The French-English AI Detection Gap Nobody Is Talking About
I coach in a bilingual school in Ottawa serving both francophone and anglophone students. the gap in AI tool quality between French and English is substantial and it is creating real equity problems that nobody in the mainstream AI-in-education conversation is addressing.
English-language AI tools: refined, multiple models, high detection accuracy on English AI text.
French-language equivalent: years behind. Most tools are English models with French “support” bolted on.
In practice this means: anglophone students can be evaluated with tools that have some accuracy claims. francophone students get evaluated with tools that have been barely tested on French academic text. the false positive rate is higher for francophone students. the detection accuracy is lower for actual AI-generated French text.
process over detection. always. but especially here – using detection-based approaches on French student writing with current tools is genuinely inequitable.
5 Replies
Join the discussion.
Log In to Reply
the bilingual dimension is critical and so underappreciated. our school in Ottawa has both francophone and anglophone students in the same building sometimes using the same detection policies applied through tools designed for one language. this is not acceptable and nobody at the board level seems to be thinking about it.
After 20 years in Quebec CEGEPs, I can give you the numbers. Over the past term I've been documenting false positive rates on francophone student writing and they are running approximately 40% higher than what the same tools produce on anglophone text from the same assignment.
GPTZero in particular performs very poorly on formal academic French. The training data bias is structural. Until there is a tool trained on substantial French-language academic text, detection on French writing will remain inequitable by design.
I've been sharing this data with my colleagues and we've moved to a strict process-documentation approach. Detection tools are not used for any formal purpose in my department.
en tant qu'enseignante de FLS depuis douze ans, ce thread me parle vraiment. j'ai eu des cas où des élèves forts - des gens qui écrivent vraiment bien en français - ont été flaggés. pas parce qu'ils ont utilisé l'IA mais parce que leur français est trop formellement correct. le modèle n'a pas appris à distinguer ça.
the fix is retraining on francophone academic data. nobody is going to do that. too expensive, too small a market.
Chloe has identified something that affects every institution with a dual-language mandate. The problem is structural: most detection tools were validated primarily on English-language corpora, and "supporting French" in documentation often means the interface translates, not that the underlying model performs equivalently. Anyone making consequential decisions based on French detection scores should verify independently what validation data was used.