Guides

How AI detectors work, and why they flag human writing

Classifiers learn what AI text looks like

Six of the seven detectors we tested are classifiers. Their makers took a language model such as RoBERTa or DeBERTa, showed it many labelled examples of human and AI writing, and trained it to output a probability that a new text is AI-written. They learn whatever separates the two piles in their training data: word choice, sentence rhythm, topics, even formatting.

That makes them only as good as their examples. The OpenAI RoBERTa detector learned GPT-2 output and caught 9.5% of our Claude texts. TMR seems to have learned something about line breaks: it flagged every multi-paragraph human text in our set and far fewer once the line breaks were removed, as the paragraph break test shows.

Zero-shot detectors measure predictability

Fast-DetectGPT needs no training examples. It runs a language model over the text and checks, word by word, whether the actual word was more predictable than the words the model itself would have picked. Language models write by choosing likely words, so their text sits close to the model’s own expectations. People are less predictable on average.

It worked better than most classifiers in our test, with an AUROC of 0.815, but it inherits the weakness of the idea. Human writing that follows a strict template is predictable too. Fast-DetectGPT flagged 40.0% of human academic abstracts and none of the human fiction passages.

Why human writing gets flagged

Both methods reward the same thing: text that looks typical. Genres with fixed conventions, such as news reports, abstracts and encyclopedia articles, produce typical text whoever writes them. In our test, Fakespot flagged 70.8% of human news reports and MAGE flagged 40.0% of human encyclopedia articles.

Editing pushes the other way. Removing common AI phrasing with a humanizer made Claude Opus 5.5 text less typical, and Desklib’s catch rate on it fell from 62.7% to 32.5%.