Skip to main content
Sign inSign up
Method

How we test our AI detector, and what the numbers mean

Every figure we publish comes from texts the detector never saw in training, and each one carries its sample size and its test date.

The test set

The texts we test on

An AI detector false positive is a human text that the detector calls machine-written. It is the error that hurts people most, so we measure it first and on its own. Our current figures come from a held-out test run on 12 October 2026 against the version of the detector the site serves today. Held out means that no text in the test was used to train or tune the detector, and before each reading we check the test set for overlap with the training data. In this test the overlap check found 0 matches.

How a test run works

  1. Keep the test texts apart

    No text in the test set is used to train or tune the detector. That is what held out means.

  2. Check for overlap

    Before each reading we compare the test set with the training data. In the current test the check found no match.

    0 matches

  3. Read at the result card's line

    Every text gets a score. It counts as flagged at the same line where the result card says a text reads as AI-written.

    65

  4. Publish counts with a date

    Each figure carries its sample size and its test date. A new test changes the date on the accuracy page.

    2026-10-12

The test set has three groups.

  • 01

    Native human writing

    150texts per language

    Everyday web text, 150 texts per language, at most 2 from any one website, spread over 8 kinds of writing.

    Languages
    • English
    • Spanish
    • French
    • German
    • Italian
    • Portuguese
    • Dutch
    • Polish
    • Turkish
    • Russian
    • Arabic
    • Hindi
    • Indonesian
  • 02

    Non-native human writing

    5languages with a learner set

    Texts by people writing in a language they learned, taken from learner collections in English, German, Spanish, French and Italian.

    Texts per language
    • English600
    • Spanish300
    • French300
    • German300
    • Italian300
  • 03

    Machine writing

    9current chat models

    Texts from 9 current chat models, plus continuations of a human opening and paraphrases.

    Texts per language
    • English75
    • Spanish76
    • French77
    • German75
    • Italian75
    • Portuguese75
    • Dutch73
    • Polish74
    • Turkish73
    • Russian75
    • Arabic73
    • Hindi73
    • Indonesian71
The line

Where a text counts as flagged

  1. Human sideThe result card readsReads as human-written
  2. In betweenThe result card readsNeither clearly AI nor clearly human
  3. Machine sideThe result card readsReads as AI-written

A text counts as flagged when it scores 65 or more. That is the same line at which the result card says a text reads as AI-written, so the test measures the verdict you would see.

False flags

How often the detector flags human writing

LanguageNative human texts flaggedNon-native human texts flagged
English0 of 150 (0.0%, 95% range 0.0 to 2.5%)3 of 600 (0.5%, 95% range 0.2 to 1.5%)
German0 of 150 (0.0%, 95% range 0.0 to 2.5%)0 of 300 (0.0%, 95% range 0.0 to 1.3%)
Spanish1 of 150 (0.7%, 95% range 0.1 to 3.7%)2 of 300 (0.7%, 95% range 0.2 to 2.4%)
Turkish0 of 150 (0.0%, 95% range 0.0 to 2.5%)Not measured

These are the false flags in the four languages of this site. Each cell gives the count and the rate, with the 95% range around it.

The English non-native set has two halves. Writers whose first language is Arabic were flagged in 3 of 300 texts (1.0%, 95% range 0.3 to 2.9%), writers whose first language is Spanish in 0 of 300. We have no non-native Turkish set, so we make no claim about Turkish written by learners. The figures for the other nine calibrated languages are on the accuracy page.

Machine text

How many machine texts it catches

LanguageAll machine texts caughtTexts written whole by a chat model
English68 of 7554 of 54
German68 of 7554 of 54
Spanish63 of 7652 of 54
Turkish65 of 7352 of 54

The machine samples are smaller, between 52 and 77 texts per language, so we publish them only as counts.

"All machine texts" includes the continuations and paraphrases, which are harder to catch than a text a chat model wrote from start to finish. Do not add a caught count and a flagged count together. They come from different texts and answer different questions.

Known limits

Where the detector is weaker

French is our weakest language at catching machine text: 53 of 77 machine texts, and 44 of 54 written whole by a chat model. The accuracy page shows this row next to the others.

Texts where a model continues a human opening are caught far less often than the rest. The groups are too small to give a fair rate, between 9 and 15 texts per language, so we give none.

Short texts give the detector less to read. Below 50 words there is no reading at all, and under 100 words the confidence rating always stays Low. Long documents have a weakness too: the score of a long text can move when the text is split into parts in a different way. We have not fixed that yet, so for a long document, look at the marked sentences as well as the overall score.

Publication rule

The rule a language has to pass

A language gets its own line only if its test set reaches a separability score of at least 0.9, measured as AUROC. AUROC is the chance that a random machine text scores higher than a random human text, so 0.5 is a coin toss and 1.0 is a perfect split. After the line is set, no more than 1 human text in 20, and no more than 1 non-native text in 20, may land at or above it. Thirteen languages pass this rule today. A text in any other language gets no score, only a note that the language is not calibrated.

Other tools

The humanizer and the other tools

For the AI humanizer we publish one result. In a test of 20 English texts on 17 October 2026, the rewrite kept the facts in 20 of 20 texts and kept the register in 20 of 20. That is a small test, and it says nothing about AI detectors. We do not publish a pass rate for the humanizer, and we do not claim that its output passes any detector.

We also run our own tests on the grammar checker, the paraphraser and the plagiarism checker. Their samples are small or built by us, so we keep those results internal until they are large enough to publish.

Our claims

What we do not claim

  • That the detector proves who wrote a text. It gives a probability with a range.

  • That it works in every language. It is calibrated for 13.

  • Any comparison with other detectors. We have not measured one.

  • Any accuracy for texts shorter than 100 words.

When we run a new test, the date on the accuracy page changes. These figures carry no personal byline. They are the company's measurements, and the editorial policy explains how our articles quote them. The team page shows who writes and checks the rest of the site.

Every language on the accuracy page Back to the detector