Sadly, all benefit systems that pay out on certain conditions are subject to fraud. For me as long as there is a human involved in the decision and the benefits

Sadly, all benefit systems that pay out on certain conditions are subject to fraud. For me as long as there is a human involved in the decision and the benefits recipient has the right to appeal or sue there is no problem in using ML models for "finding most likely candidates for fraud". I don't know about France, but I assume it is similar to my country Poland where one can sue the state very easily with minimal expense. My parents did when I was a child and we were quite poor (self represented and all that).