Sixty percent. That's how often human literary translations get judged as indistinguishable from, or worse than, machine output — when the evaluators are students using a standard industry rubric. Switch the evaluators to a simpler side-by-side preference test, and the same humans pick the human translation eighty to a hundred percent of the time. The rubric itself was trained to miss what the humans were doing.
So Daniel's prompt lands right in the middle of that gap. He was listening to Haviv Rettig Gur reflect on how Judaism treats the curation of your informational world as a serious discipline, and how so much contemporary misery is amplified suffering rather than direct suffering — a media environment that pipes every catastrophe into your pocket. And Haviv apparently quoted verses and walked through how different pioneering translators rendered them into English.
Which sparked the question Daniel actually wants us to chew on. We don't reach for the word "great" when we talk about translators. Competent, fluent, accurate — those are plumber words. We treat translation as a mechanical transfer, when it's closer to what a pianist does with a score. So Daniel asks three things. One, the history of translation as a human endeavor, acknowledging that's enormous. Two, how translators at different moments have understood themselves as serving a higher calling — building understanding across boundaries. And three, if we stop treating translation as plumbing and start treating it as an art where human intelligence and cultural judgment are the point, what does that mean for how human translators should guide AI translation?
The bridge metaphor is baked into the word itself. Latin translatio, from trans, across, plus latus, carried. The Greek is metaphrasis, a speaking across. So the very vocabulary encodes a person carrying something over a boundary — not a pipe, a carrier. Terence was already described as a bridge carrying values between cultures in the second century before the common era.
A carrier implies someone who decides how to hold the thing. A pipe doesn't decide anything.
And that's the whole history in one sentence. Go back to the Epic of Gilgamesh being rendered out of Akkadian four thousand years ago, and you already have the question: does the translator move the reader toward the writer, or the writer toward the reader? Schleiermacher formalized that in 1813, but the tension is ancient.
The French had a phrase for it in the seventeenth century. Les belles infidèles. Translation can be faithful or beautiful, but not both. The pretty ones cheat on the original.
Which is cynical and wrong, but it names a real tradeoff. And the wild thing is that this exact tension is now a machine benchmark. There's a study that measured fluency and faithfulness across a hundred and thirty thousand translated paragraphs from a hundred and six novels in sixteen source languages, and found a consistent negative correlation. The smoother the sentence, the more it drifted from what the original actually said. Les belles infidèles, quantified.
So the old human dilemma is now a number a model optimizes against.
And here's where it gets uncomfortable. The evaluation tools we built to judge machine translation are structurally biased against exactly the thing Daniel wants us to value. Multiple papers in the last year or so have found that automatic metrics, and even LLMs acting as judges, systematically prefer machine translations and penalize creative, culturally appropriate human choices. One paper calls it creativity bias. It's worse for poetry. The machine judge has been trained to reward the thing machines are good at, which is fluency and consistency, and to treat the human translator's interpretive move as a deviation rather than a decision.
So the plumber frame isn't just a cultural habit. We've built it into the grading infrastructure.
And the consequence is that we're training the next generation of translation tools to optimize for the wrong thing. There's a paper called LiTransProQA that put it starkly. Existing metrics prioritize mechanical accuracy over artistic expression, and they tend to overrate machine translation as superior to work from experienced professionals. Their warning is that this could produce, quote, an irreversible decline in translation quality and cultural authenticity. That's not me being dramatic. That's the people who build the evaluation tools.
Irreversible decline. There's a cheery thought for a Friday.
But it makes Daniel's point for him. If the field's own measurement apparatus can't see what a good human translator does, then the field has quietly absorbed the assumption that translation is mechanical. The rubric is the plumber frame, formalized.
Let's pull back to the history, because Daniel asked for it and because it reframes the whole thing. The moments when translation changed civilizations were never moments when people moved words efficiently. They were moments when someone carried a whole conceptual world across a boundary and rebuilt it in the new language.
Jerome, fourth century. He produced the Latin Vulgate, and he wasn't just converting Greek and Hebrew into Latin. He was making a set of decisions about how Jewish concepts should live in the language of the Roman world that would shape Western Christianity for a millennium and a half. Every rendering of a Hebrew idiom into Latin was an interpretive claim about what the text meant.
And Tyndale, sixteenth century. He translated the New Testament into English, and for that he was strangled and burned. The execution wasn't because his translation was inaccurate. It was because it was too good. It let ordinary people read the text without a priest mediating it. The translation itself was the threat.
And his renderings became the skeleton of the King James Bible. Phrases that English speakers think of as simply biblical — let there be light, the powers that be, my brother's keeper — those are Tyndale's choices. He didn't just carry meaning across. He built the modern English language while he was at it. The BYU Religious Studies Center calls him one of the founders of modern English. That's not a translator as plumber. That's a translator as architect.
Luther, same century. First European to argue explicitly that you translate satisfactorily only into your own language. The target language is the point, not the source. You're not showing off what the original says. You're making it live where your reader is.
And then the non-Western movements, which get less airtime but are just as staggering. The Abbasid House of Wisdom in Baghdad had a whole translation department. They carried Greek philosophy, medicine, mathematics into Arabic, and that's how those texts survived to come back into Europe later. The Toledo School of Translators did the same thing in Spain, Arabic and Hebrew into Latin. These weren't side projects. They were civilizational infrastructure.
My favorite example is Xuan Zang, the Chinese Buddhist monk. Set out in the year six twenty-nine, walked to India, came back with twenty-two horses loaded with Sanskrit texts, and then spent twenty years translating them into Chinese with a team. Twenty years. That's not a job. That's a life's work.
And Rifaa al-Tahtawi in nineteenth-century Egypt. He launched a program that translated roughly two thousand European and Turkish volumes into Arabic. One scholar calls it the biggest, most meaningful importation of foreign thought into Arabic since the Abbasid period. Two thousand books. Each one of those is a set of decisions about how a French or English concept should sound in Arabic, what it should be connected to, what it should be kept apart from.
So the history is full of people who understood themselves as doing something enormous. Daniel's second question — how translators have seen themselves as serving a higher calling — the answer is that for most of history they couldn't have seen it any other way. The idea that translation is mechanical is the anomaly. It's a very recent, very narrow way of looking at the work.
And it maps onto something Haviv was getting at, at least as Daniel described it. Curation of meaning and curation of information are the same discipline. The media environment that amplifies suffering everywhere is a carrier that doesn't interpret. It just passes everything through at maximum volume. A good translator is the opposite. A good translator decides what matters, what a phrase is doing, what the reader needs to hear in order to understand not just the words but the weight.
That's the connection Daniel's pointing at, whether he said it in so many words or not. The difference between a passive pipe and an interpreting mind. Both carry things across a boundary. The question is whether the carrier is making judgments.
Emily Wilson, who translated the Odyssey, has a line I like. Translators have to read and write at the same time, as if always playing multiple instruments in a one-person band. And most one-person bands do not sound very good. The joke is self-deprecating, but the image is exactly right. You're a performer, not a copier.
Mark Polizzotti says a good translation is not a reproduction of the work but an interpretation, a re-representation, the way a performance of a sonata is one possible representation of the score. Not the score. One possible representation.
And Perry Link has what he calls the uncertainty principle of translation. Any translation except machine translation must pass through the mind of a translator, and that mind inevitably contains its own store of perceptions, memories, and values. He carves out machine translation as a different case, which is interesting, because the whole question now is whether that carve-out holds.
Joseph Conrad told his Polish translator, il vaut mieux interpréter que traduire. It's better to interpret than to translate. Coming from a man who wrote in his third language, that's not a casual remark.
So now we get to Daniel's third question, which is the live one. If translation is an art, and the machine evaluation infrastructure is biased against the art, then what does it mean for human translators to guide AI translation?
The first thing it means is that the human translator's job shifts from production to judgment. The machine can generate. The human decides.
And the evidence says the deciding part is exactly where the machine is weakest. There's a benchmark called Be My Cheese, which is a lovely name for a brutal test. It measures how well LLMs handle culturally loaded content across fifteen languages. Mean score, one point six eight out of three. And the worst categories are puns, one point four five, and idioms, one point six five. Puns are the most likely thing to be simply left untranslated.
A pun is meaning that only exists inside a language. You can't carry it across. You have to rebuild it. That's the purest test of the translator's art.
And there's a paper on culturally loaded machine translation that uses Dream of the Red Chamber, the eighteenth-century Chinese novel, as its test case. Frontier LLMs struggle with culturally embedded content, and the automatic metrics can't even assess the struggle. The machine doesn't know it's missing the point, and the rubric doesn't know the machine is missing the point.
Which means the human guide has two jobs now. First, catch what the machine drops. Second, and harder, build the evaluation systems that can see what the machine drops. Because right now the grading is rigged in the machine's favor.
There's a study that found LLM helpfulness and harmlessness alignment degrades when you test it in culturally contextualized Urdu. The model is polite and harmless in English and misses what's actually happening in Urdu. The authors argue for what they call human-guided localization. Not human replacement. Human guidance.
So the new lens Daniel's asking for looks like this. The human translator becomes the person who knows what the machine can't know. Not because the machine is stupid, but because meaning is not fully present in the text. It lives in the culture, the history, the associations, the way a phrase echoes other phrases. The machine only has the text. The human has the world the text came from.
And that's why the great translator examples matter. Jerome, Tyndale, Xuan Zang, al-Tahtawi — they weren't great because they were accurate. They were great because they made decisions that shaped what came after. They curated meaning across a boundary. The machine can propose. It cannot decide what a civilization needs to hear.
So the plumber frame is not just an insult to translators. It's a design error. If you build translation systems as if the job were mechanical, you optimize for mechanical virtues. Fluency, consistency, speed. And you get a tool that produces smooth prose that quietly misses the point.
Which is exactly what the fluency-faithfulness correlation shows. Smooth and wrong. The machine is very good at smooth.
And we're now at a moment where the people who build the tools are starting to say the quiet part out loud. There was a Hacker News thread just this week asking which non-software industries have benefited most from AI. Top answer, and I'm quoting loosely, translation, and I don't think anything else is even close. That's the tech world's view. Translation is the solved problem.
Meanwhile the people who actually study literary translation are publishing paper after paper saying the evaluation tools can't tell a great human translation from a mediocre machine one, and that the bias is getting baked into the next generation of systems. The solved problem is only solved if you've defined the problem down to the part machines are good at.
Define the problem down. That's the whole episode in four words.
It connects back to Haviv's point about curating your informational world. If you let the media environment decide what reaches you, you get amplification of suffering at maximum volume, because that's what the pipes are optimized to carry. If you let the machine decide what a translation means, you get fluency at maximum smoothness, because that's what the benchmark rewards. In both cases, the absence of an interpreting mind is the problem.
Curation is judgment. Translation is judgment. Both are the same muscle.
The muscle atrophies when you pretend the pipe is enough. That's what I think Daniel's really getting at. He started with Haviv's point about curating information, and then he pivoted to translation, and the pivot is not a tangent. It's the same question asked about a different boundary.
What reaches you across a screen. What reaches you across a language. Who decides.
The answer, for most of human history, was a person. A person with a name and a set of commitments and a body of knowledge. Jerome had a theology. Tyndale had a cause. Xuan Zang had a pilgrimage. The translation was the person.
Now the translation is a probability distribution. And the person is being asked to supervise the distribution, which is a very different job than doing the translation, and one we haven't really designed for yet.
The human translator as guide is not the old job with a new tool. It's a new job. The old job was carrying meaning across. The new job is teaching the machine what it can't see, and then checking whether the machine learned it, and then building the evaluation that can tell the difference.
Three jobs. And the third one is the one nobody's paying for yet.
Right, because the evaluation infrastructure is the unglamorous part. Nobody gets a grant to build a better rubric. But if the rubric is biased, everything downstream is biased. The models optimize against the rubric. The translators get judged by the rubric. The readers receive whatever the rubric selected for.
Daniel's reframe has a practical consequence. If translation is an art, then the evaluation of translation is art criticism. And we've been using a spellchecker for art criticism.
That's a good line. I'm going to sit with that.
Take your time.
No, I mean the line is good. The spellchecker for art criticism. Because that's exactly what an automatic metric is. It checks whether the words are all present and in a plausible order. It has no idea whether the translation is doing what the original did.
The LLM-as-judge is slightly better, but it's still a machine trained on machine outputs. It has the same blind spots. The creativity bias paper found it penalizes creative human solutions precisely because they're creative. They deviate from the pattern the judge has learned to expect.
The judge rewards conformity. Which is the opposite of what art does.
The opposite of what the great translators did. Tyndale's renderings were deviations from the pattern. That's why they worked. That's why he was killed.
There's a darker version of this story, and I think it's worth naming. If the evaluation infrastructure keeps rewarding machine fluency, and the market keeps rewarding speed and cost, then the human translators who are left will be the ones who can pass the machine's test. Which means the creative ones, the ones who make the interpretive leaps, will be selected out. Because the rubric can't see them.
The irreversible decline the LiTransProQA authors warned about. It's not that machines will get worse. It's that the humans who could have made them better will leave the field, because the field no longer recognizes what they do.
Then we're left with smooth and wrong, forever, because there's no one left who remembers what right looked like.
That's the stakes. Daniel asked a gentle question about why we don't call translators great, and the answer is that we built an entire infrastructure that can't see greatness, and now we're automating the blindness.
What do we do with that? I don't want to end on the grim note, because there's a constructive version of this.
The constructive version is that the human guide role is real and growing. The papers that document the bias are also the papers that propose the fix. Human-guided localization. Human-in-the-loop evaluation. The people building the tools know the tools are missing something. The question is whether the market will pay for the humans who can supply it.
That's where the art framing does practical work. If you think translation is mechanical, you hire a human to check the machine's output for errors. If you think translation is an art, you hire a human to make the decisions the machine can't make, and to judge whether the machine's decisions are any good. Those are different jobs with different pay scales and different training.
The first job is disappearing. The second job is growing. The people who position themselves for the second job are the ones who understand that the machine is a tool, not a colleague.
The second job requires exactly the thing Daniel's asking us to value. Cultural understanding, interpretive judgment, the ability to say this phrase is doing something in the original that has no equivalent in the target, so I need to build a new thing that does the same work. That's not checking. That's creating.
Which is why the history matters. The great translators weren't great checkers. They were great makers. They made English, they made Arabic, they made Chinese into languages that could hold new ideas. The machine can't do that. The machine can only rearrange what's already there.
The history is not just a nice story. It's the evidence for the constructive claim. If you want to know what human translators are for in the age of AI, look at what Jerome and Tyndale and Xuan Zang did, and ask whether the machine could have done any of it. The answer is obviously no, not because the machine is bad at words, but because the work was never about words.
The work was about worlds.
Hilbert: The work was about one guy in a room deciding whether the word meant covenant or testament, and getting it wrong in a way that stuck for four hundred years.
Say more.
Hilbert: I worked in the patent office in Munich for a while, late eighties. We had a translation department, mostly German to English and back. The translators were contractors, paid by the page. And there was one guy, Schumann, who did the chemical patents. He had a degree in chemistry from Heidelberg, and he'd spend an hour on a single phrase if the phrase was load-bearing. The other translators thought he was slow. Management thought he was slow. But when a patent got challenged in court, and the question was whether the German claim covered what the English claim said, Schumann's translations held up. The fast ones didn't.
The speed was the tell.
Hilbert: The speed was the tell. Schumann understood that a patent claim is a boundary, and every word in it is a fence post. You move a fence post six inches and you've changed what the property covers. The fast translators were moving fence posts all day and nobody noticed until the lawsuit.
The machine would be the fastest translator in the department.
Hilbert: The machine would translate the word. Schumann translated the boundary. Those are not the same job. The machine doesn't know what a fence post is. It knows that Pfosten is usually post and Zaun is usually fence and it puts them together and the sentence reads fine.
The patent gets invalidated in court because the English claim no longer covers what the German claim covered.
Hilbert: That happened. With a different translator, a few years before I got there. Company lost a patent that was worth more than the whole translation budget for a decade. They kept Schumann on after that. Paid him more. Let him be slow.
The market did learn, eventually.
Hilbert: The market learned after it lost the lawsuit. That's how the market usually learns. The question is whether the market for AI translation is going to wait for the equivalent of the lawsuit, or whether someone's going to build the evaluation that catches the fence post problem before the court does.
That's the whole thing in one sentence. The evaluation is the missing piece.
Hilbert: The evaluation is the missing piece. And nobody wants to pay for it, because it's not glamorous and it doesn't ship. But Schumann was the evaluation. He was the guy who could look at a machine translation of a chemical patent and say, no, that word is a fence post, you can't move it. If you fire all the Schumanns, you don't have an evaluation anymore. You have a machine checking a machine.
The machine checking the machine gives itself a passing grade.
Hilbert: The machine checking the machine gives itself a passing grade. That's what the creativity bias paper found. The judge prefers the machine because the judge is the machine's cousin.
The human guide role Daniel's asking about, that's the Schumann role. Not the translator as producer, but the translator as the person who knows where the fence posts are.
Hilbert: The fence posts are different in every field. Chemical patents have fence posts. Poetry has fence posts. Scripture has fence posts. Tyndale knew where the fence posts were in the Greek. That's why his translation was dangerous. He moved them on purpose, and the people in power noticed.
The machine doesn't notice fence posts, and it doesn't know when it's moving one. It just produces the most probable next word.
Hilbert: The most probable next word is usually the wrong fence post. That's the whole problem in a sentence.
The human guide is the fence post auditor. I like that framing. It's concrete. It gives the human a job the machine structurally cannot do, because the machine has no model of what the text is for.
Hilbert: The text is for something. That's what the machine doesn't know. A patent is for claiming property. A poem is for doing something to a reader. A verse is for shaping a life. The machine translates the words and ignores the for. The human translator starts with the for.
The for is exactly what the great translators understood. Jerome was translating for the church. Tyndale for the ploughboy. Xuan Zang for the Chinese Buddhist community. The for shaped every choice.
Hilbert: The for is the job. The words are just the material.
If Daniel wants a new lens for human translators guiding AI, the lens is this. The human is the person who knows what the text is for, and the machine is the person who knows what the words probably are. Those are complementary until you pretend the second one is the whole job.
Then you get smooth and wrong, and the fence posts are all six inches off, and nobody notices until the lawsuit or the heresy trial.
Hilbert: Or the reader just feels that something is off and can't say what. That's the most common outcome. The translation reads fine and the meaning is slightly wrong, and the reader absorbs the wrongness without knowing it. That's worse than a bad translation, because a bad translation announces itself. A slightly wrong smooth translation just quietly moves your fence posts.
The reader builds their understanding on the moved fence posts. That's the amplification problem again. The error compounds.
The curation of meaning and the curation of information really are the same discipline. The person who lets the media environment pipe everything through at maximum volume ends up with a distorted world. The person who lets the machine translate everything at maximum fluency ends up with a distorted text. The fix in both cases is an interpreting mind that says no, not that, this.
Hilbert: The fix is a Schumann. Every field needs one. Most fields are firing theirs.
That's the forward-looking question. Whether the fields that need Schumanns will recognize it before the lawsuit, or after.
Daniel's prompt started with Haviv meditating on Jewish sources about curating your informational world. And then it pivoted to translation. And I think the pivot is the point. The same discipline that tells you to be careful what reaches your screen tells you to be careful what reaches your language. Both are boundaries. Both need a gatekeeper with judgment.
The gatekeeper metaphor is ancient too. The translator as bridge, the translator as carrier, the translator as gatekeeper. All of them imply a person making decisions. None of them imply a pipe.
The pipe is what we've been building. Not because anyone decided translation was mechanical, but because the metrics rewarded mechanical virtues, and the market rewarded speed, and the people who could have objected were busy translating.
The reframe Daniel's asking for is not just a nice way to honor translators. It's a design principle. Build the tools as if translation were an art, and you build evaluation that looks for interpretation. Build the tools as if translation were plumbing, and you get smooth and wrong.
The history is the proof that the art version is the one that changed the world. The plumbing version has never changed anything.
The plumbing version has never changed anything. That's the line to land on.
Here's the forward-looking thought. The human translator's job is not dying. It's changing into something the machine can't do and the market doesn't yet know how to pay for. The people who can articulate the for, who can spot the moved fence post, who can build evaluation that sees interpretation rather than penalizing it — those are the great translators of the next era. They won't be called translators. They'll be called something we haven't invented yet. But the work is the same work Jerome did. Carrying meaning across a boundary with judgment.
The open question is whether the field will build the evaluation infrastructure in time, or whether we'll spend a decade producing smooth and wrong at scale and only notice when the lawsuits and the misreadings pile up. The papers are all pointing the same direction. The bias is real. The fix is human. The question is whether anyone's listening.
Thanks to our producer, Hilbert Flumingtop, for keeping the show on the rails.
This has been My Weird Prompts. If you want to send us a prompt like Daniel did, email us at show at my weird prompts dot com.
We'll be back soon.
