As someone who does ML for a living I can definitely say that false positive data from the field can significantly improve the experience for those 1000. That is what should be standardized if we wanted to see systems that work for everyone.