text detection and the CNN model is trained on the CIFAKE dataset to detect AI generated images. It can be concluded that in the upcoming era, AI generated content will be omnipresent and no single model will truly be able to detect all AI generated content especially when these technologies are getting better.