Updated
Updated · magazine.sebastianraschka.com · Aug 15
Author Builds 0-100 AI Text Detector Tutorial After Substack Adds Feature
Updated
Updated · magazine.sebastianraschka.com · Aug 15

Author Builds 0-100 AI Text Detector Tutorial After Substack Adds Feature

3 articles · Updated · magazine.sebastianraschka.com · Aug 15

Summary

  • A new tutorial walks readers through building an AI text detector that scores text from 0 to 100 and can also serve as a verifier for training small language models to avoid detection.
  • Substack’s recent rollout of a built-in AI detector and reader demand for local do-it-yourself LLM demos prompted the project, which is framed as an educational case study in how such systems work.
  • The project aims to deliver a functional detector API and local browser UI that score whole documents and highlight suspicious text chunks, with potential uses in spam filtering and editing human-written text without making it sound AI-generated.
  • The method will fine-tune a DistilBERT classifier, similar to approaches the author says resemble Pangram-style models behind Substack’s feature, while stressing that AI detection remains a cat-and-mouse game prone to false positives.

Insights

If grammar checkers trigger AI detectors, could Substack's new tool accidentally penalize the platform's most polished human writers?
Could building a custom AI verifier actually provide language models with the exact blueprint needed to permanently evade detection?