Why modern AI detectors get it wrong
A detector returns a number about style, not about who wrote the text, so it can be wrong in both directions. Polished human writing gets high scores, and good AI text passes.
Reading it as a verdict is a mistake. Here is why, with our own measurements.
A detector does not know who wrote it
It cannot see how a text was made. It measures patterns: how even the sentences are, how many typical phrases there are, how predictable the text is. Indirect signs. People produce them too.
Polished human prose looks like machine prose
News, business letters and academic papers follow rules and look even. In our own measurements, neatly edited human news scored high noticeably more often than casual comments and discussions. On news texts the signal barely works at all, so we advise abstaining there. A score that ignores genre can't be trusted.
People writing in a second language
Writers using a language that is not their first often choose simple, "correct" constructions, or rely on machine translation altogether, and detectors frequently mistake such text for machine output. Because of these false alarms, several large universities, including Yale and the University of Waterloo, stopped using built-in AI detectors.
Our own check on volunteer-translated texts showed slightly more false alarms than average, though the gap is small. On non-native Ukrainian writing we could not tell the difference at the size we tested. That is no reason to relax. Be careful with anyone writing in a second language.
Short texts
There is simply not enough to measure. At about a hundred words or fewer any statistical conclusion is fragile, and an honest detector should say so. In our corpus test, texts of 80 to 299 words were close to random, while texts of 300 words and more were noticeably better, though still far from reliable.
Modern models
Our first calibration on 34 AI answers and 20 human texts looked good: 93.9% of AI texts were caught at threshold 50 with no false alarms. The set was small and limited in genre. On a much larger corpus the picture is humbler: at a 1% false-positive rate the analysis catches roughly one AI text in seven. Newer models evade the signals better than the early web chats.
Editing and paraphrase
If a person edits a machine draft or runs it through a paraphraser, the signs blur. Human text edited by a model is almost never flagged in our tests, so mixed authorship is essentially invisible. So we talk about the share of text with signs of generation, not an "AI percentage".
How to read a result
Treat it as a hint and nothing more: it tells you where to reread. Look at the specific highlighted places and consider genre and length. For disputed cases there are better ways: drafts, edit history, a conversation with the author.
What to do next
Run your own text through the checker and open the evidence list, not just the number. If you are the one judging someone else's text, work through the fair checklist. If you were the one flagged, see our story in when the detector mistakes a human for AI, and for the wider picture read how AI text detection works today.