Because cheating is easier than actually doing work, if you use this to train future models, it's likely you'll end up with cheating instead of actual generalization.