Analyzing the efficacy of Deep Learning and Transformer models in classifying Human and LLM-Generated Text

Large Language Models (LLM) are widely accepted by the humans for their powerful ability to understand, follow and create human-like text. The text generation capabilities of LLMs have reached such a height that it is now comparable to text generated by human.LLMs are a powerful subset of Artificial Intelligence (AI)models, trained on huge amount of human-written text to create indistinguishable content. AI models have taken an important step ahead in constructing hyper-realistic content which can be in any modality such as text, image, and video. The output content generated by these models can be hard to distinguish from human-written content as these models are iteratively trained on texts written by human. Further this has got serious implications across various domains, including news reporting, information fraud, academic plagiarism, data manipulation, financial fraud, theft and deep fakes. The abilities of AI-generated text pose a significant challenge, blurring the lines of differentiation between human and machine-created content. Therefore, to effectively navigate the evolving landscape of AI-generated text, we need a robust and reliable AI text detection model. This study leverages a substantial dataset of 100,000 entries. empowering our model with adaptability across various domains. By employing diverse word embedding techniques along with machine and deep learning models, we explore optimal configurations for exceptional performance. Notably, we implement pre-trained models like BERT and DeBERTa on this extensive corpus, potentially representing the first such application. The resulting models demonstrate remarkable accuracy exceeding 95% across all configurations, signifying a robust advancement in differentiating AI generated text from human written one.