What happened?

Sebastian Raschka, a well-known researcher in artificial intelligence and machine learning, published an article on August 15, 2026, addressing how to build an AI text detector from scratch. The starting point of the article was the built-in AI text detection feature that Substack recently added to its user interface.

Raschka notes that this feature likely relies on an approach similar to Pangram's models. He himself plans to develop a classifier working on similar logic, giving texts an AI probability score between 0 and 100. To do this, he will train a small language model called DistilBERT using fine-tuning.

Why does it matter?

The project isn't limited to just building a detector. Raschka also aims to use the same system as a verifier to train a small language model (SLM) that generates text capable of evading detection. This approach goes beyond classic reasoning models trained on math and code, offering an example of verifier-based large language model (LLM) applications.

The author argues that such a detector could be used in practice both to filter spam content and to correct personal writing without making it look AI-generated. For example, it could be possible to fix a text's grammar mistakes while ensuring the text is still scored as "0% AI-generated."

What's known

  • Substack integrated the AI detector directly into its user interface.
  • The system to be developed will be able to provide scores both for the entire text and for individual text segments.
  • A DistilBERT classifier will be used as the model.
  • AI text detection methods include supervised classifiers, perturbation-based probability tests, perplexity measurements, and watermarking.
  • Raschka had also written a separate article in 2023 covering different detection approaches on the topic.

What's next?

The article Raschka shared represents the first stage of the project, with subsequent plans including training the model, evaluating it, deploying it locally, and developing a user-friendly interface. The author reminds readers that AI detectors are part of a constantly evolving cat-and-mouse game, as new models may evade detection patterns and detectors need to be updated accordingly.