What happens when you paste an essay in
You paste the essay, or you drop in the Word file or the PDF, and a few seconds later you get a probability that a model wrote the text, the verdict band that probability falls in, and the rate at which that band has been wrong on our evaluation set. If the essay is under 150 words we say so and give no number, since a short piece of text does not carry enough evidence for a call worth making. The essay is scored in memory and not kept once the answer has gone back to you, and there is no account to make first.
Paste the whole essay and not one paragraph of it. The scan reads every part of the text and reports on the most machine-like stretch it finds, so it has more to go on when it has all of it, and one paragraph on its own is usually under the word floor anyway. Leave the paragraph breaks and the headings the way they were written. The model behind the scan was trained on essays with their formatting and on the same essays with the formatting stripped out, so we would rather you sent the essay as it is than a version you cleaned up for us.
What the score means on an essay
The number is a probability, and for every band of it we have published how often it turned out to be wrong on a frozen set of texts the model had never seen. On that set the false-positive rate came out at 2.6% on native-speaker writing (13 of 493), which is the share of human texts that got a verdict and were called likely AI all the same, and on writing by people whose first language is not English it was 0.6% (1 of 159). Going the other way, 16.5% of texts we called likely human were actually AI (126 of 764), so a likely-human result is evidence and not a certificate. And 5.8% of our evaluation set (88 of 1522) got no verdict at all, since the scan says inconclusive when the two sides are close and does not pick one.
An essay is the hard case for this kind of software. Careful grammar and a cleaned-up second draft are what a good essay is supposed to have, and they are also the features the models learned to associate with machine output. In 2023 a Stanford group ran TOEFL essays by non-native writers through seven popular tools and more than half came back as machine written [Liang et al., 2023], and a separate test of fourteen tools that year found none of them above 80% accuracy once the conditions changed a little [Weber-Wulff et al., 2023]. So we publish the rates with the counts behind them, and the scan says it cannot tell when it cannot.
If you are marking the essay
Use the result the way you would use a colleague's hunch, as a reason to look closer. A likely-AI result on one essay, from one tool, with a false-positive rate that is above zero, is not enough to fail a student on, and the students it lands on most often are the careful ones and the ones writing in a second language. What settles the question is the writing process, so ask for the drafts, the outline and the version history in Google Docs or Word before you say a word about the score. We wrote up what a fair process looks like for this situation, and it is short.
If it is your essay
Run it before you hand it in if that would put your mind at rest, and keep the result with the date on it. It will not make another tool agree with us, since the tools are not calibrated against each other, but it does show that you looked in good faith. What protects you if a question ever comes up is the record of the writing, and that record builds itself when you write in a document with version history switched on. The guide for students goes into this in more detail, and if you have already been flagged, start with the page for that.
Rewriting an essay to pass a scan
We do not sell a humanizer and we do not give advice on getting an essay past a scanner, ours or any other. If you wrote the essay, the drafts are your evidence and you do not need a rewriting tool. If a model wrote it, a tool that reshuffles the sentences turns work you did not do into work you also cannot explain when someone asks you about it. The scan on this site tells you where an essay stands, and that is all it is for.
Sources
- Liang, Yuksekgonul, Mao, Wu, Zou, “GPT detectors are biased against non-native English writers,” Patterns (2023), arxiv.org/abs/2304.02819 — seven detectors, TOEFL essays, 61.3% average false-positive rate.
- Weber-Wulff et al., “Testing of detection tools for AI-generated text,” International Journal for Educational Integrity (2023), arxiv.org/abs/2306.15666 — fourteen tools tested, none above 80% accuracy, and they disagree with each other on the same texts.
Common questions
Is the essay check free, and do I need an account?
It is free and there is no account. You paste the essay or upload the file and the result comes back on the same page. The only thing with a daily allowance is the certification badge, and a plain scan does not count against that.
How long does the essay have to be?
At least 150 words, and up to 10,000 words in one scan. Under the floor we give no number at all, since the evidence in a short text is too thin for a verdict we would stand behind, and most school essays are well over it in any case.
Can a teacher fail a student on this result?
Not on the result alone, and we say so on every result card. The score is a probability with a published false-positive rate behind it, which means some human essays will be called likely AI, and the students that happens to most are the careful ones and the ones writing in a second language. What a fair process looks like is on the teachers page.
Will Grammarly or a spell-check push my essay toward likely AI?
We have not measured a slice of Grammarly-polished essays on their own, so we will not put a number on it. The published research on other tools found that polish and formal structure raise the risk of a false positive, and that is a reason to keep your drafts, since a draft history shows the polishing happened on top of your own writing.
Do you keep the essays people paste in?
No. The essay is scored in memory and it is gone once the verdict has gone back to you, and we do not train on it or pass it to anyone else. The privacy page has the details and the tests that keep it true.