MT is an idiot savant in the sense that it can do very complex, large-scale tasks in a flash, but has no sense of context, the development of life, or the outside world. It’s a powerful computing brain, but it can’t think, properly speaking – it can only fulfill a range of programmed tasks.
Part of thinking is feeling, and MT can't feel. It lacks the lifetime of impressions, sentiments, and experiences that give us our knowledge of the world and our skill in managing it. To go all Humean, it lacks the impressions that it can transform into ideas. And it lacks style. It can't pick up on the social cues and hints of the many environments we traverse in our lifetimes.
Let's look at an example. Take the very first sentence of Gustave Flaubert's 1857 classic Madame Bovary. Here it is in the original French: Nous étions à l'Étude, quand le Proviseur entra, suivi d'un nouveau habillé en bourgeois et d'un garçon de classe qui portait un grand pupitre.
Here's the opening of the novel in Francis Steegmuller’s English translation: “We were in study-hall when the headmaster entered, followed by a new boy not yet in school uniform and by the handyman carrying a large desk.” Running the original through Google, we wind up with the following: “We were under study, when the headmaster entered, followed by a new dressed in bourgeois and a class boy who wore a large desk.”
This is a different picture entirely. First, the handyman has been transformed into a mere “class boy,” losing his significance in this vignette. More comically, the boy, instead of carrying the desk, is wearing it. This is due to the fact that the same verb is used in French for both actions. In “a new dressed in bourgeois,” the subject disappears: a new boy? a new girl? a new donkey? The French adjective nouveau contains a clue: it’s masculine, so we know the new entity is not a girl. But English lacks gendered adjectives, so you can’t tell the gender of the noun from the adjective.
To make things worse, it is clear that the computer lacks understanding of the cultural context of this passage. It doesn't know that “study hall” is the correct translation, rather than “under study,” a phrase that suggests they are being observed or monitored. And it doesn't know how to interpret the word “bourgeois,” which in this case refers not to a social class or even to a person, but to a style of dress (that's why he's “not yet in school uniform” in the standard translation).
So the picture we have is approximately as follows. A boy pathetically staggers around a research clinic with a desk strapped to his back, accompanied by a ghostly, indeterminate entity dressed for some reason like a member of the 19th-century European upper middle class. We, the lab rats on which a new medicine is being tested, can only hope that the headmaster will offer us an explanation. Is it Halloween yet?
Google Translate has obviously never spent a day of its life in school; has never had to wonder what to wear today; and has no experience dealing with people. And that's just one sentence. Would you trust it to translate the whole book for you?
This inability to sense a larger context means that computer translation is also inadequate for specialized fields. Many fields use their own jargon. They also break down further into various subfields, each with its own peculiar terminology.
One way MT responds to this type of complexity is by bundling terms together. Recently I was dealing with a German contract that had been machine-translated into English. It included this string of words: “Installation/installation and commissioning of the complete installation at the site.” That’s three different iterations of “installation,” each one translating a different word in the original (respectively Aufstellung, Einrichtung, and Anlage). Yes, any of these can be “installation,” but they can also be “facility,” “setup,” “plant” or a number of other things. It all depends on context. Conversely, MT doesn’t just bundle, it fragments. The same program couldn’t decide if Auftraggeber was “client,” “contracting authority,” or “customer” – it used all three, seemingly at random.
MT knocks up against the real world with all the obliviousness of Chauncey Gardiner in Being There. The personal name Timo Kloster is rendered as “Timo Monastery” because the computer can’t distinguish a person from a building. BAG-Rechtsprechung – the case law of the German Federal Labor Court – becomes “bag case-law,” or jurisprudence relating to bags. Idiomatic expressions are rendered literally: es liegt auf der Hand (“it’s obvious”) becomes “it lies on the hand,” while keine Zukunftsmusik (meaning something that is already a reality) is “no future music.”
Despite these howlers, increasingly I find myself doing less straight translation, and more post-editing of machine translation. The translation profession as a whole is going in this direction. But while MT is certainly no future music, it lies on the hand that only the astute mind of the human translator can make the text look like a human product.