Multilingual AI Detection – Which Tools Actually Work in French?
Wrote the French language AI tools thread for our school last year and I’ve been updating my notes since. Here’s what I know now about which tools actually work for French, not just claim to.
Genuinely tested on French academic text:
– Proofademic: multilingual model with actual French training data. performed best in my informal testing on French text specifically.
– Compilatio: French-origin product, better calibrated for formal academic French than English-origin tools.
English tools with French “support”:
– GPTZero: poor performance on academic French. Better on conversational French.
– Turnitin: marginally better than GPTZero but still trained primarily on English data.
The bilingual Canadian reality is that we need tools that perform well on both, and we don’t really have a good mainstream option yet. Compilatio is French-origin but not well-known outside Quebec. If anyone has tested it in detail I’d love to compare notes.
4 Replies
Join the discussion.
Log In to Reply
helpful breakdown. the distinction between "supports French" and "trained on French" is exactly the one teachers need to understand. GPTZero's French "support" is essentially applying an English model to French text, which doesn't work.
Compilatio est en effet meilleur pour le français académique québécois. Je l'ai utilisé pendant deux ans au cégep. L'inconvénient est le coût et le fait que le support anglophone est limité. Pour les écoles vraiment bilingues, les compromis restent difficiles.
Ce que je dirais de Proofademic sur le français: j'ai maintenant six semaines de données comparatives et c'est l'outil anglophone qui performe le mieux sur nos textes français. Pas parfait, mais significativement au-dessus des autres.
the multilingual problem compounds with bilingual students. a student who thinks in one language and writes in another produces patterns that detection models have almost certainly never seen in training data. the false positive risk for that population specifically is very high.
au Québec on a testé quatre outils en contexte francophone cette année. les résultats variaient tellement qu'on a finalement décidé de ne plus utiliser les scores comme élément central d'une décision - seulement comme signal initial qui doit être corroboré. GPTZero en particulier était très inconsistant sur les textes en français québécois académique.