My two kroner:
I was EVP/CMO of the largest language translation company in 2001 (BGS), and just exited as CEO of a small translation company (Elanex). What Holger says is largely correct - from the perspective of absolute accuracy of language. If a machine translates something, you can't be 100% sure it is 100% right. If a trained professional translator does, you are 99.999% there.
MT (machine translation) has always been a "next ten year" technology for the past forty years. However, last Fall Google MT made a step-change improvement in MT by using a
neural network model. It was most notable in Japanese where translations that were largely laughable became more passable. Not good enough to put on your website, but the gisting was pretty good. The major language pairs - DE/IT/ES/EN - are remarkable good. It will continue to get better as the previous SMBT (Statistical) approach seemed to plateau even with gigantic corpora (database of sentence pairs).
But Holger brings a good point about gunged up TM's and certainly there is a lot of garbage out there and it is growing. If I'm publishing customer facing content that relates to my brand, health and safety information, instructions, etc. it will for sure be done by a professional translator. If the language pair, TM is big, and MT results are good enough, I'd be happy to use something called PEMT (post edited machine translation) to speed the output and lower the cost. But again, I'll always want a human involved.
The hardest part about automating translation is teaching a machine to understand context. That's why rules-based MT systems failed early - there are not enough rules to understand the implied subject of a sentence, etc. For example, in English, the subject can be implicit and understood by the listener/reader. In Japanese, the subject is always explicit, and repeated over and over. Those sorts of structural differences, plus general nuance of spoken language (previous sentence hints) makes this a big leap.
Anyone that uses Siri or the android equivalent knows that voice recognition gets better and better. Once you have speech to text working well, then plug in an MT engine and the appropriate text to speech system and you have automated verbal translation. Would I want to use this for a medical conversation or a business transaction? Not today, not for a very, very long time. Would I want to use it to find the best tacos in the neighborhood? Yes.
Microsoft released
Skype Translator. That's pretty much how the earpiece thing works. You can give it a try with your friends overseas. Also, if you have not tried the Google Translate App with Wordlens - it translates text your smartphone camera is reading on the fly. Pretty amazing for menus and road signs when you travel and have the right data plan.
I'm quite happy to be out of the day-to-day slog of convincing companies to spend large sums on translating content when some executive that's never left the country says, "just use Google translate, it's free."
I also think that the perfection we once demanded for printed materials has really degraded over the last decade. Part of this is a 140, oops, 280 character way of expressing ourselves; people that think by putting "sent from my iPhone" is an excuse not to proofread; and also because of agile development and publishing moving to an online medium where the cost to fix is zero (unlike a printed manual). All of these forces plus technology is accelerating the adoption rate and acceptance of MT.
Perhaps the easier it is for us to communicate with people we don't understand the better off the world might be. That's a reason to hope this technology really does become mainstream.