Methodology
What we measure, what the number means, and where it fails. The short version: this is evidence, not proof, and it should never be the only thing you act on.
What the score is
Your text goes to a third-party AI-detection model, which returns a probability that the writing was machine-generated, overall and sentence by sentence. We show that probability, hedged, as "Our confidence this is AI-generated: N%". Nothing is rounded up, and the number you see is the number we got.
What the score is not
It is not a record of what wrote your text. No such record exists. Detectors work by measuring how the writing is put together: predictability, sentence-shape regularity, word-choice patterns, and comparing that against how machine-generated text tends to look. A high score means "this is written the way AI tends to write." It does not, and cannot, mean "a machine wrote this."
That gap is why we never print a verdict. "This is AI" is a claim about the world; "our confidence is 92%" is a claim about our measurement. Only the second one is ours to make.
Where it fails
- False positives are real. Genuine human writing scores high sometimes. Independent testing consistently finds real-world accuracy well below what detector marketing claims: ours included, which is why we don't print an accuracy figure to be impressed by.
- The errors are not evenly spread. Writing by non-native English speakers is flagged disproportionately, because the same regularity and restricted vocabulary that a detector reads as "machine-like" is also what careful second-language writing looks like. Formulaic prose: lab reports, legal boilerplate, five-paragraph essays: trips it for the same reason.
- Editing defeats it. AI text that a human has rewritten, or that has been run through a paraphrasing tool, often reads as human to every detector on the market. A low score is weak evidence of anything.
- Short text is noise. The less text there is, the less signal there is. A couple of sentences is not enough to say much.
- The target moves. Detectors are trained on how models wrote yesterday. Newer models write differently, and accuracy against them is worse than against the ones the detector has seen.
How to use it
Treat a high score as a reason to look closer: read the work, ask about the process, compare it to other things the person has written. Treat a low score as almost no information at all.
Do not use this score by itself to discipline a student, reject a candidate, deny a claim, or end a contract. It is not built for that and no detector is. If you're making a decision that affects someone's education, job, or money, the score is one input among several, and the weakest one.
If you have been accused because of a detector score, including ours, the points above are the honest state of the technology, and you're welcome to cite this page.
What we don't claim
- We don't claim to be the most accurate detector, or to publish an accuracy percentage. Anyone quoting you one number is quoting you a number from conditions that aren't yours.
- We don't claim to identify which AI wrote something. That is not a thing detectors can do, whatever the marketing says.
- We don't claim to catch every AI text, and we say plainly that edited text usually gets through.
Languages
English only, for now. Not because the underlying model is necessarily blind to other languages, but because we have not verified how it performs on them, and a detector that quietly accepts your Spanish and hands back a number it cannot stand behind is worse than one that says no. Detection accuracy varies by language, and the false-positive problem above is already worse for writing by non-native English speakers. We will add a language when we can show what it does, not when the box accepts the characters.
Your text
Submitted text is sent to our detection provider to produce the score, stored only long enough to serve your report: deleted automatically within 24 hours, or 30 days if you bought it, and never used for training or sold. Details in Privacy.