It could be that the people's typing (and therefore the training data) includes so many different unique movements that an "average" of all of them doesn't actually look like any real person's individual typing pattern. Sorta how some one-size-fits-all designs don't fit any one person perfectly.