Yeah it gets tricky. I think eventually it has to be about tradeoffs - no ML system can be 100% correct. I do think there's a "right" decision (up to a point) in the context of the product or business.