Training a model to identify AI-generated web content from structure alone

Researchers developed a method called SlopShape that identifies AI-generated commercial web content by analyzing structural patterns—such as how information is organized, presented, and voiced—rather than individual words. The model achieved 98% accuracy in detecting AI-generated posts from five frontier models, and could attribute posts to their source model with 79.3% accuracy, while remaining robust even when AI-generated text is reworded. The findings reveal that AI-generated content shares distinctive structural characteristics that differ markedly from human-written posts.