I agree, but since you don't know which part of the data is for validation and which one is for training, and since you can assume such a system will only accept data that makes consensus, you tend to only accept data that is valid. That's the beauty of it.
