Skip to content

Our method

Almost every online prep course promises "an AI score". None of them shows how close that score is to a real examiner's. We want to measure it and publish the result. Here is where we stand, with nothing dressed up.

TL;DR

  • The corrector applies the three criteria of the TCF candidate handbook: linguistic, pragmatic and sociolinguistic.
  • An internal check has been done; a comparison with qualified human examiners is planned and will be published.
  • Until then, the level shown is an estimate for practice, not an official score.

What is already done

  • A first internal check: about twenty practice texts and recordings written by our team, from A2 to C1, in writing and speaking, run through the corrector several times. It helped us fix flaws, such as mistakes flagged by error.
  • This check has a limit: the level of these texts was set by us, not by a certified examiner. So it is not yet a measure of reliability.

What we plan to do

  • Gather practice texts and recordings at every level, with their authors' consent.
  • Have them marked by human examiners qualified in French as a foreign language, using the official TCF criteria, without seeing the AI's score.
  • Compare: the average gap in levels, the share of texts given the same level and within one level, and how much the examiners agree with each other.
  • Publish these figures here, with the date of the measurement and the version of the feedback being assessed.

Until then, the level shown by the corrector is an estimate to help you practise, not an official score.

The criteria we use

We apply the criteria published by France Éducation international in the TCF candidate handbook (Manuel du candidat): linguistic (vocabulary, grammar, spelling, sentence building), pragmatic (coherence, development of the topic) and sociolinguistic (fitting the situation). The rules that lead to "below A1" (A1 non atteint), such as word count, off-topic answers and memorized texts, are checked first.

Your data

A text only enters our checking corpus if its author agreed to it with a separate checkbox, unchecked by default. It is then anonymized.