Researchers train classifier to detect AI-written blog posts with 98% accuracy using structure

Startup Sitefire, founded by Stanford and TUM alumni and backed by Y Combinator, trained a classifier to distinguish AI-generated blog posts from human-written ones based purely on structural features rather than word choice. The team collected 2,250 pre-ChatGPT blog posts from 268 B2B websites and had five leading AI models rewrite each one, then used a secondary AI to answer 214 structural questions about every post. The resulting classifier achieved 98% accuracy on unseen posts, correctly identifying 1,721 of 1,740 examples. A key finding was that AI-generated posts follow a predictable pattern — repeating the main point in the title, introduction, and conclusion — with 77% of AI posts ending by restating their thesis versus only 12% of human posts. The structural approach proved robust even when AI models heavily paraphrased their own output, suggesting simple rewording cannot evade detection.