Currently there is only the way to tell it what is right – by creating enough correctly identified observations.
There are lots of ongoing discussions about shortcomings and potential improvements to Computer Vision, as listed in this search of forum topics.
One in particular the developers are considering is how to make it smarter geographically.
I would suggest chiming in on some of the above topics if you are interested. I think this thread (correct me if I’m wrong) was intended to be more about human-generated IDs.