Language is an incredibly difficult problem. Natural language processing can get you to 80% very easily, but we've been struggling with that last 20% for a long time.

As the saying goes, once you're done with the first 80%, you just need to complete the last 80%.

The reason is that intelligence is needed for the remaining 20%, and that's hard to emulate.

I think more likely there will be "tiers" of translators, with highest accuracy going first to languages with accessible and common Rosetta Stone-style word-for-word translations. Think French to German. Or Polish to Russian. Or English to everything else.

Langagues without vast translated data sets will probably be stitched together by various translations of multiple degrees. If you want to translate Farsi to Creole, for example, you probably won't find many direct translations to start with, but you'll probably have a better chance getting "mostly" there by going through a "hub" language like French or English.