What Translators Can Teach Us About AI
How a semiotician's insights on translation reveal the limitations of our most powerful language models
When ChatGPT effortlessly converts a marketing brief from English to Japanese or Google Translate instantly renders a German news article in flawless Spanish, we witness something that feels like magic. It seems that the machines have mastered the language. They've conquered Babel.
But these technological marvels obscure a profound truth that the late Italian philosopher Umberto Eco spent decades unpacking: translation is not a mechanical act of substitution. It's a deeply human cultural negotiation process that requires judgment, creativity, and a willingness to betray literal meaning to preserve deeper truth.
I recently revisited Eco's Experiences in Translation, a collection of lectures from 1998 that feels startlingly prescient in our AI-saturated world. What surprised me wasn't just the Italian semiotician's keen insights about the art of translation, but how perfectly his framework explains the limitations of even our most sophisticated language models. The distance between what AI can do and what human translators actually do illuminates something essential about meaning itself and why certain forms of human judgment remain irreplaceable.
The Joke That Couldn't Be Translated
Consider this scene from Eco's novel Foucault's Pendulum. Three characters are bantering, and one makes a joke involving the Italian word lettera, which means both "letter" (a mailed document) and "epistle" (as in the biblical letters of Saint Paul). The punchline depends entirely on readers catching both meanings simultaneously.
When the novel was translated into English, Eco and his translator William Weaver faced a problem: English lacks this dual meaning. "Letter" and "epistle" are distinct words. A literal translation would render the joke incomprehensible.
Their solution was radical by conventional standards. Weaver completely rearranged the dialogue, redistributing lines between characters and crafting an entirely new punchline that preserved the spirit of the joke while abandoning its literal structure. Eco notes, "The English text says that Casaubon said q while the Italian text says that Casaubon said p."
Yet Eco approved this seemingly unfaithful rendition because it accomplished something more important than word-for-word accuracy: it preserved the novel's narrative effect. Readers in both languages would experience the same flash of humor, the same sense of the characters' personalities. The surface was sacrificed to preserve the depth.
This is precisely what today's AI systems struggle with. They excel at substitutional accuracy—finding the statistically most probable equivalent for each term. But they falter when meaning exists not in the words but in the cultural, intertextual, or emotional space between them. Even when prompted to be creative, they rarely make the kind of interpretive leap that Weaver did—to completely reimagine a passage while preserving its essence.
Translation as Cultural Adaptation
The challenges multiply when we move beyond wordplay to cultural references. Eco describes another passage from his work where a character mentions seeing endless horizons "beyond the hedge." Italian readers immediately recognize this as an allusion to Giacomo Leopardi's famous poem "L'infinito"—a touchstone of Italian Romanticism that evokes specific emotional and philosophical associations.
But English readers would see just a hedge—an unremarkable shrub with no literary resonance. The solution? Weaver replaced the hedge with a reference to "Darien," evoking Keats's sonnet "On First Looking into Chapman's Homer" instead. The hedge would have the same emotional register, but it would have an entirely different cultural marker.
This is cultural translation at its most sophisticated: it moves not just between languages but between entire systems of cultural reference.
I was curious how ChatGPT would handle this kind of challenge, so I asked it to translate a simple American cultural reference—"It's like waiting for the ball to drop on New Year's Eve"—for audiences in India, Japan, and Brazil, while preserving both the temporal meaning and emotional anticipation.
Its answers were grammatically flawless but culturally tone-deaf. For India, it suggested: "It's like waiting for the diyas to be lit during Diwali"—technically correct in referencing a festive waiting period, but missing the specific New Year's context and the sense of countdown to a precise moment. The AI was thinking categorically (festival → festival) rather than experientially.
What would a human translator do? They might refer to waiting for the first bell at Jama Masjid to announce the end of Ramadan fasting or the countdown to the cricket ball being bowled in the opening of the India-Pakistan match—references that capture both the suspenseful waiting and the cultural weight of the original.
Where AI Excels—And Where It Falters
Modern language models are marvels of pattern recognition. They can reliably produce grammatical sentences, maintain consistency across thousands of words, and adopt different stylistic registers. Given enough examples, they can mimic the voice of Hemingway or recreate the cadence of legal documents.
But Eco's framework helps us see precisely what they cannot (yet) do: make meaningful interpretive bets. Good translators, Eco argues, must sometimes gamble on the author's intent, choosing cultural resonance over lexical fidelity. He writes:
"Interpreting means making a bet on the sense of a text, among other things. This sense that a translator must find—and preserve or recreate—is not hidden in any pure language...It is just the outcome of an interpretative inference that other readers can or cannot share."
This betting—this willingness to make a bold interpretive choice—requires a form of cultural and contextual intelligence that remains beyond AI's reach. It demands pattern recognition and pattern imagination: the ability to intuit what will resonate with a particular audience in a specific context.
Consider the marketing slogan "Just Do It." A literal translation preserves the words but likely loses Nike's punchy, rebellious ethos. A human translator might lengthen or completely reconceive the slogan to capture its cultural force in a new context. In Russian, it might become more like "Challenge Yourself" to avoid associations with impulsiveness (which carries negative connotations in Russian culture).
When I asked an AI to translate this slogan for Russian audiences, it offered grammatically correct options. Still, it couldn't articulate why certain cultural adaptations might be necessary—or make a confident bet on which adaptation would work best.
Beyond Words: The Substance of Expression
Eco's insights extend beyond word choice to what he calls "the substance of expression"—the material quality of language. Translation isn't just about semantic meaning; it's about rhythm, sound, and the physical experience of language.
He offers the example of Poe's poem "The Raven" with its hypnotic refrain "nevermore"—a word Poe chose specifically for its somber, drawn-out vowel sound and the resonance of that final "r." Translating "nevermore" into French as "jamais plus" or Portuguese as "nunca mais" loses the sonic quality that gives the poem much of its power.
This matters beyond poetry. Think of political slogans like "Yes We Can" or "Make America Great Again"—their effectiveness lies partly in their rhythm and sound. Or consider user experience design: the difference between "click here" and "tap here" isn't just semantic but experiential, reflecting different physical interactions.
Current AI models can be instructed to consider rhythm and sound, but they struggle to feel the embodied quality of language—the way certain sounds resonate in the chest or linger on the tongue. They process language as information rather than as embodied experience.
The Translator's Mind in an AI World
Eco's framework doesn't just help us understand translation—it illuminates what remains distinctly human in an age of increasingly capable language AI.
First, there's the matter of cultural fluency. True cultural understanding isn't just knowing facts about another culture; it's inhabiting a different way of seeing. It knows which jokes land, which references resonate, and which metaphors illuminate. AI can simulate this knowledge but lacks the lived experience that gives cultural references their emotional weight.
Second, there's interpretive courage. Good translators—like good leaders, designers, and strategists—must sometimes break rules to preserve deeper principles. They must be willing to seem unfaithful to the surface to remain faithful to the depths. AI systems, designed to minimize error rather than maximize meaning, are inherently conservative in their interpretive choices.
Finally, there's contextual judgment. The right translation depends not just on the text itself but on its context: who will read it, why, and in what circumstances. AI can be given contextual information but struggles to weigh competing contextual factors without explicit guidance.
These qualities—cultural fluency, interpretive courage, and contextual judgment—define good translators and effective humans in many domains. They represent not just skills but a way of thinking—a mode of intelligence that values meaning over accuracy.
Learning the Translator's Mindset
Eco's insights offer a roadmap for developing this translator's mindset, whether you're actually translating languages or simply navigating different cultural, professional, or disciplinary contexts:
Focus on effects, not just equivalence. Ask not "What does this say?" but "What does this do?" A direct translation that fails to produce the same effect is no translation at all.
Recognize when to sacrifice fidelity. Sometimes, preserving the letter of the text means betraying its spirit. Know when to prioritize the more profound truth over surface accuracy.
Be willing to make interpretive bets. Translation requires confidence in your interpretation, even when multiple readings are possible. You can make choices based on your best understanding of context and intent.
Attend to the substance, not just the form. Consider how language feels—its rhythm, texture, and sound—not just what it denotes.
These principles apply far beyond traditional translation. They're valuable for anyone who communicates across contexts: the product manager explaining technical concepts to non-technical stakeholders; the therapist helping patients reframe their experiences; the leader articulating a vision across different organizational cultures.
The Future of Translation in an AI World
What does this mean for the future of translation—both the profession and the broader human activity?
AI will increasingly handle what we might call "substitutional translation"—the conversion of straightforward content where literal accuracy is sufficient. Technical manuals, financial reports, and other primarily denotative texts will be translated by machines with increasing accuracy.
But "interpretive translation"—where cultural resonance, emotional impact, and contextual judgment matter more than word-for-word accuracy—will remain a human domain. Literature, marketing, diplomacy, and other fields requiring cultural sensitivity will still need human translators, though they'll likely use AI as a collaborator rather than a replacement.
The translator's mindset—the ability to move fluidly between contexts while preserving meaning—will become increasingly valuable in a world where AI handles routine communication. However, the uniquely human ability to make interpretive leaps, understand cultural nuance, and preserve meaning across contexts will remain essential.
Umberto Eco never witnessed the AI revolution, but his insights anticipated its limitations. Translation isn't just about words; it's about worlds—the cultural, emotional, and experiential realities that words evoke. Until AI can truly inhabit those worlds rather than simply modeling them statistically, the translator's mind will remain one of our most valuable human capacities.