Why Translation Must Have Fingerprints

Christianity is, and has always been, a faith in translation.

Our Scriptures were given across three languages: Hebrew, Aramaic, and Greek. In that sense, the Bible itself is a book that constantly moves its readers between languages. Christianity’s spread throughout the world has meant constant translation from one language to another. In private conversations, on papyrus and scrolls, in sermons, songs, and heavy tomes, Christianity has never been hemmed in by any particular language. Rather, as Scripture says, our God will be served by all languages just as he will be worshiped by all peoples (Dan. 7:14).

The Gambian theologian Lamin Sanneh put it well:

Christianity is a translated religion without a revealed language. . . . Without translation there would be no Christianity. Translation is the church’s birthmark as well as its missionary benchmark.

What should we think about that ancient birthmark and historic benchmark in our own day, where the task of theological translation is increasingly outsourced to AI?

Three Camps on AI Translation

This week, Mere Orthodoxy launched The Faith Received (TFR), a vast online repository of theological texts. For the most part, those texts are in English as a result of AI translation, although the project invites human readers to correct mistakes by comparing the AI rendering to the source texts.

The immediate online response has been deeply polarizing—notably so, given that translation sits so deep in Protestantism’s bones.

For some, TFR is an obvious win: People can read obscure texts (for free), most of which were previously only in other languages. Even if you know no Latin or Syriac, the idea goes, you can now read Voetius or Bardesanes for yourself. In that view, AI translation is seen as a great and efficient equalizer of knowledge.

Others express a more cautious pragmatism: AI has good and bad sides, so let’s be careful to check translations for accuracy and not forget that human-led translation is important too.

Another camp has more profound reservations and sees something like TFR more as a curse than a blessing.

For full disclosure, I’m (mournfully) in the third camp. I appreciate and admire Mere Orthodoxy. Like those who run it, I want to see Christians gain wider access to historic theological literature. I also believe in the importance of translation. I’ve invested much of my adult life in learning Dutch to make Herman and J. H. Bavinck available to English speakers. For years, I’ve trained PhD students to do the same. And yet I find it hard to muster enthusiasm for a project like TFR, which I cannot help but see as one step forward, and many more steps back, for theology at large in the English-speaking world.

The project’s most enthusiastic instant online adopters are also, unfortunately, sometimes the least imaginative in asking why someone like me might see this project as concerning for theology. To some online commenters, we’re haters, dinosaurs running scared from our own tech-apocalypse, Luddites, doomers, and so on. As I read from one vocal supporter on X, our lack of enthusiasm is “pathetic.”

It’s easy to imagine why we might look like this if AI translation is imagined in simple terms as a quantum leap forward in both efficiency and equality. If you can read English and have an internet connection, you can now read Voetius, who has finally been freed from his ivory-tower-Latin captivity.

Like the more cautious pragmatists, of course, this group acknowledges that AI translation can make mistakes. Both groups offer the same solution to that problem: Readers can compare the AI translation to the original text and check for themselves if the AI’s rendering is “accurate.” So far, so democratic.

Something about it also seamlessly carries along a deep, historic impulse within Christianity, and not least of the Reformed sort: Give as many people as possible access to theology in a language they understand.

Alongside that continuity, however, I fear that AI translation changes something significant in the historic nature of Christianity as a “religion in translation.” The shift from Christianity as a human-translated religion to an AI-translated religion is far from seamless and shouldn’t be overlooked glibly. The nature of that change, though, is harder to imagine for some than for others.

Defining Translation

What is “translation”? The answer depends on which language you speak. Translate “translation” into different languages, and you might well find its meaning shifts.

My colleague Hephzibah Israel has written on how differently it works in English and in Tamil: “Derived from Latin (translatio), ‘to translate’ in English implies ‘to carry across’ or transfer from one language to another. The Tamil verb ‘peyar’ instead conveys: to dislocate, to uproot, or to unseat with force.”

The shift from Christianity as a human-translated religion to an AI-translated religion is far from seamless and shouldn’t be overlooked glibly.

At that basic level, English conditions us to imagine translation as a relatively straightforward process. In a steady pair of hands, a word, a line, a paragraph, or a book can easily be moved from one language to another, its meaning never wobbling for a second. (To English-speaking AI enthusiasts, AI translation represents a hand breathtaking in steadiness and speed.)

In Tamil, the whole process is more dramatic—a disassembly of meaning in one language, and its reassembly in another. That notion of translation prompts questions of how the uprooted, reconstructed meaning relates to the original. Phrased in that way, translation is an attempt to understand what something means in one language and an attempt to reconstruct that meaning in another. Unlike English, Tamil foregrounds the idea that translation is interpretation, rather than just relocation.

At that basic level of vocabulary, English gives us a suggestive prompt that translation is easy and that ideas across languages are only superficially different. Alongside this, anglophone culture treats monolingualism as the norm. For many of us, in any meaningful sense, English is all we know, our sole frame of reference in imagining other languages. That the same idea might also be somehow different in another language is often difficult for us to imagine—like the abstract thought of a far-off planet where the laws of gravity are different, or where people walk on their hands and see with their ears.

To people who live and think in more than one language, though, this is the stuff of life under the sun: Languages are meaningfully different, and moving from one to another is ordinary and complex all at once.

What’s Lost in Translation

When I talk to monolingual English speakers, I often find they imagine translation like converting a recipe from one measurement system to another. Recipe A uses cups, whereas recipe B uses milliliters. Translation is taken to be the straightforward, mechanical task of following the same recipe in either system. The recipe and end result are the same, and the difference in cooking either way is superficial.

To multilingual people, translation is a different kind of experience. It’s more like starting with a recipe rooted in one culture, with its local ingredients, culinary traditions, methods, and tastes, and trying to create a faithful representation of it in a different culture, where some of the original ingredients are unavailable, and where locals have different techniques and tastes (or allergies). As the cook (or translator), your attempt to represent the original dish faithfully invariably entails choices that change the dish somehow.

Often, those choices are straightforward and the changes minimal: Recipe A uses sugar but is being translated into a culture that uses honey. In either case, the dish stays sweet, even if the taste subtly differs. Sometimes, the choices are necessarily more complicated: Recipe A needs tamarind, a local ingredient that simultaneously tastes sweet and sour, which I cannot locally source for recipe B. However, I can more or less replicate that effect by combining two local ingredients: lemon juice and sugar.

Sometimes, recipe A has an ingredient that cannot easily be replicated with local substitutes—like saffron. No combination of substitutes will match its color, blend of floral aroma, and earthy taste. A blend of turmeric and sweet paprika can take its place, representing what was there in the original, but it isn’t the same. As the chef/translator, to be truthful I should probably let you know that in this case, there wasn’t a direct way to move from one recipe to the other.

What does this mean for languages and the translation of theological texts? To use Dutch-to-English as an example, some translation is indeed simple: Moving from God is goed to God is good is more like an English sense of translation-as-relocation than a Tamil sense of translation-as-dislocation.

Sometimes, translation is more challenging and involves subtler interpretation on the translator’s part. For example, ik lees contains two possible English meanings: I read and I am reading. In English, these are different. Knowing which to use is often a source of frustration to Dutch speakers when speaking English. In translating ik lees, English forces me to make an interpretative choice, as I cannot carry over the ambiguity in its original form.

As is true of languages at large, Dutch has terms with no direct English equivalent—the most famous being the feeling of gezelligheid. Like a chef swapping tamarind for a combination of lemon juice and sugar, English doesn’t prevent me from telling you about gezelligheid, or even helping you experience it. It’s the warm feeling you get when in cozy places with people you like.

As a translator, however, I sometimes find that Dutch presents me with its own saffron. This content requires a Tamil sense of translation as total deconstruction and imaginative reconstruction. Consider the Calvinistic poetry of Willem Bilderdijk.

Geleerden! ’t is vergeefsch, met klemmend taalvoldingen
Het vrede- en zoenverbond den wijsgeer op te dringen.
’t Is de overriding niet van ’t wikkend onverstand,
Waaraan de Vredegod Zijn heilkracht heeft verpand.

What does it mean to “translate” this into English? The stanza is an elegant appeal to believing scholars not to think they can coax unbelieving philosophers into Christian faith by arid arguments alone. To write this, Bilderdijk combines unusual vocabulary and rhyming couplets in a way that doesn’t carry over into English: taalvoldingen and dringen reinforce each other through rhyme, whereas the English equivalents, “verbal persuasions” and “to impose,” do not.

Whatever it would mean to translate this text, I cannot simply carry over its stylistic composition and its conceptual meaning. Both are integral to the original. At my hand, by my choice, one of them must be sacrificed.

I cannot simply carry over its stylistic composition and its conceptual meaning. Both are integral to the original. One of them must be sacrificed.

When I do so, what should I call my rendering? Bilderdijk in English? An attempt to represent Bilderdijk in English? Or, in effect, have I written an entirely new English poem prompted by reading Bilderdijk in Dutch?

If someone else produced an alternative English version, making different choices about vocabulary or style, which version would be the “accurate” one? In any case, as the translator, if I were to meet an English speaker who read my translation and told me he or she had “now read Bilderdijk and love his poetry,” I would have to say, “No, you haven’t, but thanks for enjoying my interpretation of it.”

What Happens When AI’s Translation Becomes Ours

At its best, theological writing aims to tell the truth about God, and all else in the light of God. The translation of that material is sometimes simple, often complex, and sometimes almost impossible.

In every case, translation remains an act of interpretation. Pre-AI, this was the practice by which the faith was received—person to person, language to language, sometimes by smooth relocation, and sometimes by more difficult (and often error-strewn) passage.

That pursuit of truth means the recognition of many things: that all theological translation is interpretation, that the translation and the source text are different texts, and that somewhere between the two, the text was covered in the translator’s fingerprints. A way of imagining theological translation that ignores this is linguistically naive. More seriously, it pushes that use of theological texts toward ambivalence to the pursuit of truth.

Once, I was teaching an undergraduate seminar on the German idealist philosopher Hegel. In the discussion, a brilliant German student pointed out that reading Hegel in English was easier and more challenging than reading him in German. In the original, she said, Hegel is often deliberately ambiguous, whereas our English translator tended to resolve those ambiguities and give his writing a more concrete feel. To the student sitting beside her, a monolingual Brit, this was a lightbulb moment—a pointer to a reality that hadn’t struck him yet. He was reading an interpretation of Hegel, but Hegel himself was something subtly different. In a small way, their interaction challenged the British student’s framework in considering what it meant to read a text, and to do so as someone who wanted to seek truth.

I have no doubt that some will use TFR like that German student. Appeals to “check the translation alongside the original” are meaningful if you can read the original with some skill. If you can read the original as a richly textured unit of meaning, you might well interact with it fruitfully. Truth might well be pursued.

For many English speakers in our monolingual English world, unfortunately, the invitation to compare AI translation to source text will be less edifying. If you want to read the AI translation of Bardesanes, a third-century Syriac theologian, and are curious about how AI has interpreted his original text, you need to know Syriac. (If you can read Syriac, good for you!) For the rest of us, we’re left to take the AI rendering at face value.

As we do so, perhaps, we’ll relay errors and confuse the AI’s translation biases for whatever Bardesanes himself wrote down. As we do so, the AI enterprise subtly nudges us into a murky world. Some corner of your mind now has information that prompts you to say, “You know, I once read in Bardesanes of Edessa that . . .” However, if you can’t read Syriac, and if no one who can takes the time to check AI Bardesanes, the reality is you haven’t actually read Bardesanes. Whatever you do with your AI translation–derived knowledge, it’s a lose-lose situation.

Imagine you’re a preacher, and in your sermon you say, “I have this great quote from Bardesanes . . . although it’s currently unverified information from an AI translation,” your honesty will deflate the moment. Use the quote without attributing it to unverified AI, and you might inspire your congregation, but you yourself are being shaped by that practice, with all its underlying assumptions about truth.

Use the quote without attributing it to unverified AI, and you might inspire your congregation, but you yourself are being shaped by that practice, with all its underlying assumptions about truth.

Many users of TFR, I expect, will have a more disciplined approach than that. Using the available tools to view the side-by-side presentation of AI translation and source text, they’ll remember that the English may be an uncorrected AI translation, thus shifting the onus onto readers to verify for themselves whether the English is a good translation.

Unfortunately, while that setup reflects well on the integrity of the project’s designers, its effectiveness is constrained by anglophone culture at large. In that culture, we fail abysmally at teaching languages other than English. From the get-go, those languages are presented as inherently “foreign,” rather than things that might become familiar. Few of us learn them in ways that make them intuitive or imaginative to us.

For most of us—like for Augustine reflecting on why he hated learning Greek as a child in Confessions—they fail to take root in our minds. Rather, even among the highly educated, anglophone culture often produces people who interact with other languages in a monolingual way.

Many of us imagine the English and foreign-language texts side by side like a recipe in cups and a recipe in milliliters. We only know how to judge the task of translation by tracing a line from each familiar English word to the foreign-looking equivalent that (presumably) should mean and do the same thing in the other language—like the cup of milk that I expect to be matched by however many millilitres of milk in the foreign recipe.

In that mindset, what does it mean to “check a translation for accuracy”? It often means identifying that the English text has the word “God,” so we check if the Latin line also has “Deus,” and so we hover back and forth from translation to original, imagining the English text atomistically and confirming to ourselves whether the foreign words seem to map onto this English grid.

In a way, this is a kind of truth seeking, and it may sometimes bear fruit. Here and there, perhaps, we’ll catch the AI misreading a foreign script and giving us the wrong English word. But it’s truth seeking of a blinkered sort, and it’s particular to people formed in English-speaking cultures in our time.

Old Problem on a New Scale

TFR’s most vocal online supporters have been quick to describe it as an epochal moment for our reading of theological sources. Given Christianity’s long history of dealing with translation, I view it more cautiously. The problems it creates aren’t entirely new.

Between the 12th and 14th centuries, the medieval translation industry circulated a vast amount of Latin translation of Greek and Arabic texts, often erroneous, that was received and circulated uncritically by readers who knew no Greek or Arabic. AI translation amplifies that practice, perhaps, but didn’t create it.

In the Middle Ages, a fruit of that industry was a general laziness toward the study of languages by otherwise serious Christian thinkers—a detrimental cultural norm criticized, notably, by the English polymath Roger Bacon, who traced out its effects on the quality of biblical exegesis, scholarship at large, evangelism on the frontiers of Europe, and much more.

In our day, the AI translation industry is our direct equivalent. Making the project of theological retrieval dependent on it is no small development in the long history of our “translated religion.”