Though I generally agree with your point, the tree vs stopsign example may not be the best because it would arguably work equally well on humans.
The point is that humans would see one thing whereas computers would be highly confident it is something else.