---
title: "SlopShape spots AI-written blog posts by their structure"
description: "An arXiv paper that identifies AI-generated company blog posts from how they are structured."
date: 2026-09-26T12:00:00+02:00
lastmod: 2026-11-03T00:46:22+01:00
url: https://klimek.link/posts/2026-09-26-slopshape-ai-web-content/
author: "Marcin Klimek"
source: https://klimek.link/posts/2026-09-26-slopshape-ai-web-content/
---

# SlopShape spots AI-written blog posts by their structure

Source: [SlopShape on arXiv](https://arxiv.org/abs/2609.15369)

Jochen Madler wanted a detector that still works after someone rewords an AI text. Madler stopped looking at words and looked at structure: what a post says, in what order, with what evidence and in what voice. On company blogs the structural features found AI posts almost every time, even after each post had been reworded by the model that wrote it. The paper describes the AI posts as having "a tidy, self-announcing shape", and the human ones as scattered across rare shapes. The code and prompts are public.

**TL;DR**

- Author: Jochen Madler; arXiv, 14 September 2026
- Replicates StoryScope (Russell et al., 2026); tests commercial blog posts, not fiction
- Data: 2,250 human posts (pre-ChatGPT, 268 company domains) vs 11,250 AI mirror posts, five frontier models
- Instrument: 203 features, applied by an LLM
- Gold-session agreement: human-human kappa 0.939, human-model 0.951
- 176 structural features alone: 97.0 macro-F1 on held-out companies
- 96.1 when each AI post is reworded by its own model
- Source-model attribution: 68.6% vs 16.7% chance
- AI posts: tidy, self-announcing shape; human posts: rare structural configurations
- Released: pipeline, instrument, prompts, code, aggregate artifacts

Source: https://klimek.link/posts/2026-09-26-slopshape-ai-web-content/