Abstract:
This study focuses on the growing difficulty of recognizing text that produced by machines in a time when artificial intelligence is extensively used. The study proposes a novel method based on a Long Short-Term Memory (LSTM), GRU and Hybrid architecture to distinguish AI-generated content and human-written text with remarkable accuracy. By employing advanced techniques for text preprocessing, vectorization, and embedding, we achieved an efficient design with average computational demands. We tested the models on a large dataset, the model demonstrated outstanding performance. The best model achieved an accuracy of 98.31% and the best F1-score is 0.98. These findings show the outstanding ability of the model to generalize well on unseen data, proving the potential of using it in real-world applications. The model's stability and reliability are backed by highly similar outcomes of the training and validation phases with minimal overfitting due to excellent regularization strategies. The confusion matrices and the full classification reports gave in-depth insights into the model's strengths and weaknesses, thus enhancing its applicability.
