Building an AI Text Detector From Scratch

Summary

Sebastian Raschka presents an educational project for building an AI text detector from scratch, prompted partly by Substack’s built-in detection feature and interest in local small-language-model projects. The project has two linked aims: explain how AI detectors work and use the detector as a verifier while training a small language model to produce text that avoids detection. The tutorial frames this as a broader example of building a scorer or verifier for an LLM application, rather than as a definitive test of authorship. The proposed system is intended to support both human users and agents through an API and a local browser interface that reports a whole-text AI score and can highlight scores for individual chunks. The article discusses several detection approaches, including supervised classifiers, perturbation-based probability tests, perplexity measures, and watermarking, but focuses on a classifier similar to the models believed to support Substack’s detector. The planned implementation fine-tunes DistilBERT to return a 0–100 score representing the classifier’s estimated probability for the AI-generated class under its training distribution. The author stresses that this score is not a general probability that a text was written by AI. The project also targets practical uses such as filtering spam and checking whether grammar or editing tools have made a person’s writing appear overly AI-like. Its central limitation is the adversarial, changing relationship between detectors and language models: a detector may learn one model’s patterns, while another model may avoid them deliberately or incidentally, forcing future updates. Human-written text can also be falsely flagged, so detection results require cautious interpretation.