I have long remembered a conversation I had 20 years ago with one of my professors, an expert in what we then called artificial intelligence, which, in many ways, is wildly different to what we now call AI. In this exchange, he confidently told me there was no point learning a second language.
Computers would soon erase language barriers, he said. Not only by translating written text, but also in real time using audio, to make conversations flow smoothly even if everyone involved is speaking in a different tongue.
While he may have been a couple of decades premature, he was correct. AI can now translate, in real time, between virtually any two languages. But while learning a language requires hard work and perseverance, and AI demands only that you download an app, there are still plenty of reasons to do things the old-fashioned way.
Advertisement
is a translation expert at the University of Nottingham, UK, so certainly isn’t unbiased. But he believes there is enjoyment and satisfaction to be had in learning a language that alone justifies the effort, even if AI can already do it flawlessly. “Think of chess,” says Mével. “Computers can handle the game perfectly. And yet people still play.”
The thing is, AI translation is actually still a long way from flawless. Translation models have little or no concept of wider context, idioms, slang or humour, says Mével. While the instructions for a mundane gadget – your internet router, say – might be translated quickly and cheaply by AI with relatively little harm, Mével believes we still need to rely on humans to convert the best novels and films from language to language.
“The top-tier translations are all produced by humans,” he says. “AI is not very good at picking up on what might be described as beauty in a text. All it’s doing is looking for the next most plausible words.”
In essence, AI can take a life-affirming, moving French novel and turn it into a distinctly average English one, but a talented human translator can retain the emotional resonance. It is a talent that is part art and part craft, says MĂ©vel.
The white whale of AI translation might, ironically, be Moby-Dick, says computer scientist at the University of Cambridge. While AI models may be able to handle the nuts and bolts of the task, they simply aren’t equipped to grapple with the themes, allusions or metaphors inherent to Herman Melville’s classic novel.
“What does it mean that the whale is white? White is used in the Christian Western tradition to mean cleanliness and purity,” says Holden. “Now, if you want the actual subtlety and all that part of Moby-Dick to carry over into a translation to a language that’s from a completely different culture, then you have to have someone who really understands both languages. Not just a stochastic machine.”
Holden says that what current AI models do is impressive, but they haven’t yet made learning a language obsolete. He is optimistic that it will get there in time, but steadfastly refuses to be pinned down to a date – while he believes there is nothing inherent to the human brain or its abilities that can’t be replicated, he thinks we could still be in for a very long wait.
More prosaically, mediating translation through machines involves some serious privacy implications, says at the University of Oxford. Everything an AI model listens to and translates is open to surveillance by the company that made it, who may use it to serve you targeted adverts or simply sell your data on to others. “That doesn’t mean we shouldn’t use it, but it is to say that there are trade-offs and we need to decide if convenience is worth the loss of privacy,” says Wachter.
That could be particularly concerning when speaking in some countries, says , also at the University of Oxford. “It’s easy in a conversation to say more than you meant to, and if you live in a country that’s not democratic, and you might have said something sensitive… you can see where this is going.”
Despite these negatives, the allure of convenience, cost and speed mean there is already a clear move away from manual translation to machines. at the University of Leeds, UK, who lectures on AI and linguistics, says he sees a noticeable drop each year in the number of students applying to study language and translation. He puts this down to both AI’s reduction in demand for human translators and AI-powered tools like Duolingo making it easier and cheaper to learn a language if it is still needed.
But, crucially, this drop is only seen among undergraduates: demand for more intense postgraduate study remains steady. On these courses, students not only delve deeper into a language, expanding their vocabulary and accuracy, but also learn about differences in culture to precisely capture the sort of nuances and social references that AI can’t. Demand for these skills will be longer-lasting and robust, says Sharoff.
“Unless we have a completely homogeneous culture emerging in which there is no difference between people living in Yorkshire and people living in Patagonia, I don’t think that the need for some kind of intercultural communication actually disappears,” he says.
But when AI is so immediate and easy, it is difficult not to be tempted, even for someone like Sharoff who loves language, enjoys learning and has made a career of linguistics. He relied on AI apps to translate on a recent trip to China. “If I had an infinite amount of time, I would have learned more of Chinese, right? But in the absence of infinite time, the app did a decent job,” says Sharoff.
