New Method Detects AI-Generated Text by Analyzing Sentence Transitions

Detecting text generated by artificial intelligence (AIGT) remains a significant challenge, as many existing methods struggle with overfitting to specific models or failing under shifts in data distribution. Current approaches often rely on token-level statistics or isolated stylistic features, which can be brittle. Researchers have identified a novel structural signal at the sentence-pair level: large language models (LLMs) tend to produce inter-sentence transitions with inflated variance, characterized by recurring similarity bursts at paragraph boundaries and templated connections, a phenomenon termed "Relational Over-Regularization" (ROR). This ROR signal indicates that LLMs deviate from human writing patterns in how they connect sentences. To exploit this, a new graph-based framework called the Cross-Source Stylometric Fingerprint Graph (CSFG) has been developed. CSFG encodes various signals, including positional, sequential, semantic, and crucially, transition deviation, as learnable edge features within a Graph Neural Network (GNN). The per-edge signed deviation directly operationalizes ROR without requiring manual thresholds, also serving as a false-positive calibrator. The CSFG framework demonstrates impressive performance, achieving 97.14% accuracy in binary AIGT detection, significantly outperforming previous graph-based baselines. It also maintains a low false-positive rate and shows robust generalization to LLMs not seen during training, particularly those exhibiting the inflated-variance characteristic. This method offers a more resilient approach to AIGT detection by focusing on the structural coherence of text rather than superficial features.