The most common languages in the training corpus will output the most reliably.