I have been sitting with a single sentence from Merethe Lindstrøm’s Dager i stillhetens historie (Days in the History of Silence) for the better part of a morning. The sentence, in Norwegian, reads: Hun satt ved vinduet og så ut, men det var ikke noe der, bare stillheten som la seg over hagen som et lag med noe som ikke kunne sees. In Anne Bruce’s English translation, it becomes: “She sat by the window and looked out, but there was nothing there, only the silence that settled over the garden like a layer of something that could not be seen.” I have written it out twice now, once in each language, and the difference between them is not a matter of vocabulary or syntax. It is a matter of what the Norwegian knows and the English must learn: how to let silence be a substance, not an absence.
This is the kind of problem that occupies me as a translator and critic. It is also the kind of problem that, I suspect, a machine trained on vast corpora of bilingual text would handle with a certain competence. Feed the Norwegian into an AI story generator that promises coherent prose from a prompt, and it might produce something very like Bruce’s version—perhaps even identical. The word order is straightforward. The vocabulary is unremarkable. The grammar is compliant. And yet the sentence, in its English form, trembles with decisions that no algorithm made. Bruce chose “settled” over “lay,” “layer” over “film” or “veil,” “something that could not be seen” over “an invisible thing.” Each choice is a small ethical act, a negotiation between the source language’s metaphysics and the target language’s habits. The machine, scanning millions of aligned sentences, would recognize the pattern but not the pressure behind it. It would not feel the weight of stillheten as a presence, because presence is not a statistical property.
I want to be precise about what I mean by “machine” here. I am not speaking of some hypothetical superintelligence, but of the class of tools now readily available: large language models fine-tuned for narrative generation. These systems are, at their core, pattern-matchers of extraordinary sophistication. They have ingested terabytes of text and learned to predict, with startling accuracy, which word is likely to follow which. They can produce paragraphs that scan as fluent, stories that satisfy genre expectations, dialogue that sounds plausible. What they cannot do is hear the silence between words, because silence is not in the training data. It is in the body of the translator, who has lived with a language long enough to know what it withholds.
Lindstrøm’s novel is, in many ways, a book about withholding. It concerns an elderly couple, Eva and Simon, who have sealed off whole chambers of their shared history. The narrative moves through their present-day routines—meals, walks, visits from a home help worker—while the past leaks in through fissures they have tried to caulk. The prose is quiet, declarative, almost flat. And yet it is precisely this flatness that makes the book so difficult to translate. Norwegian can be a plainspoken language, but its plainness carries a different cultural resonance than English plainness. When Lindstrøm writes a sentence like De snakket ikke om det—“They did not talk about it”—the Norwegian reader hears not just a fact but a whole social architecture of avoidance, a specifically Nordic relationship to confrontation and disclosure. The English reader, encountering “They did not talk about it,” may hear something closer to a temporary omission, a topic deferred. The translator must decide whether to amplify, to add a word like “ever” or “anymore,” or to trust that the accumulation of such sentences will teach the English reader a new grammar of silence. Bruce, for the most part, trusts. This trust is a craft decision that no AI, however well-trained, can replicate, because it arises from a human assessment of what a reader can bear and what a text can survive.
The Authors Guild, in its AI Best Practices for Authors, articulates a principle that has stayed with me: “AI outputs, by contrast, are generic mashups of pre-existing works ingested during training. When you claim authorship in a work, it means you contribute your unique view and thoughts and your unique voice.” The document is, rightly, concerned with the ethics of using AI in one’s own writing practice—the danger of ceding voice to a statistical average, the theft embedded in the training data. But I think the principle extends further, into the act of translation. A translation is not merely a transfer of semantic content; it is an authorship of a second original. The translator contributes a unique view and a unique voice, shaped by a specific history of reading, a specific ear for rhythm, a specific tolerance for ambiguity. When I read Bruce’s Lindstrøm, I am reading Bruce as much as Lindstrøm, in the way that a performance of a Bach cello suite is the cellist’s as much as Bach’s. The AI, by contrast, can only produce a composite of every translator it has ever ingested, smoothed to the most probable rendering. It cannot have a view. It cannot have an ear. It cannot, in any meaningful sense, have read the book.
Let me return to the sentence. Hun satt ved vinduet og så ut, men det var ikke noe der, bare stillheten som la seg over hagen som et lag med noe som ikke kunne sees. I have been turning it over not only as a translator but as a reader who has lived in Helsinki for fifteen years, who knows what it is to sit by a window in November and look out at a garden that offers nothing to the eye but a kind of thickened gray. The silence Lindstrøm describes is not metaphorical. It is meteorological, almost geological. It has weight and extension. It legger seg—lays itself down—like snow, like dust, like the particular gloom of a Nordic winter afternoon when the sun sets at three. The English “settled” does some of this work, but it also domesticates the image, makes it cozier. A machine translation engine, optimizing for fluency, might choose “settled” for exactly that reason: it collocates well with “silence” in English. But the translator who chooses “settled” does so knowing what is lost—the more active, more bodily verb la seg, which suggests an agency in the silence, a will to cover. The machine does not know what is lost. It only knows what is gained in probability.
This is not a Luddite’s lament. I am not arguing that AI has no place in the literary ecosystem. There are tasks—terminology management, consistency checking, the drafting of routine correspondence—where pattern-matching is genuinely useful. What I am arguing is that literary translation belongs to a category of human activity that is defined by its irreducibility to pattern. The creative writing resources assembled by Purdue University’s Online Writing Lab, in their introduction to creative writing, emphasize foundational skills like voice, style, and the development of individual thinking. These are not mystical properties; they are teachable, discussable, refinable. But they are also fundamentally human. They arise from a person’s accumulated experience of language as lived, not language as data. A student in a creative writing workshop learns to hear the difference between a sentence that merely communicates and a sentence that breathes. A translator learns the same thing, across a wider and more treacherous gap.
The gap is what interests me. In Lindstrøm’s novel, there is a scene where Eva, the wife, recalls a moment from decades earlier, when she stood in a doorway and watched her husband Simon with their infant daughter. The memory is presented without commentary, without emotional notation. The prose does not tell us what Eva felt; it tells us where she stood, what the light was like, how long she remained there. The silence in the scene is not a gap in the narrative; it is the narrative. To translate such a passage, a translator must resist the urge to explain, to psychologize, to add the kind of emotional signposting that English-language literary fiction often expects. This resistance is an ethical act. It is a decision to preserve the foreignness of the text, to let the English reader experience the same disorientation the Norwegian reader experiences. An AI, trained on a corpus that overwhelmingly favors explicitness over reticence, would likely nudge the passage toward legibility. It would add a “she felt” or a “she remembered” or a “she realized.” It would, in other words, correct the silence. And in correcting it, it would destroy the book.
I want to be fair. I have experimented with AI translation tools, not as a replacement for my own work but as a way of understanding what they see. I have fed them paragraphs from Dag Solstad, from Tove Jansson, from Kerstin Ekman. The results are often surprisingly competent. They get the denotation right. They rarely make grammatical errors. They even, occasionally, produce a phrase that I might have arrived at myself. But they never produce a phrase that makes me stop and think, I would not have thought of that, and it is better than what I would have thought of. They never surprise me. And surprise, in translation as in all writing, is the signature of a mind at work. It is the evidence that someone has not merely processed language but has listened to it, has heard something in the original that the rest of us missed and found a way to carry it across.
Consider the word stillheten again. Norwegian forms abstract nouns with the suffix -het, much as English uses -ness. But stillhet is not quite “stillness.” “Stillness” in English leans toward the absence of motion; stillhet leans toward the absence of sound, a quiet that is almost a substance. The English “silence” is closer, but it lacks the morphological relationship to stille (quiet, still) that a Norwegian reader feels. The translator cannot reproduce that relationship. She can only compensate elsewhere, by building a prose texture in which silence accumulates weight through repetition, through the careful placement of sentences that end without resolution, through a rhythm that teaches the reader to pause. This is not a local problem with a local solution. It is a global problem that requires a global sensitivity. The AI, which processes text in windows of tokens, has no global sensitivity. It has only local predictions, stitched together by an algorithm that does not know what a novel is.
I am aware that this essay risks sounding elegiac, as if I were composing a eulogy for human translation before it has died. That is not my intention. Human translation is not dying. It is, if anything, more vital than ever, because the forces that threaten it also clarify its value. When a machine can produce a plausible version of any sentence, the question shifts from “Can it be translated?” to “What kind of translation matters?” The answer, I think, is the kind that bears the marks of a specific human encounter with a specific text. The kind that makes choices a machine would not make because it cannot see the choices. The kind that understands silence not as empty space but as a material to be shaped.
I close the book. Outside my window in Helsinki, the November garden is doing what Nordic gardens do: withdrawing into gray, offering nothing to the eye but the bare branches of a rowan tree and the dark earth where the perennials have died back. The silence here is not the silence of Lindstrøm’s sentence, but it is related. It is a silence that has weight, that legger seg over the city like a layer of something that cannot be seen. I sit with it for a while. Then I open the book again, to a different page, and begin the work of listening.