They're picking among most likely tokens, which are ones with least entropy and so least information per token.