AI-Generated Text Detection using ISGraphs and Graph Neural Networks

The growing presence of AI-generated text in online environments has raised concerns around misinformation, academic fraud, and content manipulation. To address this, we propose a graph-based detection system that combines Integrated Syntactic Graphs with Graph Neural Networks to distinguish between human and machine-generated text. Our approach leverages syntactic dependency structures and contextual embeddings from pre-trained language models, showing strong performance across multiple test scenarios, including clean, short, Unicode, and paraphrased variants. Our results demonstrate the robustness and adaptability of the text graph approach in different AI-generated text detection scenarios. This study was part of our PANCLEF 2025 Voight-Kampff AI Detection Sensitivity submission, which ranked 2nd of 27.