I don't want to be overly dramatic and overly negative about the AI translations I've been working with. They are bad, yes, but I don't want to overstate their badness because that would obscure the specific points I'm making. Some AI translations, the best AI translations, are not that bad. Some of them are still bad to the point of being unusable, but others are better. They're not good, but they're mostly serviceable, and it's extremely impressive that a machine can come up with something serviceable, something comparable to the work of a very mediocre human translator.
A client who hires a subpar translator who accepts being underpaid, in order to avoid paying professional rates for a professional, is getting subpar work. A client who uses AI to get work cheap and fast is getting worse than subpar work. But AI is getting better, it might soon be at the point where laymen can't tell the difference, and then, using AI instead of paying a human will mostly be a labour rights issue, and that's a far thornier question. (Note that I'm not talking about using AI translators to read something for yourself, or to communicate in your daily life: I'm talking about AI translation for publication, using AI for something you expect other people to pay money for.)
My actual point about AI translation is that even when it's fairly good, when it makes few errors and conveys the message intelligibly, it lacks something. I'm not talking about heart and soul here, nothing to do with some intangible human quality: I'm talking about specificity. AI works with great averages, and so it automatically irons out nuance. If you write something unusual, AI will assume it's an error, instead of an intentionally unusual statement. This is regression to the mean, and based on the texts I'm working with, it's an Anglophone, American mean. If you say something that's true of 1980s Hungary, it might slightly alter the sentence to "make sense" for 1980s US. Some alterations are factual, these are more serious errors but also easier to edit out. But other things are harder to catch, slight shifts in tone and valence, an erasure of the original, specific, non-American perspective, and the end result is a text that doesn't have anything wrong with it, but is markedly simpler and dumber than it should be. And flattening complex, knotty, peripheral perspectives into something closer to a monoculture is, in the long term, intellectually devastating.
A while ago, I was proofreading AI translations of subtitles. The video was an interview with a couple of game devs talking about their game. âThe gameâ was frequently mentioned. And then, all of a sudden, the translation talked about football. I did a double take. Where had that come from?
The original sentence went somewhere along the lines of âfrom the veterans who have been with us for a long time to young people only just getting into the gameâ and the AI translation assumed that there was only one game young people could be getting into. It had to be football. So thatâs what it put in.
That moment really clarified for me this regression to a cultural mean described above. Only one thing made cultural sense, right? Too bad the actual video was about something entirely different.
Yes that's such a great example of what I was talking about! And this one is obvious enough and weird enough that you could catch it in proofreading, but if the mistake is subtler (or if there's no proofreading, and let's face it there usually isn't), someday soon we'll end up in a world where everyone likes football and drinks beer and does, says, thinks, believes only the most statistically average things to do, say, think and believe.
