Building an AI Text Detector From Scratch

A tutorial project outlines how to build an AI text detector from scratch, presented as an educational exercise for studying the limitations of such detectors and for exploring verifier-based LLM applications beyond reasoning models trained on math and code.

The write-up notes that Substack recently launched an AI detector feature in its UI, and that the project's method is similar to Pangram models, which the author says are, as far as they know, behind Substack's AI detection feature.

The detector is described as a supervised classifier that returns a 0-100 score, essentially a classifier with an estimated probability score denoting how likely a text is AI-generated according to the classifier. The author adds that the score is the classifier's estimated probability for the AI-generated class based on its training distribution, and should not be interpreted as a general probability that the text was written by AI. The plan is to fine-tune a DistilBERT classifier.

The stated goals include explaining how AI detectors can work, illustrating an end-to-end LLM project covering evaluation, training, and local deployment, and producing an AI-detector API usable by humans and agents along with a user-friendly UI. A preview of the local browser interface returns a whole-text AI score and can highlight scores for individual text chunks. The detector is also to be used as a verifier to train a small language model to produce text that avoids detection.

A disclaimer states that AI checkers are a cat-and-mouse game: a checker may learn a pattern indicative of AI-generated content, the next LLM may incidentally or deliberately not exhibit that pattern, and the checker then has to be updated, and so forth. It adds that false positives — human-written text flagged as AI-generated — are likely, with more to come later.

The author also mentions possible practical uses such as filtering spammy content and improving personal writing without turning it into AI-generated text, including the idea of asking a tool to fix grammar while ensuring text still scores 0% AI-generated. The piece references an earlier 2023 article on approaches for detecting LLM-generated content, listing supervised classifiers, perturbation-based probability tests, perplexity measures, and watermarking, and an early Substack article on fine-tuning large language models.