How accurate it really is.
Measured on our own test set and published here in full, every time we change the model. No AI detector is close to perfect, this one included. These are the numbers to know before you trust a result.
When a person wrote it
5 in 100 human-written texts score above 50%.
Each amber dot is a real person we would read wrongly. They are the reason a number from us is never proof, and the number we work hardest to bring down.
When AI wrote it
37 in 100 AI-written texts score below 50%.
Each blue dot is AI text we read as human. Short texts, heavily edited drafts and the newest models slip past most often. We publish this number too, because a detector that only reports its hits is hiding half the story.
What moves the number
Length
Short texts give the model little to go on. Under about 150 words we show a wide bracket and say so. Every extra paragraph narrows it. Longer text, steadier reading.
Genre
Formulaic writing scores higher whoever wrote it: lab reports, cover letters, legal boilerplate, abstracts. A tidy five-paragraph essay sits closer to that than a diary does.
Non-native English
Simpler vocabulary and regular sentences can read as predictable. We test this group on its own and publish the gap rather than pretend it is zero.
Polished, not written
A grammar pass by an AI tool moves human text up a little; a full rewrite moves it a lot. We can show which paragraphs changed character. We cannot tell you which of the two happened.
How we test
The way we would want to be tested: on texts the model has never seen, with the misses counted and published.
- Human texts Essays, emails, cover letters, posts and articles written before 2022, when AI writing tools were rare, from many writers, including people writing in English as a second language.
- AI texts Generated from the same prompts by several current models, in several styles, with and without an instruction to sound human.
- Refreshed on a schedule The test set is refreshed after every model or threshold change, and whenever a major new model ships. This page shows the result before we change anything.
- 50% means 50% We do not move the line to make the numbers look better. A text at 50% is as likely one way as the other.
What this means for you
If you are a student
A high number on your own writing happens to about 5 in 100 people. Keep drafts, notes and version history. They are stronger evidence than we are, in either direction.
If you teach or grade
Treat a result as a reason to start a conversation, never as the end of one. Ask for drafts. A score alone is not evidence of anything, and it should never decide a grade.
If you edit or hire
Use the paragraph view to see where a text changes character, then read those paragraphs yourself. The signals we list are things you can check with your own eyes.
If a decision rests on this, do not let it rest on this alone.
Check a text