Every AI detector gets some texts wrong. It misses some AI writing, and it flags some writing a person did themselves. The second mistake is the one that hurts, because it lands on someone's name, and how often a detector makes it is called its false-positive rate.
How we set Detect's lines
We took human writing the detector had never seen, cut it into passages of a few hundred words and scored every passage. Then we drew three lines, where 1 in 10, 1 in 20 and 1 in 100 of those human passages score higher. When Detect flags your text, the result says which line it crossed.
How to read a result
| Result | What it means | What it doesn't mean |
|---|---|---|
| No AI signals found | No line crossed | That no AI was used |
| Slight AI signals, 1 in 10 | Scores higher than 9 in 10 human passages | That the text was generated |
| Strong AI signals, 1 in 100 | Scores higher than 99 in 100 human passages | That anyone cheated |
| Too short | Under about 100 words | Anything about the text |
"1 in 100" isn't a 99% chance your text is AI. It's how often human passages score that high.
Why it depends on what you write
The same lines apply to every kind of writing, but some kinds cross them more often. Formal, encyclopedic and technical prose gets flagged the most. In scientific papers, about 3 human passages in 100 cross the strictest line instead of 1. If you write papers, expect that line to catch your own writing about three times as often as it says.
Why short texts get no score
Under about 100 words there isn't enough to judge, so short texts get no score at all. Lots of social posts are that short. Any hidden characters in them are still listed.