For many languages, the translation matrix doesn't work well directly, so you end up with Japanese -> English -> Spanish when you select Japanese -> Spanish. You lose details at every step, but the end result is still better than the rather unreliable direct translation.
Translations are also a lot better when the languages share more grammar and cultural history. Japanese has grammatical constructs that simply do not exist in English (and vice versa), so providing accurate mapping is an impossible task.
English <-> (insert Germanic or Romance language here) is always pretty solid, but should that be a surprise? This is also the case that is easiest for people.
I wonder if anyone has played around with Language A <-> Common Intermediate Representation <-> Language B. If not, I nominate Lojban for the CIR.
Translations are also a lot better when the languages share more grammar and cultural history. Japanese has grammatical constructs that simply do not exist in English (and vice versa), so providing accurate mapping is an impossible task.