Human and AI Written Text Detection Using Deep Learning and Machine Learning

This research investigates classifying human and AI-generated text, which is crucial in this era of AI technology and the evolution of large language models (LLMs). This highlights a comparative approach using classical machine learning and deep learning models to enhance detection accuracy. A private dataset with 1.6 million text samples (1,000,001 human-written samples and 608,397 AI-generated samples) has been introduced for model training to detect human and AI-generated text. Four models—Decision Tree, Gradient Boosting, LSTM (Long Short-Term Memory) and Bi-GRU (Bidirectional Gated Recurrent Unit) are trained and evaluated to ensure optimal performance. Decision Tree and Gradient Boosting have been chosen for their ability to capture non-linear relationships, where LSTM can capture long-term dependencies and Bi-GRU understands the context through bidirectional processing for sequential data. The Bi-GRU model and LSTM model outperform, where Bi-GRU has the highest validation accuracy of 99.69% and LSTM has the validation score of 99.68%. And both have perfect F1-Score of 100%. However, the classical models, the Decision Tree and Gradient Boosting models have lower accuracy of 97.51% and 95.95% respectively. The outstanding performance of recurrent neural network variants proves their ability to work on complex sequential tasks like text classification. Visualization techniques like word clouds reveal distinct linguistic patterns in AI versus human texts. A web application has been developed using Streamlit enabling detection of content in the real world, such as plagiarism and content authenticity.