Students switching between French and English prompts to dodge detection, seeing this?
noticed a pattern this week that i havent seen named anywhere else yet. a few students in my french second language course submitted essays that read like they were drafted in english by chatgpt then translated to french, either by the student or by the same tool. the french comes out grammatically correct but stylistically flat in a way thats hard to describe, missing the small idiomatic touches a genuine french draft usually has even from a weaker writer. detection tools we have access to didnt flag any of these as ai generated at all, i think because they’re mostly trained to catch patterns in english generation, not translated french. this feels like a gap thats going to get exploited more as word gets around. anyone else seeing this specific pattern of draft in english translate to french as a detection workaround.
5 Replies
Join the discussion.
Log In to Reply
there was some discussion of this exact translation workaround at a conference i attended in the spring, presenters called it cross language laundering informally. nobody had a clean detection solution for it yet as of a few months ago.
@Greg Simard cross language laundering is a good name for it, wish i had that term when i was writing up my notes for our department this week.
YES i think i caught something similar without realizing what it was, a students french essay that felt technically fine but completely lifeless. going back to compare it against their earlier in class writing samples now.
the reliable fix here isnt a better detector, its knowing your students baseline french writing well enough that the flatness stands out to you directly, which is exactly what nathan and angela are describing. keep dated in class french writing samples from early in the term for every student, ideally handwritten or in a locked environment, so you have a genuine reference point when something later feels off. no tool is going to catch cross language laundering reliably any time soon.
process over detection, always, but this is a good example of why. even if a tool eventually catches translated ai french, a student who actually engaged with drafting in french from the start writes with a different rhythm you can hear immediately once you know your students.