A new Show HN project trains a model to classify web pages as AI-generated or human-written using only structural features such as tag nesting, class patterns, and layout ratios. The work was posted on Hacker News where it received 38 points and 8 comments, and the underlying paper is available on arXiv.
The method extracts features from the DOM tree and CSS grid properties rather than token distributions or image artifacts. This makes detection possible even when text is heavily edited or images are swapped.
What It Is / How It Works
The model ingests parsed HTML and builds a feature vector from element depth histograms, attribute entropy, and grid-template ratios. These signals are fed into a lightweight classifier that outputs an AI probability score.
Because the approach ignores textual content, it remains effective against paraphrasing attacks that defeat token-based detectors.
Benchmarks and Early Results
No public accuracy numbers were released in the initial post. Early testers on Hacker News noted that the structural classifier caught several commercial AI site builders that text-only tools missed.
The project emphasizes low compute requirements, running inference on CPU in under 50 ms per page.
How to Try It
The arXiv paper at https://arxiv.org/abs/2609.15369 contains the feature extraction code and training scripts. Clone the linked repository, run the provided preprocessing script on a Common Crawl sample, then train the classifier with the supplied config.
Community nodes for integration into scraping pipelines are already appearing on GitHub.
Pros and Cons
- Works on pages where text has been manually rewritten
- Low latency and no need for large language model inference
- Limited to web pages; does not generalize to documents or code
- May degrade if AI tools adopt more varied structural templates
Alternatives and Comparisons
| Approach | Input Features | Speed | Robust to Paraphrase |
|---|---|---|---|
| Structural classifier | DOM + CSS metrics | <50 ms | High |
| GPTZero / Originality | Token statistics | 200 ms | Low |
| Watermark detectors | Model logits | Varies | Medium |
Who Should Use This
Developers building content moderation pipelines or search engine filters will find the structural signal useful as an additional feature. Researchers studying AI adoption on the web can apply it at scale without heavy GPU resources.
Skip it if the target content is primarily non-HTML, such as PDFs or source code repositories.
Bottom Line / Verdict
Structural signals provide a fast, content-agnostic complement to existing AI detectors and deserve inclusion in any production pipeline.
The approach highlights a practical direction for detection research that keeps pace with rapidly improving generative tools.
Top comments (0)