What is it measuring, exactly?

Most AI text detectors look for statistical patterns: predictable word choices, sentence rhythms or features learned by a classifier. Some compare a passage with the behavior of language models. None can see the author's notes, browser history or half-finished drafts. The final text is all it has.

That limitation becomes obvious with mixed writing. A person may draft the argument, use a grammar checker, ask a model for three alternatives and then rewrite every sentence. Calling the outcome simply 'human' or 'AI' hides the process the user probably wants to understand.

Short, familiar phrases provide little evidence

A brief email saying 'Thank you for your message; please find the invoice attached' is predictable because the situation is predictable. Lists, headlines, template letters and heavily quoted passages offer similarly thin evidence. A detector may return a number anyway, but there is not much authorship signal in the sample.

Longer continuous prose is usually more informative, though genre still matters. Technical instructions can repeat themselves. Exam answers may share the vocabulary of a marking scheme. Writers using a second language sometimes choose safer, more regular constructions. Those are ordinary human reasons for text to look statistically plain.

Editing ruins the clean laboratory question

Published research has shown that paraphrasing can weaken detection. Outside a study, text is also translated, shortened, expanded, proofread and moved between tools. A detector evaluated on untouched model output is facing a different task when it meets a polished document from a real workflow.

The reverse error matters just as much: predictable human prose can be flagged. A threshold that helps researchers screen a million documents may be completely unsuitable for deciding whether one student cheated.

How to review a disputed passage

Start with drafts, notes, references and version history. Ask the writer to explain an argument or describe how they developed it. Earlier work can provide context, although people improve and change style. Treat detector output as one additional source of information.

For an editor, the score may point toward fact-checking or a disclosure conversation. It cannot prove plagiarism, factual accuracy or ownership; those are separate questions. A confident number is not a shortcut through them.

Getting a less misleading result

Submit a substantial passage in its original language. Remove long quotations that were written by somebody else, and note whether translation or heavy editing took place. Do not repeatedly trim the sample until the tool gives the answer you expected.

If a person's education, job or reputation is involved, never let the score stand alone. Sometimes the only responsible conclusion is that the final text does not reveal its authorship with enough confidence. That may be unsatisfying, but it is much fairer than pretending the detector knows more than it does.

Automated results require source and context review.

Continue with independent verification.

Check a text sample
Sources

Primary reading

We use original standards, regulators, public institutions and research papers wherever possible. Sources were last checked on 12 August 2026.