A Novel Approach to Automated Detection of AI-Generated Text

摘要

Detecting machine-generated text involves identifying whether text has been created by artificial intelligence models or written by humans. This task has become increasingly significant due to the potential misuse of AI-generated text for producing fake news, reviews, or spam that can mislead people. The aim of this study is to develop a model capable of determining if a tweet's author is human or a robot. To achieve this, we utilized a zero-shot prompt with a pre-trained model and fine-tuned SBERT using various transformer models. Additionally, we employed graph attention network and graph convolutional network models to analyze the author's writing style. The findings indicate that using the graph convolutional network model to extract writing style characteristics yields the highest accuracy, reaching 93.60%. Detecting machine-generated text is vital for preventing the abuse of AI models and ensuring the reliability of content on online platforms by effectively distinguishing between human and AI-generated text.