AI Detector for Teachers: Reading a Result Fairly

What a detector result can and cannot support

A detector compares a text with patterns it learned from human and machine writing and returns a model signal. Textumi names the result with one of five labels: “Human-written”, “Mostly human”, “Mixed: human and AI”, “Mostly AI” and “AI-generated”.

Each label is an estimate from the style of the text, not a finding about who wrote it. “Human-written” does not confirm that a person wrote the text, and “Mixed” or “Mostly AI” does not show what share of the text a machine wrote: the detector does not measure that. No label is the probability that the student used AI.

So “Human-written” clears nobody, and “AI-generated” convicts nobody.

A result can help you decide which submissions deserve a slower, closer read than the rest.
A result can point you to passages worth asking about when you talk with the student.
A result cannot establish who wrote a text, and it should never be the only basis for a grade penalty or a misconduct report.
A result cannot settle a disagreement, because a student has no way to argue with a model signal on its merits.

Detectors go wrong in both directions: they flag human writing and miss AI writing. In 2023 OpenAI withdrew its own AI text classifier, citing low accuracy. Check a result from any tool against other evidence.

A fair review starts from how the student wrote

The strongest evidence of authorship is how a text came to exist, and a detector sees none of it. When a submission worries you, work through these steps before you form a view.

Read the work yourself first and note what seems off: a change of voice, sources that do not exist, claims the course never covered.
Compare it with the same student's earlier work, ideally something written in class: vocabulary, sentence length, typical mistakes.
Ask for drafts, notes and an outline. A document written in an online editor usually keeps a version history that shows the text growing over hours or days.
Talk with the student in private and without accusation. Ask them to explain the argument, describe how they found a source, or say why they built a paragraph the way they did.
Write down what you looked at and what you found, including anything in the student's favor.
Follow your institution's procedure and decide on all the evidence together, not on one detector result.

A student who can show drafts and discuss the work in detail has answered the question the detector raised. If they cannot, a different conversation may follow, but the conclusion then rests on that conversation and the drafts, not on the detector’s label.

Short answers and non-native writers need extra care

A short text gives a detector very little to work with. Textumi accepts texts from 5 words, but under 50 words it shows “Too little text to judge” instead of a label, and even a slightly longer answer gives a less reliable result than a full essay. Textumi only assesses paragraphs of 50 words or more on their own, so a one-paragraph answer or a discussion post is too thin to judge by a detector.

Students who write in a second language are flagged wrongly more often. In 2023 Stanford researchers found that detectors labeled essays by non-native English writers as AI-generated more often than essays by native speakers. A smaller vocabulary and regular sentences look predictable to a model, and predictability is what a detector reads as machine writing.

Read results with the same caution for formulaic genres such as lab reports and summaries, and for texts that were translated, heavily edited or written partly with someone's help.

Language matters too. Textumi's English and Ukrainian detectors are the most thoroughly tested. The other eighteen languages were trained and tested on a narrower set of texts. French, Vietnamese, Turkish, Indonesian and Dutch often miss AI texts in other genres, so a “Human-written” or “Mostly human” result in those languages says little.

A short course policy on AI and detection

Students cope better when they know the rules before the first assignment. One page, in line with your institution's regulations, can cover these points.

Which uses of AI are allowed in the course, which are not, and whether the answer changes from one assignment to another.
How students should disclose permitted AI help, for example in a short note at the end of the work.
A request to keep drafts, notes and version history until final grades are in, with the reason you ask.
That you may use a detector, that its result is a signal and not proof, and that no penalty will rest on it alone.
What happens after a flag: a private conversation, a look at the drafts, and the formal procedure only if concerns remain.
How a student can raise a concern or appeal if they believe they were flagged wrongly.

How to check a text in Textumi

Sign in to your Textumi account, open the workspace and choose AI Detector.
Paste the text, or import a DOCX, PDF, ODT, RTF, TXT or MD file of up to 10 MB. Text in a screenshot or photo in PNG or JPEG can be recognized, but review it before you check it.
Leave the language on automatic detection, or choose it yourself if the text is short or mixed.
Select Analyze text. One check takes up to 12,000 characters, so check a long paper in parts.
Look at the label beside the original and at the highlighted paragraphs. In a long text the highlights do not cover every paragraph.

The detector is Textumi's own development: we built and tuned its models, and checks run on our own servers without passing the text to an outside AI provider. You can use it in the web workspace, in the macOS app on Individual and Professional, or through the Textumi connector right in a conversation with your AI assistant. We keep improving the detector and regularly add tools that make everyday work with it easier.

How many words you can check depends on your plan; the pricing page lists every allowance. On Individual and Professional, results stay in a private History for 30 days, and you can delete them sooner. On plans without History you can open a result for about an hour after it finishes; after that it is no longer available to you, and it is deleted within 7 days.

Open the AI Detector Compare the plans Read the privacy policy

Common questions

Can teachers detect ChatGPT?

Sometimes. A detector may flag text from ChatGPT or another model, but it often misses it, especially after editing, and it sometimes flags work a student wrote entirely alone. No tool can identify where a text came from with certainty. A result tells you where to look; a conclusion needs drafts, earlier work and a conversation with the student.

Is a detector result evidence?

Weak, indirect evidence at best. A result shows only how closely a text resembles patterns the model has learned. It is a signal, not proof of authorship, and it cannot be questioned like a witness. A fair decision needs evidence about the writing process: drafts, versions, sources and the student's ability to explain the work.

How long does a text need to be for a reliable check?

The longer, the better. Textumi accepts texts from 5 words, but under 50 words it shows “Too little text to judge” instead of a label, and a result on a text just above that still rests on little material. Check full essays or long passages. One check takes up to 12,000 characters, so a long paper may have to be checked in parts.

Why would a student's own writing be flagged?

Detectors react to predictable wording and sentence structure. Careful formal or formulaic writing, and writing in a second language, often look that way. Short, translated and heavily edited texts are also misjudged more often. A flag on honestly written work is a known failure of these tools, not a sign that the student is hiding something.

Should I run every submission through a detector?

That is for you and your institution to decide, but first think about what you will do with the results. Checking everything produces false flags through volume alone, and each one means an uncomfortable conversation for a student. Many teachers check only when their own reading raises a doubt, and say so in the course policy.

Which languages can I check in Textumi?

The AI Detector works with 20 languages. The English and Ukrainian detectors are the most thoroughly tested. The other eighteen were trained and tested on a narrower set of texts, so read their results on other kinds of writing with more care. French, Vietnamese, Turkish, Indonesian and Dutch often miss AI texts outside the genres they were tested on.

What happens to the student work I check?

Submitted texts and results are stored encrypted. On plans without History you can open a result for about an hour after it finishes; after that it is no longer available to you, and it is deleted within 7 days. Until then it is kept encrypted, and only Textumi's administrators can open it, for support and abuse prevention. On Individual and Professional it stays in your private History for 30 days or until you delete it. Detection runs on Textumi's own servers. Check your institution's rules before uploading work that identifies a student.

What should I do when a student denies using AI?

Take the denial seriously, because the detector can be wrong. Ask for drafts, notes and version history, and have the student walk you through the argument and the sources. If the evidence about the writing process supports them, the question the detector raised is settled. If concerns remain, follow your institution's procedure and pass on everything you looked at, not just the result.