Far from a restrictive act of copying, a translator restores the meaning of a text by means of an elaborate process that requires imagination, ingenuity, and freedom.
—Jhumpa Lahiri, “In Praise of Echo”[T]rust is a hard commodity to build, in any interpersonal communication, and all too easy to ruin. No one likes taking another person’s word, and yet in translation, that is literally what the reader is asked to do.
—Mark Polizzotti, Sympathy for the Traitor
The project of machine translation was already in its fifth decade when the search engine AltaVista introduced Babel Fish, at the end of 1997. Named after the “leech-like” creature that functions as a universal translator in Douglas Adams’s The Hitchhiker’s Guide to the Galaxy (1979), it broke new ground by offering translation for free online. Previously, machine translation (MT) was something for which you had to pay and wait, since humans generally intervened to tidy up what the machine produced. AltaVista promised instant results—no human lag required—whether you wished to have an entire Webpage translated (without altering the graphics) or input an inscrutable chunk of foreign text. Here, in other words, was real-time translation of and on the Web. Digital Equipment Corporation, AltaVista’s parent company, declared that Babel Fish had “broken the Internet language barrier.”11xQuoted in Victoria Shannon, “The End User: The Power of Babel—Technology,” New York Times, May 3, 2006; https://www.nytimes.com/2006/05/03/technology/03iht-ptend04.1654790.html.
“Barely breached” would be more accurate. Initially, Babel Fish could translate English text only into German, French, Spanish, Italian, and Portuguese, and vice versa (but not, say, French to Spanish), though its capabilities would expand in the ensuing years. Even within that circumscribed domain, Babel Fish often stumbled, especially with names, technical terms, and idiomatic expressions. Its hold on grammar, too, was shaky. First-time users raved, but professional translators balked. Mischief-makers played a game—called “round-tripping”—in which a translated text is rendered back into its original tongue to see what distortions arise. In this regard, Babel Fish was very obliging.
Shortly after the service’s debut, Umberto Eco, the novelist, semiotician, and satirist, spotted the invitation on AltaVista’s homepage and took the bait, asking Babel Fish to translate English sites into his native Italian. His “first shock,” he reported a few months later in his “La Bustina di Minerva” column in the Italian newsmagazine L’Espresso, came when he saw a page titled “Gli impianti di Shakespeare”—or “Shakespeare’s Plants.”22xUmberto Eco, “La vera storia dei pali del Papa,” L’Espresso, January 15, 1998. The rules-based system that AltaVista employed (created by the Parisian company Systran) had latched on to the wrong sense of the English “works.” The Italian translation should have been opere, the “works” of an artist; instead, it selected impianti, the “works” of an industrialist. Eco noticed other errors: An author was credited with having many ventilatori (the whirring kind of “fans”); a publisher was referred to as Harcourt “Support” (i.e., “Brace”); the Polish people (i polacchi) were reduced to poles (i pali). In a subsequent column, Eco reported on round-tripping, using the opening lines of Dante’s Inferno as a test case. Round-trip Dante, he explained, and you get proof that the machine poses no threat to “il divin poeta.” Send Dante’s lines through several more permutations, and you get modern poetry.33xUmberto Eco, “Trionfante ritorno a Babele. Como giocare seriamente con Altavista,” L’Espresso, February 19, 1998. I am grateful to the good people at L’Espresso for responding promptly to my request for a scan of this column.
Eco kept playing with Babel Fish, and in Mouse or Rat?, his 2003 book on “translation as negotiation,” he again brought up its shortcomings to illustrate the qualities of effective translation. To begin with, Eco pointed out, translation does not consist of mechanical synonym-swapping. If that were the case, Babel Fish’s execution would have been flawless. The words we rely on most have multiple senses, and to determine the specific one invoked on any given occasion, the translator must decipher contextual clues. This Babel Fish could not do—even when given the larger window of multiple sentences. Inputting the opening verses of the Authorized Version’s Genesis, Eco was amused to find the English “spirit of God” transformed into the Spanish “el alcohol de dios.”44xUmberto Eco, Mouse or Rat? Translation as Negotiation (London, England: Phoenix Paperback 2004), 14. Moving between languages, furthermore, the translator needs to honor the grammatical and syntactic mores of both. No competent English user would say “divide waters of waters” (as Babel Fish rendered “aguas de aguas”) or begin a sentence, “In the God who began created heaven…” (as Babel Fish did in multiple languages). Even contextual and grammatical wherewithal could get the translator only so far, however. To decide whether “works” ought to be opere or impianti, the translator needed to know a simple fact that Babel Fish lacked—“that Shakespeare was a poet and a playwright and not an industrial tycoon.”55xIbid., 13.
The machine’s miscues clarified that translation depends on more than a large vocabulary and grammatical proficiency in two tongues. The translator must also possess extensive “world knowledge.”66xIbid., 18. Only one so equipped can undertake the multiple negotiations—with languages, with the author, with the imagined reader, with a very real publisher—that translation entails. Babel Fish might have its uses (and amusements), Eco granted, but it could never handle all that. Real-time machine translation was not ready for the world.
New Techniques, Old Ambitions
Is it now? Within a decade of its release, Babel Fish began to sink to the bottom of the Web, along with AltaVista, DEC, and its later owner, Yahoo, all doomed by the ascendance of Google. Google Translate (launched in 2006) demonstrated that a corpus-based statistical approach was superior to Babel Fish’s often cross-wired rules, though it too was prone to embarrassing gaffes, leading Google to implement a neural net upgrade in 2016. Generative AI has now shaken up the field again, initial results being so promising that some within the industry are speaking of machine translation as “almost a solved problem” (to echo a recent Economist headline).77xMachine Translation Is Almost a Solved Problem,” The Economist, December 11, 2024; https://www.economist.com/science-and-technology/2024/12/11/machine-translation-is-almost-a-solved-problem. Once again, we hear rumors of a forthcoming Babel fish. “Apple Is Turning Its AirPods Into the Babel Fish From Hitchhiker’s Guide to the Galaxy,” the news site Quartz reports.88xEce Yildirim, “Apple Is Turning Its AirPods Into the Babel Fish From Hitchhiker’s Guide to the Galaxy,” Quartz, March 14, 2025; https://qz.com/apple-airpods-translate-ios-19-1851770018. “Meta’s New Translation AI Is Nearly a Babel Fish,” the engineering site IEEE Spectrum declares.99xCharles Q. Choi, “Meta’s New Translation AI Is Nearly a Babel Fish,” IEEE Spectrum, January 15, 2025; https://spectrum.ieee.org/machine-translation Others promise that Star Trek’s universal translators will soon materialize in our palms.
Having now collectively ridden through several AI hype cycles, we know to be wary of such grandiose claims. For one, the world hosts more than seven thousand languages, and even the best of the current tools can handle only a tiny fraction of that total. The new Meta product SEAMLESSM4T (those four Ms standing for “Massively Multilingual & Multimodal Machine Translation”), for example, can take in speech and text from around one hundred languages, but its outputs fall to the thirties depending on what you ask it to do.1010xSeamless Communication et al., “SEAMLESSM4T: Massively Multilingual & Multimodal Machine Translation,” October 25, 2023; https://arxiv.org/pdf/2308.11596. That is still a remarkable achievement, especially given that the system works across modalities—from, say, Dutch text to Romanian speech. But you are out of luck if you wish to generate Serbian speech from Tamil speech or Xhosa text from any source. In thousands of cases, data is too scarce to make contemporary MT techniques feasible.
The “nearly solved problem” is not machine translation simpliciter, then, but the automatic, real-time translation of “high-resource” languages, meaning those for which training data has been generously provided by the Web, such as English and Spanish. To prove this point, researchers wave their systems’ scores on tests with acronyms like BLEU, MQM, and XCOMET in the air and proclaim that the machines are catching up to human translators. But the real test is the average user’s experience, yours and mine. While writing this article, I have run Italian, French, Spanish, Latin, and German passages through multiple machine translators, and while all made mistakes (more on that in a moment), their performance was consistently serviceable—at times, delightful—and undeniably convenient. The Belgian traveler who can’t read an apparently urgent Hindi sign, the Korean scientist writing an article in her third language, the Peruvian parent whose baby monitor comes with Chinese directions, the Saudi business owner arguing with a Norwegian contractor—all these parties, and many others, will find the new and improved MT a godsend.
Nor should we be surprised that machine translators can, in many everyday instances, rival their human counterparts. In 2011, as alarm bells sounded about the old Google Translate, the translator and academic David Bellos wisely noted, “Whatever a language may be in principle, in practice it is used most commonly to say the same things over and over again.”1111xDavid Bellos, Is That a Fish in Your Ear? Translation and the Meaning of Everything (New York, NY: Faber and Faber, 2011), 257. Professional translators are not constantly reinventing the wheel, Bellos stressed; rather, through daily practice and exposure to the conventions of their field, they develop “automatisms” that help them to make quick work of recurring issues. In this respect, human translators are not so different from Google Translate, Bellos argued, “scanning their own memories in double-quick time for the most probable solution to the issue at hand.”1212xIbid. Even human translation has its mechanical side.
Bellos was thus more than happy to cede routine tasks to computers. Doing so, he believed, would free up human translators to concentrate on the many tricky problems that seemed to defy statistical solutions. He imagined that demand for human translation services would grow as businesses and consumers around the globe came to “expect more and more communication between languages” and found MT inadequate on its own.1313x“What’s Lost (And Found) In Machine Translation,” interview with David Bellos; https://bigthink.com/videos/whats-lost-and-found-in-machine-translation/. Yet he also admitted to having worries that Google Translate and tools of its ilk might instigate a cultural shift in which translation would come to be seen as “a task fit only for machines.”1414xInterview with Translator David Bellos,” Gengo, February 2, 2012; https://gengo.com/business-insights/david-bellos/.
Bellos failed to see the enduring attraction of the Babel Fish formula: cheap translation at the speed of the Web. That project is now charging ahead with machine learning, and industry leaders are unsurprisingly eager to reduce their dependence on slow, expensive, break-taking humans, if not remove them from the equation entirely. Vasco Pedro, the outspoken CEO of Unbabel, has forecast exactly that: “It’s hard for me to see right now,” he told CNBC last year, “how three years from now, you will need humans to be translating anything.”1515xArjun Kharpal, “Startup CEO Says Humans Won’t Be Needed for Translation in 3 Years as It Launches AI App,” CNBC, November 13, 2024; https://www.cnbc.com/2024/11/13/unbabel-launches-ai-translation-app-looks-for-fresh-funding.html. Unbabel’s services are representative of the emerging paradigm for fully automated translation: Machines generate text, edit that text, and score the text for accuracy, length, coherence, and even more intangible qualities like fluidity and style. Clients are offered, in turn, not only rapid translation across a broad menu of languages but also “real-time quality scores” of the machine’s outputs. Human translators are pushed to the fringes, serving as a fail-safe mechanism for exceptionally difficult use cases and scorers in the development of gold-standard benchmarks. Several of the new MT companies present human editing as an optional upgrade.
In 2011, Bellos cautioned readers to use Google Translate only “to translate into a language in which you are sure you can recognize nonsense.”1616xBellos, Is That a Fish in Your Ear?, 256. Unbabel’s products are designed to surmount those inhibitions: Thanks to the company’s trusty internal metrics, you can allow the machines to churn away in ninety languages—from Afrikaans to Vietnamese—with the reassurance that line graphs and pie charts provide.
Minerva’s Mystery
Unsurprisingly, I have reservations about the drive to push humans out of the business of translation. In the first place, you don’t need to wade very deeply into the technical literature to grasp that, while the new MT has surpassed the old benchmarks by Olympic leaps and bounds, the machines are far from the “super polyglots” that Mr. Pedro has been peddling for smooth sailing in the metaverse. While adept at translating words, sentences, and paragraphs, their performance degrades as assignments lengthen; automatic book-length translation is still out of reach. The machines can forget previous renderings of a repeated term or phrase. They can slip into, and sometimes get mired in, “translationese”—clauses in which in the wrong places sit. They can be “verbose” (a semi-technical term), yielding more words than the prompter needs (say, for a headline) or fabricating additional verbiage beyond what the source text states. They sometimes refuse reasonable assignments. They make mistakes about (or fail to ask for context regarding) sensitive grammatical matters such as gender and person. They still garble idiomatic expressions, metaphors, and wordplay. And, of course, they hallucinate—which is to say, they make stuff up.
Eco provides a useful gauge of MT’s current capabilities and hazards. As expected, all the systems that I tested romped through the puzzles Eco posed to Babel Fish a quarter century ago. In fact, when round-tripping (immaculately) the opening words of Inferno, several of the machines earned bonus points for noting the lines are Dante’s. Yet when I submitted passages from the essays in which Eco related his Babel Fish experiments, results were mixed. The opening sentences of his piece “How to Play Seriously with Altavista” are illustrative. First, Eco’s Italian:
Tornare per la terza volta alle traduzioni via Internet potrebbe sembrare indisponente. Ma questa volta non voglio ironizzare sulla ottusità delle macchine, bensì vedere come sfruttarla a fini creativi. Devo il consiglio a Lee Marshall.1717xEco, “Trionfante ritorno a Babele. Como giocare seriamente con Altavista,” L’Espresso.
Claude 3.7 Sonnet came in dead last, though at a quick first glance you might not see why:
Returning for the third time to translations via Internet might seem indisposing. But this time I don’t want to focus on the senselessness of machines, but rather see how to use it for creative purposes. I gave the advice to Leo [sic] Marshall.
There are two inexcusable blunders here. The first is the choice of “indisposing” for the Italian “indisponente.” In addition to being outdated English, “indisposing” is untrue to the author’s meaning. It suggests a lack of fit, or even illness, whereas Eco’s word choice was meant to anticipate the reader’s chagrin at finding another Eco column on Babel Fish. “Annoying” would work, as would “irritating,” “off-putting,” “exasperating,” and “irksome”—those words being the other machines’ proposals. The second issue is the verb in the last sentence. Claude got the gesture backwards: “devo” means not “I gave” but “I owe.” Now to be fair to Claude’s developers at Anthropic, I should acknowledge that when I tried the same prompt a few days later, the bot translated the verb correctly. Yet the fact that the machine translator can be right one day and wrong another is, as I’m sure you’ll agree, irritating.
The other responses were quite similar, structurally speaking, and while one might nitpick here or there, they all gave clear-enough windows into Eco’s thinking. One of Gemini 2.0 Flash’s outputs stood out for another reason. First, the translation:
To return for the third time to internet translations might seem off-putting. But this time I don’t want to make fun of the machines’ obtuseness, but rather see how to exploit it for creative purposes. I owe the advice to Lee Marshall.
After this, and without my prompting, Gemini proffered notes on key words in the passage, beginning with “indisponente.” “While ‘off-putting’ works,” the bot explained, “it can also mean ‘unpleasant’ or ‘reluctant.’ The idea is that it’s a bit annoying or tedious to keep going back to online translations.” While the second sentence is true, the first is not. “Unpleasant” is in the right neighborhood, although Italian lexicographers stress that indisponente signals a strong reaction that, to my ear, “unpleasant” does not convey. “Reluctant,” however, is simply incorrect. When I followed up to ask Gemini if “reluctant” would be an appropriate translation of indisponente, it replied, “No, ‘reluctant’ is not a direct or common translation of the Italian word,” and added that while a person who is “indisponente (annoying) might cause someone else to be ‘reluctant’ to interact with them, the words themselves describe different qualities or states.” Annoying.
The most troublesome phrase for the machines, though, was among the first that I input—the title of Eco’s column, “La Bustina di Minerva.” The most popular answer was “Minerva’s sachet.” Gemini suggested “Minerva’s little envelope” or “Minerva’s small pouch.” ChatGPT replied, “The Little Minerva’s Packet” or, more naturally, ‘Minerva’s Little Notebook,” and, on a subsequent trial, related, “It refers to the small pouch (or satchel) associated with Minerva, the Roman goddess of wisdom,” and then pointed out, “The phrase was used by the Italian writer and philosopher Umberto Eco as the title of a long-running column in L’Espresso magazine.” Despite all that bonus information, ChatGPT was, like the rest, mistaken.
“Bustina,” mind you, is applied to small envelopes, pouches, and packets. But our author was referring to a specific kind, the bustina di fiammiferi—the matchbook (or matchbox). In his inaugural column, in March 1985, Eco explained that he had chosen the name “not as a reference to the goddess of wisdom, but to the matchbooks that go by that brand name.”1818xEco, “Che Bell’Errore,” L’Espresso, March 31, 1985; https://lespresso.it/c/idee/2016/2/26/umberto-eco-che-bellerrore-ecco-la-sua-prima-storica-bustina-di-minerva/19016. Translation my own. He was recalling the old practice of using the blank inner flap of a matchbook to scribble an idea before it faded, the phone number of a potential love interest, or the title of a book to purchase or bypass. That’s what his column would offer—not the wisdom of the ages but thinking at its earliest stages.
To get the title—“Minerva’s Matchbook”—right, the translator needs to possess that “world knowledge” that Eco, having sampled Babel Fish, newly appreciated in translators of his own species. A translator who dwelled in Eco’s Italy—where people regularly needed a light—might catch the reference immediately. Yet even one remote from Eco’s world can easily glean the necessary information by starting in the obvious place: the debut column. And even that effort isn’t necessary. Googling will solve the mystery in a few clicks. The last two steps—searching the author’s writings or the Web—will happen, though, only if you first admit that Eco’s meaning eludes you, that if translated in the most straightforward fashion the title is odd. Humility here is an asset. ChatGPT went the other way: To support its version of Eco’s title, it invented a backstory, adding an accessory to Minerva’s traditional garb. Irksome.
Faithfulness, Trust—and Maybe a Little Humility?
It will get better, you say, whether in hope, fear, or fatigue. It may. By the time you read this, the machines may have sorted out their confusions about indisponente and devo and gotten a hint about the brand name behind Eco’s column title. On the other hand, due to the growing complexity of these systems, ironing out the current kinks may have ripple effects that give rise to other headaches, mild or severe. Some degree of duplicity and hallucination may just be part of the deal. Umberto Eco may always be a nuisance to translation machines.
Either way, we ought to be wary of the efforts of the proponents of MT automation to define translation as a purely technical activity, one whose success or failure can be unfailingly measured in real time, at all times, by metrics determined and administered by company machines. Translation is a technical activity, both in the more rarefied sense of dealing with specialized domains (e.g., scientific papers) and the basic one of depending on tekhnē, craft. But that craft is, and has always been, employed to mediate between those who otherwise would be reduced to pantomime or frustrated silence. Translation is thus, as translators themselves have been admitting with growing confidence in recent decades, a profoundly ethical pursuit.
That was what Eco was trying to get at with his notion of translation as “negotiation.” Translation does not happen in a vacuum; it is a social act that takes place at a given time and in a given milieu. It occurs on behalf of multiple parties—the author, who otherwise wouldn’t be read in the target language; the reader, who otherwise wouldn’t have access to the author’s questioning, imagining, reasoning, fact-finding, forecasting, declaiming, etc.; the publisher, who commissions the work and hopes to turn a buck; and perhaps to “culture” or “posterity” in some vague or certain sense. Accordingly, Eco portrayed the translator as the party who seeks a compromise between the several claimants in the exchange.
“World knowledge” helps. For only by having a thick understanding of the author, the author’s culture, the source language, the target language, the conventions of the genre in question, and the intended audience can the translator make wise linguistic choices. The process hinges, though, on faithfulness—not as a patented method but as a manner of approach. Faithfulness, on Eco’s telling, consists of faith in the possibility of translation in the first place, a wager on meaning despite obvious impediments. It extends to the translator’s interpretive excavations to grasp the text’s “deep sense” (which becomes the basis for subsequent decisions to be made about style, tone, rhythm, sentence length, word order, punctuation, and so on). It finds expression in “the goodwill that prods us to negotiate the best solution for every line.” “Among the synonyms of faithfulness,” Eco argued in 2003, “the word exactitude does not exist. Instead, there is loyalty, devotion, allegiance, piety.”1919xEco, Mouse or Rat?, 192.
Eco’s bar is, unquestionably, high. Others would place it lower, or stress other dimensions of translation. Even so, his account is valuable now because it brings to light the fact that the translator, being answerable to multiple parties, faces a series of ethical challenges as she works word by word, sentence by sentence, through her assignment.
Meanwhile, one is hard-pressed to find any mention of faithfulness, devotion, allegiance, or piety in the reports issuing daily on MT breakthroughs, though synonyms of exactitude abound. In industry and academic papers on MT, the “ethics” section (if there is one) often consists of nothing more than the details of how the team sought to tamp down bias, especially gender bias, and root out offensive language. Developers, for example, have been alarmed to note added “toxicity” in their outputs—meaning the machines sometimes introduce toxic language even though the original contained none. The favored solution is, of course, to build better diagnostics and add more safety protocols. Eco endeavored to renew faithfulness; Big Tech offers us “automatic toxicity detection.”
But the campaign to automate translation has its weightiest consequence not in the machine’s sphere but the user’s. As the translator Mark Polizzotti has observed, the “stale” Italian pun traduttore, traditore (“translator, traitor”) conveys a cultural truth: In most situations, we’d prefer not to trust the middleman.2020xMark Polizzotti, Sympathy for the Traitor: A Translation Manifesto (Cambridge, MA: MIT Press, 2018), 12. Translation is, for the audience, a vulnerable state: We must put our trust in a stranger who stands between us and the source that we cannot hear or read for ourselves. We must (to echo the second epigraph above) “take another person’s word for it.” Receiving a translation is an act of faith. Translation lays bare the grating fact of our finitude.
How much more appealing is the empowerment promised by the new MT systems. Now you are the controls. You can have your translation right now, tailored to your specifications. Which style would you prefer, DeepL Translate asks users—Simple? Business? Academic? Casual? What about tone? Friendly, enthusiastic, confident, or diplomatic? Or perhaps you’d like the text rendered as rhyming couplets? ChatGPT stands ready. And we have already seen a preview, thanks to Gemini, of the machines’ ability to supplement their renditions with commentary. With these tools at our fingertips and plugged into our ears, we may at last enjoy the linguistic freedom of our favorite sci-fi characters. We may go anywhere and never be lost for words.
Of course, as things currently stand, the new Babel Fish may prove riskier companions than today’s boisterous headlines admit. At our cultural moment in which a single ill-timed word or misplaced character can ignite a career-ending social-media storm, I suspect that some hard lessons await the most vociferous early adopters, especially those who pay no heed to Bellos’s warning against straying into languages in which you can’t spot rubbish (or excrement). We may find, moreover, that the ability to pontificate at any time, in scores of languages, creates new problems as words get out that we might, on second thought, rather not have shared or heard. More communication is not necessarily better communication. (Such was Douglas Adams’s lesson: “The poor Babel fish, by effectively removing all barriers to communication between different races and cultures, has caused more and bloodier wars than anything else in the history of creation.”2121xDouglas Adams, The Hitchhiker’s Guide to the Galaxy (London, England: Pan Books, 1979), 61.)
The old, slower-paced translation, meanwhile, offers those willing to attend a salutary disquiet. Where Babel Fish would render humanity’s diverse articulations on the user’s handpicked terms, the old translation reminds us that the whole world cannot be accounted for using the satchel of words that we ordinarily bear about with us. The old translation exposes the limitations of our language, even while gesturing toward the possibilities of others through shrewd incorporations. Translation, at its most human, tugs familiar words into foreign realms of meaning, renewing and extending them, and inviting us to follow there that we might sample the local flavors for ourselves. The developers of machine translation have achieved remarkable technical feats, and surely more are to come; their labors will release a torrent of words, and many of them will be good ones. But we ought to be apprehensive about what the Babel Fish whispers in our ears. Automatic machine translation is being marketed as a means to expand our little worlds. It may just as easily render the world back to us on ever more narrow terms.