But using the outputs to train would make their statement false, since they are influenced by the inputs