The model can generate translations and humans can verify that they're accurate enough to be used for training, making corrections if needed. Verifying work is much faster than creating it.
b) The claim that verifying is faster than creating is unsupported. (E.g. reviewing programmer code is tougher than writing from scratch.)