AI detector and Humanizer languages

Supported languages

In each of these languages you can check a text with the AI detector and rewrite a draft with the Humanizer. The table shows the state of the detector and the longest text the Humanizer takes.

Textumi does not work with other languages: for a text in any of them, the detector returns an error message instead of a result.

How reliable the detector is in each language

Each language has its own classifier, and a text is never scored by the classifier of another language. Whatever the language, the result is a signal about how a text is written, not proof of who wrote it.

English and Ukrainian

These two detectors have been in service longest and are the most thoroughly tested. They still make mistakes in both directions, most often on short, formulaic, translated or heavily edited texts.

The other eighteen languages

Each of these detectors was trained and tested on Wikimedia prose (news, encyclopedia articles and discussions) paired with texts on the same subjects written by two AI models. A version is offered only if it meets a fixed rule on its test, described on the methodology page.

Dutch is the exception: its version fell just short of that rule, and Textumi decided to offer it anyway.

On other kinds of writing these detectors differ a lot. On a small set of AI-written notices, complaints, instructions and similar texts, the French, Vietnamese, Turkish and Indonesian detectors missed most of them, and the Swedish and Dutch ones about half. Romanian, Korean, Hindi and Spanish found most. Portuguese, Arabic, Polish, German, Italian, Norwegian Bokmål and Danish have not been checked outside their test material.

So in any of these eighteen languages, read “Human-written” or “Mostly human” as “little that looks like AI was found”, not as proof that a person wrote the text.

How Textumi's AI detector is built and testedOpen the AI detector

What the Humanizer does in each language

The Humanizer rewrites a text in the language it was written in, so that it reads more naturally and keeps its meaning. In every language it checks facts, names, numbers and exact names against the original, along with the logic: negations, contrasts and words such as "unfortunately". A version that fails the check is not shown.

The Humanizer was built for English first. A text in another language goes through an English intermediate version and back, so its limit is lower and each word counts as two. Compare the result with your original before you use it.

Through your own AI assistant, the Humanizer works in English only. The other languages are available on textumi.com.

Textumi's own rewriting pipeline does most of the work, and an external AI model named in the privacy policy also takes part. In no language does the Humanizer promise a particular result in detectors: another detector may still flag the rewritten text.

Open the Humanizer

Word limits for each language

In the Humanizer, each word of a text that is not in English counts as two: a Spanish text of 400 words uses 800 words of your allowance. The detector counts every word once in every language.

A text of fewer than 50 words gets no label in any language: Textumi shows “Too little text to judge” instead. From 50 words the estimate is more reliable.

Allowances in each plan

How the language of a text is set

Textumi recognizes the language from the text itself. You can also choose it by hand, which is safer for short texts and closely related languages: a short Indonesian instruction, for example, can be mistaken for Malay.

The detector does not guess. If it cannot tell the language with confidence, if the language is not supported, or if the text is clearly in a different language from the one you chose, it returns an error message instead of a result.

The interface language is a separate setting. It changes menus and messages but never decides which classifier reads your text, so you can check a Korean text in the Polish interface.