What a result means, and what to do with it
Understanding results
If your text was flagged
“Likely AI-written” on my own writing — what now?
Start with the sentence on your result card: at your length band, in our sealed test about 2 in every 1,000 human texts of 150 words or more received a wrong AI signal, and about 2.4 in every 100 at 100 to 149 words. A wrong signal is rare, but it is not impossible, and the score is never proof. The guide walks through what to check, what to keep, and what to show a reviewer.
Read the full guide →Technical readers
Model identity, method summary and the sealed test are on the Evidence page; the full methodology has its own page. Evidence → · Methodology →
Frequently asked questions
The eight questions we are asked most often, answered with the measured figures from our sealed test.
What does the score actually mean?
It is an AI-writing signal score from 0 to 100: how strongly the text matches the patterns the frozen detector learned from AI writing. It is not a probability that a person or a model wrote the text, and it is not a measure of quality. The verdict band beside it is decided by thresholds that were fixed before our sealed test and have not moved since.
Can a result prove someone used AI?
No. A score is an estimate of AI-writing signal strength, not proof, and never enough on its own for a decision about a person. In our sealed test some human texts still received a wrong AI signal: about 2 in every 1,000 at 150 words or more, and about 2.4 in every 100 at 100 to 149 words.
Why is short text held to a stricter bar?
Short human writing is easier to misread, so texts of 100 to 149 words use a stricter threshold. Even so, about 2.4 in every 100 human texts of that length received a wrong AI signal in our sealed test, against about 2 in every 1,000 at 150 words or more. If you can, check a longer passage.
What is the score stability range?
The range reflects variation in the score mapping across resampled source families. It is not a probability that AI wrote the text. The verdict is decided by the frozen thresholds on the internal signal; the displayed number never changes it.
My own writing was flagged. What should I do?
Keep the evidence you already have: drafts, version history, notes and sources. Check a longer passage if one exists, because longer text is measured against a stricter-tested threshold. Then show the reviewer the measured false-flag rate for that length, which is printed on the result card and on our evidence page.
Does it work on languages other than English?
It is validated on English text of 100 words or more only. Other languages are accepted but nothing about those results has been checked, so treat them as unvalidated.
Which kinds of writing are hardest for it?
Short texts, technical writing such as software bug reports, and volunteer replies written in the style of an AI assistant are harder for the system; the figures for each type are on the Evidence page.
What happens to the text I paste?
Submitted text is not retained. Limited operational metadata may be logged for reliability and aggregate product monitoring.