Assessing Bi-LSTM Model’s Performance in Identifying AI Generated Text in Digital Media

Abstract

The rapid growth of AI-generated content from ChatGPT, Gemini, and many others poses significant challenges in digital media. There are increasing instances of text generated by machine learning models being indistinguishable from human-written text. This calls for an effective way of detecting and distinguishing text generated by AI from human text. Using a Cross-Industry Standard Process for Data Mining (CRISP-DM) methodology, a Bidirectional Long Short-Term Memory (Bi-LSTM) model for classifying AI-generated text in a digital media context was developed. A carefully curated dataset of human and AI-generated text samples was used. The Bi-LSTM model without an embedding layer was implemented, optimizing the model to capture complex linguistic patterns apparent in each text type. An experimental setup was used to evaluate the effectiveness of the model. It was noted that the model achieved a remarkable test accuracy of 99.79%, with a loss of 0.009. To facilitate practical implementation, we developed a web application using Next.js with our model served from a Flask server that enables real-time AI text detection. Our results highlight the model’s ability to accurately identify AI-generated text, providing valuable insights into deploying such models to verify content on media platforms. This study highlights the potential of neural network-based classifiers to address the pressing need for automated AI text detection in an increasingly AI-influenced digital ecosystem.