Tracking language development when students use AI translation first

Hit a wall this term trying to genuinely assess my French immersion students’ actual language growth. More and more of them are drafting their written work in English first, running it through Google Translate or a similar tool, then cleaning up the French output themselves before submitting. The final product often reads like real, if imperfect, French, so on paper it looks like progress. But when I give the same students an in class, no tools, timed writing sample, the gap between that and their translated homework is enormous, sometimes shockingly so. I genuinely dont know anymore what their true proficiency level is because the two data points tell completely different stories, and report cards are coming up. Is anyone weighting in class, tools free writing much more heavily than take home work for exactly this reason? Feels like the only honest signal left but it also feels unfair to base a grade mostly on a nervous timed sample instead of the fuller process work.