Our idea of machine translation often resembles that of a Babel fish — the fictional animal imagined by Douglas Adams in his “Hitchhiker’s Guide to the Galaxy“. Once inserted into the ear, this odd creature “excretes a telepathic matrix” by which its host can instantly understand any language. Transposition between languages, in that sense, could achieve perfect one-to-one correspondence, while the device simply fades into the background. Translation just happens, as if by magic.

I must admit, using contemporary digital translation tools does occasionally feel like magic. Who can forget using a camera to read a restaurant menu in a foreign language for the first time? Yet, the image of a seamless universal language machine never sat well with me. What would happen in the Galaxy, for example, were a toddler to use the Babel fish from an early age, before acquiring a native tongue? What language would the child learn? And what would it hear from other people — perhaps not a language, but the telepathic sense of dwelling within pure thought, unmarred by culture?

Of the dreams humans have dreamt, the dream about a single, universal human language was a vivid one. John Wilkins, the British linguist and one of the founders of the Royal Society, published his “Essay Towards a Real Character” in 1668. In this heavy tome, Wilkins proposed a novel writing system, which would more perfectly correspond to the world as we know. By learning to write in it, the nations of the world would “improve knowledge,” “facilitate mutual commerce,”  and even “clear the differences between religions.” Several of those pretended mysterious and profound notions, he wrote, expressed in a great swelling of words, on closer scrutiny, will appear to be “either nonsense or flat or jejune.” What a wonderful word, jejune — though not found in the book’s limited lexicon.

Wilkins thought his artificial language to harbinger a new philosophy, unable to express any false ideas. A precursor to modern calculus, it had a direct influence on the “universal characteristic” of Leibniz and Newton. Although we tend to think of calculus today in the context of mathematics, the 17th century conversation about its possibilities concerned metaphysics more generally. Such an artificial language, according to Leibniz, would align the kind of truths found in encyclopedias — describing the world — with the kind of proofs found in geometry textbooks — revealing its underlying logic. Translated into an artificial language, the messiness of human thought would give way to the ordered beauty of mathematics. The ultimate goal for universal character was to facilitate a conversation without the possibility of misunderstanding.

Imminent Research Report 2025

A journey through localization, technology, language, and research.

The ultimate resource for understanding AI's impact on human interaction. Written by a global community of multidisciplinary experts. Designed to help navigate the innovation leap driven by automatic, affordable, and reliable translation technology.

Secure your copy now

Perfect or “lossless” communication may be useful for science or commerce. It isn’t always so for social purposes. In a perfect world, a telepath would “get” the sense of a foreign phrase without much effort or friction. But the brute getting of sense can also be quite boring. Communicating with someone perfectly would be a little like talking to oneself. The very interest of sharing something lies in the slight incompatibility of views and experiences. I am grateful you are not me. In conversation, we get to step outside ourselves. Those who speak multiple languages savor the small quirky ways particular to each. Language doesn’t just contain meaning, it also represents a particular style of thinking. Each language holds its own sense of flare and pizazz. I don’t just speak, I think differently across languages. The joy of translation comes in approximating these verbal intangibles, in a process that produces entirely new and often unexpected resonances.

Centuries later, in the 1930s at Cambridge, two brilliant students sat in a lecture on the foundations of language and mathematics. One of them was the young Alan Turing, who went on to create the model of a general-purpose computer. The other was Margaret Masterman, a pioneer of machine translation. Typical for Wittgenstein, lectures devolved into a series of thought experiments: What if I was saying something that made sense by accident, he asked, without myself understanding the meaning of my words? What if I took drugs and accidentally started speaking Russian? What if I was hallucinating and by chance guessed what you meant, without speaking your language?

His point, I think, was to question the ease with which we sometimes pretend to understand each other. I am hungry, my niece says, and I feed her leftover pancakes which she seems to enjoy immensely. But isn’t hunger also such a personal, private feeling, Wittgenstein wondered? How can we really understand what another thinks or feels?  My hunger is only the approximation of a child’s hunger. Worse yet, the word “hunger” expresses some imperfect average of hungers previously encountered. Some people feel tired when they are hungry, others energetic. Some languages don’t make a neat distinction between “hunger” and “thirst” in the way that English does. Others say “I have a need for food.” Is that the same as hunger or different? A gap always exists between the word, world, and my private experience of what it is meant to represent. And the more we zoom into that gap, the more improbable the task of translation becomes. How can we hope to get any message across cultures, if we can’t even do it well between two people speaking the same language?

Wittgenstein wasn’t in the habit of offering solutions. His lectures usually fizzled out in a haze receding metaphors. Turing and Masterman solved their teacher’s language puzzles distinctly. Turing famously imagined a universal machine capable of converting written instructions into its own inner physical states. In this way, one could ensure that a machine’s internals would correspond exactly to its received programming. A Turing machine could never lie in that sense. There could be no “misunderstanding” between programmers and their devices. Universal machines perform understanding in its ideal, totally faithful form. Any misshapen or incorrect instruction simply does not compute. When a machine language interpreter “doesn’t understand” it returns a compilation error, unable to continue.

How can we hope to get any message across cultures, if we can’t even do it well between two people speaking the same language?

Later, after WWII, Turing would also propose an “imitation game,” by which the machine could deceive others, by misrepresenting its assigned gender. And though now programmed to lie, the machine would still remain true to its instructions. Turing’s universal computer ultimately represented a universal translation device, enacting the perfect transposition between human ideas (software) and machine components (hardware).

Margaret Masterman took another approach to translation from that of Turing. Unfortunately, her influence on computer science has been all-too-often neglected. In thinking about translation in the 1950s, Masterman abandoned the search for ideal or grammars which characterized much of prescriptive linguistics of the time. Instead, she embraced the pragmatic “ordinary language” philosophy pioneered by the likes of Wittgenstein, J.L. Austin, and Gilbert Ryle. At the time, she was particularly taken with a popular series of language books for first-time learners called English Through Pictures, authored by I. A. Richards and Christine Gibson. The pragmatists believed language to be discovered through use. Therefore, learners could be best taught the meaning of new words by referring to simple pictures, depicting stick figures in ordinary situations. There was no need to explain the meaning of words like “here,” “there,” “hat,” “hand,” and so on. “A word is known by the company it keeps,” the pragmatists liked to insist, and “a picture is worth a thousand words.”

If children were best taught through pictures, how would we go about teaching computers to speak or translate languages? Inspired by Wilkins and his “universal character” writing system, Masterman’s lab (called the Cambridge Language Research Unit) proposed an intermediary machine notation they initially called “mechanical pidgin” and later “interlingua.” The interlingua was to machine what the stick figure was to a child. The machine notation, Masterman’s group thought, would represent the “naked ideas” expressed in any language, without the vestments of any specific culture. In this way, the English word “puppies” could be rendered into something like the machine equivalent of “young dog plural,” ignoring the specifics of the English juvenile form and plural construction. In a surprising (for me) turn of events, that computer language was developed by R.H. Richen, the Assistant Director of the Commonwealth Agricultural Bureaux and Director of the Cambridge Bureau on Plant Breeding and Genetics. Masterman and Richens called this intermediary language NUDE. Once denuded, a word from one language could then be re-dressed in another.

A word is known by the company it keeps.

Let’s remember also that computers were still a novelty in academic research in the 1950s. Cambridge University, where Masterman’s lab was housed, built its first computer — the Electronic Delay Storage Automatic Calculator (EDSAC) — in 1948, which was replaced by EDSAC 2 only in 1958. This was a beast of a machine, powered by thousands of vacuum tubes, mercury delay lines, and accumulators, all taking up several rooms in a laboratory. Revolutionary for the time, it had the available memory, for both data storage and programming instructions, of only 512 18-bit words (roughly equivalent to a page of printed text).

It is incredible to think then, that Masterman saw the potential of machine translation decades before it was practically feasible. Few of her lab’s ideas could be programmed into working computer programs at the time, due to severe hardware limitations. For about a decade in the 1950s, Masterman and Richen were basically groping in the dark. And what they found there was once again the dream of a universal language — the interlingua — a language of Babel towers and Babel fish, a language free from human error and ambiguity, a language perfectly true to the world, proper and correct everywhere for all times.

Though built on a set of different, much more advanced technologies, modern AI-enabled machine translators share in that universal vision. A translation of any message from one language to another, the thinking goes, should work like a formula — where A always leads to B, and B back to A. Languages, in that model, can correspond to each other exactly, without ambiguity, mano-a-mano. But that doesn’t sit quite right, does it?

A direct, one-to-one congruence may hold when giving directions or ordering a meal. The social world however can be messier than a map or a menu. Translation in the wild also requires a sense of discernment, tact, and grace. Just because a machine praises your cookery, I am paraphrasing René Descartes loosely here, doesn’t mean you would invite it over for dinner. Words don’t just mean things on paper. The same sentence spoken to a friend means something different when spoken to a work associate or to a family elder. Other things are better left unsaid. Languages such as Korean, for example, have several layers of honorifics not found in English. It isn’t that Americans cannot logically express or understand that same level of politeness. The social rituals themselves, which depend on age and status, differ drastically between the two cultures. The underlying value systems and their history differ. As such, they cannot ever be rendered “perfectly” across languages. Context matters. And contexts are incongruent by definition. Sometimes wrong and never exactly right — at best, a translation can hope to become “appropriate” to the situation.

The beauty and inadequacy of machine translation was made clear to me recently, during a three-hour interview given by the Ukrainian President Volodymyr Zelenskyy to Lex Fridman, an American podcaster, on January 5th, 2025. The event was billed to be “translated by AI,” erroneously as it turns out. The trouble began long before AI. In advance of the event, Fridman lobbied to hold the interview in Russian — a language he speaks with the strained difficulty of a heritage speaker. The podcaster, it seems, held lofty ambitions to act as a mediator between Zelensky and Vladimir Putin, “on a boat in the middle of the Black Sea,” in his words.

Ukraine’s president was visibly amused if not annoyed by the suggestion, which laid bare Fridman’s ambition and naivete. Nobody is sailing anywhere amidst the drones on the Black Sea. Neither could the leader of a nation at war speak freely in his enemy’s tongue. Language lay at the origins of this conflict: Russia alleged the suppression of its language among Russian speakers on Ukrainian territory, mandated centrally, alongside Nazism and antisemitism. The facts — that Zelenskyy was both Jewish and a native speaker of Russian — didn’t stand in the way of good propaganda, repeated almost verbatim by the host. Fridman so badly wanted to enter diplomatic history at sea, he forgot to facilitate the historic occasion in his studio.

And what they found there was once again the dream of a universal language — the interlingua — a language of Babel towers and Babel fish, a language free from human error and ambiguity, a language perfectly true to the world, proper and correct everywhere for all times.

Forty-some awkward minutes later, an uneasy compromise was reached: Zelenskyy proposed Fridman use any combination of Russian, Ukrainian, or English to ask his questions. A president’s duty required answering in Ukrainian. To make things more complicated, this mixture of languages was funnelled through the official translator, into Ukrainian, off camera. The translator struggled to keep up, given the frequent code and language switching, as evidenced by several of the president’s apologies to the audience.

Yet most of the English-speaking audience heard neither the translator nor Zelensky nor Fridman directly. It took my brain some time to register that the conversation was actually dubbed, uncannily, in the voice of the participants. Alongside the default English language, listeners could also select Russian, Ukrainian, or the mixed original track, labeled (for the lack of a better option) as UK English. In every case, these “translations” retained each of the speaker’s distinctive vocal color, tone, and even accent in the received language.

As it turns out, ElevenLabs, the company behind this neat technological trick, wasn’t really in the business of translation at all. Rather, it specialized in creating realistic audio deep-fakes, which, according to Recorded Future (“the world’s largest threat intelligence company”), were also used extensively by Kremlin to spread disinformation across the globe.

The translation itself was likely done by using a mixture of automated tools and human input. I can further surmise that even the seemingly “simple” task of transcription couldn’t be fully automated either, given the unprecedented tri-lingual nature of the conversation. The translation was probably a hybrid effort as well, involving AI and human translators, alongside editors from both teams. At one point, when asked about corruption during the war, Ukraine’s president answered with something like “we slapped every hand,” meaning “we punished everyone.” Misunderstanding the idiom, Fridman’s team rendered the Ukrainian into the English “we slapped them on the wrist,” implying the opposite of the conveyed message (meaning, “we punished them lightly”). Later, Fridman thoughtfully appealed to his Telegram channel, offering his audience several possible options, including:

1. We slapped everyone’s hands
2. We hit everyone’s hands
3. We beat everyone’s hands
4. We cracked down on everyone
5. We cracked down hard on everyone
6. We firmly reprimanded everyone
7. We punished everyone

The audience chose option five, “we cracked down hard on everyone,” bringing the text closer to its intended meaning. Further, ElevenLabs could reinsert the crowd-sourced correction into the original video, in Zelenskyy’s own voice, where it stands now, spliced seamlessly by AI.

The exchange, advertised to be “translated by AI,” under examination, reveals to have involved professional translators, public relations experts, software, a team of audio engineers, linguists, presidents, YouTube and YouTubers, alongside crowd contributions from the audience.

Looking back at that tangled message between recipients – separated by language, medium, war, and technology – I cannot help but revisit the evolving history of machine translation. No, it can never become fish, magic, or telepathy, as Douglas Adams imagined it. The task of the translator remains a doggedly human, analog enterprise, mired in aesthetic and political conundrums. “Half-way between poetry and doctrine,” as Walter Benjamin had it, translation cannot be solved in the way of checkers, restaurant menus, or sudoku. Its mystery deepens by the measure of entrenched difference between two strangers, so similar yet unable and unwilling to become one.

Dennis Yi Tenen

Associate Professor at Columbia University

Dennis Yi Tenen is an associate professor of English and comparative literature at Columbia University, where he also co-directs the Center for Comparative Media. His recent publications include Literary Theory for Robots (Norton, 2024). He lives, writes, and codes in New York City.

1. Jones, Karen Sparck. R.H. Richens: Translation in the NUDe. Early Years in Machine Translation: Memoirs and Biographies of Pioneers, edited by John W. Hutchins, John Benjamins Publishing Company, 2000, pp. 263-278.
2. Liu, Lydia H. After Turing: How Philosophy Migrated to the AI Lab. Critical Inquiry, vol. 50, no. 1, Sept. 2023, pp. 2–30.
3. Masterman, Margaret, and Yorick Wilks. Language, Cohesion and Form. | Cambridge University Press, 2005.
4. Wittgenstein, Ludwig. The Blue and Brown Books: Preliminary Studies for the Philosophical Investigations. Harper & Row, 1965.