How it works

Why AI text detectors do worse in Czech and Slovak than in English

Detectors learn mostly on English, so in Czech and Slovak they work with a smaller and less represented sample. Free word order and inflection further dilute the signals detection rests on. Treat a result in these languages as a weaker hint than the same result in English.

A smaller corpus, less certainty

The models detectors rest on have seen orders of magnitude more English text. Czech and Slovak are minority languages for them, and the decision boundary sits more loosely there.

In practice that means a wider grey zone — more texts end up somewhere between clearly human and clearly machine-made.

Free word order dilutes the signal

Detection leans heavily on how predictable the next word is. English has a fixed word order, so predictability is easy to measure. In Czech and Slovak you can assemble the same sentence five ways without changing its meaning.

The model then sees a smaller difference between human and machine writing than actually exists.

What to do about it in practice

Don't work with a single number. Look at the marked sentences and read them out loud — you will recognise a clunky rhythm and template-like turns of phrase yourself.

For longer work, check chapter by chapter. An average across a hundred pages hides exactly the place you care about.

FAQ

Is detection in Czech and Slovak any use?

Yes, as a basis for a conversation. It is not proof, and in these languages that holds even more than in English.

Does translating the text into English help?

It does not. Translation changes the construction of the sentences, and the result then says something about the translation, not about the original text.

Does DetekceGPT account for Czech and Slovak?

Its own style layer accounts for things specific to both languages, for instance the share of diacritics and the frequency of native function words. The tool works out which of the two a text is on its own.

Try it on your own text

Paste a text or upload a file and look at the score and at the specific sentences that came out suspicious.

Run a detection

More articles