The End of Foreign-Language Education – The Atlantic

Thanks to AI, people may no longer feel the need to learn a second language.

A few days ago, I watched a video of myself talking in perfect Chinese. I’ve been studying the language on and off for only a few years, and I’m far from fluent. But there I was, pronouncing each character flawlessly in the correct tone, just as a native speaker would. Gone were my grammar mistakes and awkward pauses, replaced by a smooth and slightly alien-sounding voice. “My favorite food is sushi,” I said—wo zui xihuan de shiwu shi shousi—with no hint of excitement or joy.

I’d created the video using software from a Los Angeles–based artificial-intelligence start-up called HeyGen. It allows users to generate deepfake videos of real people “saying” almost anything based on a single picture of their face and a script, which is paired with a synthetic voice and can be translated into more than 40 languages. By merely uploading a selfie taken on my iPhone, I was able to glimpse a level of Mandarin fluency that may elude me for the rest of my life.

HeyGen’s visuals are flawed—the way it animates selfies almost reminded me of the animatronics in Disney’s It’s a Small World ride—but its language technology is good enough to make me question whether learning Mandarin is a wasted effort. Neural networks, the machine-learning systems that power generative-AI programs such as ChatGPT, have rapidly improved the quality of automatic translation over the past several years, making even older tools like Google Translate far more accurate.

At the same time, the number of students studying foreign languages in the U.S. and other countries is shrinking. Total enrollment in language courses other than English at American colleges decreased 29.3 percent from 2009 to 2021, according to the latest data from the Modern Language Association, better known as the MLA. In Australia, only 8.6 percent of high-school seniors were studying a foreign language in 2021—a historic low. In South Korea and New Zealand, universities are closing their French, German, and Italian departments. One recent study from the education company EF Education First found that English proficiency is decreasing among young people in some places.

Many factors could help explain the downward trend, including pandemic-related school disruptions, growing isolationism, and funding cuts to humanities programs. But whether the cause of the shift is political, cultural, or some mix of things, it’s clear that people are turning away from language learning just as automatic translation becomes ubiquitous across the internet.

Within a few years, AI translation may become so commonplace and frictionless that billions of people take for granted the fact that the emails they receive, videos they watch, and albums they listen to were originally produced in a language other than their native one. Something enormous will be lost in exchange for that convenience. Studies have suggested that language shapes the way people interpret reality. Learning a different way to speak, read, and write helps people discover new ways to see the world—experts I spoke with likened it to discovering a new way to think. No machine can replace such a profoundly human experience. Yet tech companies are weaving automatic translation into more and more products. As the technology becomes normalized, we may find that we’ve allowed deep human connections to be replaced by communication that’s technically proficient but ultimately hollow.

AI language tools are now in social-media apps, messaging platforms, and streaming sites. Spotify is experimenting with using a voice-generation tool from the ChatGPT maker OpenAI to translate podcasts in the host’s own voice, while Samsung is touting that its new Galaxy S24 smartphone can translate phone calls as they’re occurring. Roblox, meanwhile, claimed last month that its AI translation tool is so fast and accurate, its English-speaking users might not realize that their conversation partner “is actually in Korea.” The technology—which works especially well for “high-resource languages” such as English and Chinese, and less so for languages such as Swahili and Urdu—is being used in much more high-stakes situations as well, such as translating the testimony of asylum seekers and firsthand accounts from conflict zones. Musicians are already using it to translate songs, and at least one couple credited it with helping them to fall in love.

One of the most telling use cases comes from a start-up called Jumpspeak, which makes a language-learning app similar to Duolingo and Babbel. Instead of hiring actual bilingual actors, Jumpspeak appears to have used AI-generated “people” reading AI-translated scripts in at least four ads on Instagram and Facebook. At least some of the personas shown in the ads appear to be default characters available on HeyGen’s platform. “I struggled to learn languages my whole life. Then I learned Spanish in six months, I got a job opportunity in France, and I learned French. I learned Mandarin before visiting China,” a synthetic avatar says in one of the ads, while switching between all three languages. Even a language-learning app is surrendering to the allure of AI, at least in its marketing.

Alexandru Voica, a communications professional who works for another video-generating AI service, told me he came across Jumpspeak’s ads while looking for a program to teach his children Romanian, the language spoken by their grandparents. He argued that the ads demonstrated how deepfakes and automated-translation software could be used to mislead or deceive people. “I’m worried that some in the industry are currently in a race to the bottom on AI safety,” he told me in an email. (The ads were taken down after I started reporting this story, but it’s not clear if Meta or Jumpspeak removed them; neither company returned requests for comment. HeyGen also did not immediately respond to a request for comment about its product being used in Jumpspeak’s marketing.)

The world is already seeing how all of this can go wrong. Earlier this month, a far-right conspiracy theorist shared several AI-generated clips on X of Adolf Hitler giving a 1939 speech in English instead of the original German. The videos, which were purportedly produced using software from a company called ElevenLabs, featured a re-creation of Hitler’s own voice. It was a strange experience, hearing Hitler speak in English, and some people left comments suggesting that they found him easy to empathize with: “It sounds like these people cared about their country above all else,” one X user reportedly wrote in response to the videos. ElevenLabs did not immediately respond to a request for comment. (The Atlantic uses ElevenLabs’ AI voice generator to narrate some articles.)

Gabriel Nicholas, a research fellow at the nonprofit Center for Democracy and Technology, told me that part of the problem with machine-translation programs is that they’re often falsely perceived as being neutral, rather than “bringing their own perspective upon how to move text from one language to another.” The truth is that there is no single right or correct way to transpose a sentence from French to Russian or any other language—it’s an art rather than a science. “Students will ask, ‘How do you say this in Spanish?’ and I’ll say, ‘You just don’t say it the same way in Spanish; the way you would approach it is different,’” Deborah Cohn, a Spanish- and Portuguese-language professor at Indiana University Bloomington who has written about the importance of language learning for bolstering U.S. national security, told me.

I recently came across a beautiful and particularly illustrative example of this fact in an article written by a translator in China named Anne. “Building a ladder between widely different languages, such as Chinese and English, is sometimes as difficult as a doctor building a bridge in a patient’s heart,” she wrote. The metaphor initially struck me as slightly odd, but thankfully I wasn’t relying on ChatGPT to translate Anne’s words from their original Mandarin. I was reading a human translation by a professor named Jeffrey Ding, who helpfully noted that Anne may have been referring to a type of heart surgery that has recently become common in China. It’s a small detail, but understanding that context brought me much closer to the true meaning of what Anne was trying to say.

But most students will likely never achieve anything close to the fluency required to tell whether a translation rings close enough to the original or not. If professors accept that automated technology will far outpace the technical skills of the average Russian or Arabic major, their focus would ideally shift from grammar drills to developing cultural competency, or understanding the beliefs and practices of people from different backgrounds. Instead of cutting language courses in response to AI, schools should “stress more than ever the intercultural components of language learning that tremendously benefit the students taking these classes,” Jen William, the head of the School of Languages and Cultures at Purdue University and a member of the executive committee of the Association of Language Departments, told me.

Paula Krebs, the executive director of the MLA, referenced a beloved 1991 episode of Star Trek: The Next Generation to make a similar point. In “Darmok,” the crew aboard the starship Enterprise struggles to communicate with aliens living on a planet called El-Adrel IV. They have access to a “universal translator” that allows them to understand the basic syntax and semantics of what the Tamarians are saying, but the greater meaning of their utterances remains a mystery.

It later becomes clear that their language revolves around allegories rooted in the Tamarians’ unique history and practices. Even though Captain Picard was translating all the words they were saying, he “couldn’t understand the metaphors of their culture,” Krebs told me. More than 30 years later, something like a universal translator is now being developed on Earth. But it similarly doesn’t have the power to bridge cultural divides the way that humans can.