How it works

What a check does

The standard detector reads the text. When it reads the text as human, the result is "Verified human". When part of the text reads as machine-written, the result is "Not verified", with the sentences and the reason.

Verified human: the standard detector read this text as human. It covers this text on this date. It does not show who wrote it or how.

The mark is for words a person wrote, as handed in. Text rewritten by a humanizer does not qualify, even if a person wrote the first draft. Our detector does not catch every rewrite: in our tests it missed most rewrites by one tool.

What it reads

Our detectors are language models trained on academic papers and student essays. Some were written by people and some by AI models.

They are not built or tested for other languages or other kinds of writing, such as fiction, email or business documents.

How a reading is made

The standard detector reads the text sentence by sentence and in longer passages, and marks the passages it reads as AI-written.

From those marks it works out one share: how much of the text it reads as written or edited by an AI model. At or under 6%, the reading is Human.

Treat the marks as a reading, not a finding. We have not measured how exactly they fall on the sentences an AI model wrote.

The second reading

A second detector, trained on student essays, also read the text. Its reading is shown as a note and does not decide the mark.

Why it is a note: before launch, a test on real student essays found one that the student-essay detector read as partly machine-written at every line we could set. Withholding the mark on its reading would have withheld it from that student, so the rule was changed before release.

Deep checks

On a free account and every plan, a deep check adds notes a person can check: references we could not find in Crossref or OpenAlex, unusual stock phrases, and pairs of numbers, names or dates worth a second look. The pricing page says whether up to 8 dates and numbers are also checked against Wikipedia today.

It shows the sentences it found and never marks a text as AI-written. Today its findings are notes and do not change the mark. They mark honest writing too: the model card lists how often, and what a reviewer made of the marks.

A lookup service can be unavailable. When it is, a reference reads as not checked, not as not found.

A deep check sends up to 8 sentences from the text to Google (Gemini, in our own cloud project) to be judged, and search words drawn from them to the Wikimedia Foundation (Wikipedia search). The titles in its reference list go to Crossref and OpenAlex.

What a flag means

A flag is any result other than Human.

A flag should be reason to look closer and not verified proof of cheating.

Settings

A verification reads at the standard setting (6%). In the app, a single-model reading under Advanced also offers accusation-safe (15%) and sensitive (2%); what each costs is on the model card.

Changing the setting never turns an AI verdict into Human. It only moves the line for text that is partly machine-written.

When it does not answer

Under 50 words there is no reading, and the result is "Not scored". From 50 to 149 words the reading is weak. Machine-written text under 300 words has not been measured.

If a detector is not answering, the check says so. Nothing else reads the text in its place.

Limits

• Paid humanizer tools. It called 20 of 176 humanized AI papers human and, in a newer test of one humanizer tool, 231 of 255 rewritten long papers and 122 of 136 rewritten essays human, so a humanizer tool can get text past it.
• AI-edited papers. At the standard setting the standard detector called 38 of 140 partly AI-edited papers clean. Some had only a small AI-written share.
• Recent writing. Every human-written document we tested was written before 2022, so we know it is human. Writing from after AI tools became common has not been tested, and neither has writing done with grammar checkers or translation tools.

How the live detector was chosen

The standard detector was chosen by a rule written before the results, and the second reading became a note by a rule written before its test.

Its default line moved from 10% to 6% on 2026-09-15.

We run and grade our own tests. No outside party has checked these results yet.

The model card lists each result with its date and its source.