The Synthesized Polyglot: How AI is Erasing the Concept of a ‘Foreign’ Language

Diego Marani, a writer and interpreter, recently described a moment that sticks in your memory. Serving as a simultaneous interpreter at an altar during an ecumenical council, he came to the conclusion that translation was not his calling. The position, the clerical tone, and the raised arms in the particular gesture of liturgical address were all part of the job of becoming a priest. He pointed out that what he accomplished that day was beyond the capabilities of any machine. It’s a very specific, almost personal example. However, it highlights a point that is often overlooked in the current discussion about AI translation: language is more than just information transferring between containers.

These days, AI is truly adept at transferring that data. A decade ago, the state of real-time voice-to-voice translation would have seemed unthinkable. It is based on neural networks that capture phonetic characteristics, regional vocabulary, and emotional register. When a person in a Chicago meeting room speaks English and their Tokyo counterpart hears Japanese, the gap closes almost instantly. The friction that once made communicating in a foreign language difficult is disappearing. The language barrier is no longer an obstacle in the majority of practical interactions, such as business negotiations, travel arrangements, and customer service. It would be strange to act otherwise; that is a real accomplishment.

But pay attention to what’s going on beneath it. When major AI systems are tested for multilingualism, researchers at Brown University, the University of Oregon, and Hong Kong University of Science and Technology consistently find that these models translate into English much more accurately than they translate out of it. When working in Korean, Vietnamese, Indonesian, or Tamil, they are more likely to produce false information and perform significantly worse when responding to factual questions in non-English languages. One well-known chatbot accurately responded in English when asked in Spanish where the next White Lotus season would take place, but it only provided “somewhere in Asia” when asked in the translated version of the same question. The technology is not as smooth. It is approximate in one direction and fluent in the other, which is toward English.

Beyond the technical, this asymmetry is significant. It’s possible that AI translation is subtly strengthening English’s hegemony in international trade and communication rather than undermining it. More English text on the internet, more English in scholarly publications, and more English in the data sets that teach these models how language functions are examples of training data that reflects current power disparities. “If you don’t publish papers in English, you’re not relevant,” stated Pascale Fung, director of the Center for AI Research at the Hong Kong University of Science and Technology and a seven-language speaker. The AI tools being developed to overcome linguistic barriers are partially the result of the same structural bias that they are meant to address.

AI Language Translation & the Synthesized Polyglot — Key Facts & Context

There is also the cultural dimension, which is more difficult to quantify but may have greater long-term implications. Regional humor, metaphor, the particular meanings of words that have no clear equivalent, and the silence that can mean one thing in a Japanese business conversation and another in a Brazilian family dinner are just a few examples of the things that language carries that word-for-word or even sentence-level translation cannot consistently replicate. AI systems trained on English-dominated data are likely to completely overlook the fact that the Basque word “uso,” which means dove, can serve as an insult, giving it the serene symbolism that the English-speaking internet associates with the bird. These are not examples of edge cases. They represent the actual use of language.

There is a sense that the discussion is being held at the incorrect level of abstraction as this technology advances. Will AI eliminate the need to learn a foreign language? This is the topic of discussion in scholarly articles and editorial columns. is true, but it takes a backseat to a more unsettling one: what happens to regional and minority languages around the world when the main impact of AI translation is to make English more accessible while further lagging behind Swahili, Quechua, and Tagalog? According to Stanford research from 2025, the communities are completely shut out of AI tools because their languages don’t produce enough training data for the models to care about, not because translation is flawed.

The reversibility of this trajectory is still unknown. For Latin American, African, and Southeast Asian dialects, some researchers are creating new training data sets. With its PaLM 2 model, Google has made credible attempts to train on over 100 languages. These are real steps. However, the amount of English text continues to increase, and large language models’ statistical architecture will inevitably favor the language with the greatest representation. There is a synthetic polyglot. The question of whether it speaks for everyone or primarily for those who have already been heard is still up for debate.