AI text detector

Signs, not secrets

Many AI text detectors return a single percentage from a model you can’t inspect. GPTTrace starts from the opposite end: a catalogue of observable signs, most of them documented by Wikipedia editors who have spent years cleaning chatbot-written text out of the encyclopedia. Every sign GPTTrace finds is highlighted in your text and explained, so you can judge for yourself whether it fits.

The catalogue merges two sources. One is an era-weighted vocabulary — words like “delve”, “tapestry” and “showcasing” that chat models used far more often than people, weighted by how recently each became a tell. The other is the structural and formatting patterns from Wikipedia’s Signs of AI writing: inflated claims of significance, “not just X, but Y” parallelism, superficial “-ing” clauses, vague attributions, collaborative phrases left in, Markdown and chat-interface citation tokens, and invisible Unicode characters.

Statistics and the deep scan

Alongside the signs, GPTTrace measures properties of the whole text: how much sentence length varies (people mix short and long sentences; models are more even), how uniform paragraph sizes are, how varied the vocabulary is, and how much the text shares phrasing with typical chatbot prose, measured by compression. These catch texts that avoid the obvious words.

The optional deep scan adds a RoBERTa neural classifier fine-tuned on the RAID benchmark, which covers many generators and adversarial rewrites. It runs locally after a one-time download. In our tests it is far stronger than the signs alone on modern chatbots, but it over-flags encyclopedic and technical human writing when used by itself. So GPTTrace combines it with the other evidence using weights fitted on labelled data, with the threshold set to keep false accusations rare.

Detectors for specific chatbots and uses