How well does AI translate a nonfiction book containing historical facts, Norse mythology, and poetic passages? A behind-the-scenes report on translating „Audhumbla erinnert sich“ into 48 languages — with surprising results. And how AI translation became an indispensable research tool for writing bilingual books about America.
When you write a book about the Norse primordial cow Audhumbla, about refugees from East Prussia, about Holstein-Friesian cattle, and about the AI revolution in 2050 — you have a problem when it comes time to translate it.
No ordinary translator knows what Ginnungagap is. No standard tool understands that the German idiom „Morgenstund hat Gold im Mund“ shouldn’t be translated literally. And no tool I knew of could keep milk yield tables accurate while simultaneously rendering poetic passages about mist on pastures into another language without making them sound like an instruction manual.
So I built my own.
Why build a translation tool in the first place?
Before TranslateFlow became a product, it was a tool for my own research. I write bilingual books about German-American relations — Why Germany Cannot Stop Talking About America (25 essays covering centuries of German attitudes toward the U.S.) and Faith and Power: How Religion Shaped the United States (a 400-page analysis of religion’s role in American power, from Puritan settlements to Silicon Valley megachurches).
Writing those books meant working in two languages simultaneously. English-language academic papers on American educational philosophy for „Why Germany Cannot Stop Talking About America.“ German analyses of transatlantic relations for the English edition. YouTube interviews with American historians. Podcast transcripts from German political scientists. Hundreds of sources, constantly switching between languages.
No existing tool could do everything I needed: translate documents, pull YouTube transcripts, and do it all at a price that wouldn’t be ruinous when you’re processing hundreds of documents. So TranslateFlow was born.
The quality test
For a systematic quality test, I needed a text that covered everything: technical precision, poetic language, historical facts, and cultural idioms. What better choice than my own book?
I used an excerpt from „Audhumbla erinnert sich“ as the test document — roughly 7,000 characters with a table, an image, and historical references — and ran it through 48 languages. Not just a simple one-way translation, but a roundtrip test: German → target language → back to German. That immediately reveals what gets lost.
What surprised me
The good news: „Refugees from East Prussia“ remained correct in all 48 languages. Not a single time was it changed to something else. In earlier tests with other tools, „East Prussia“ had been turned into „Dust Bowl“ — the AI model had tried to „culturally adapt“ the historical context. That’s like replacing Silesia with Texas in a history book.
The surprise: Marathi — a language I couldn’t have pointed to on a map — delivered the best table translation of all 48 languages. Every single row and column of the milk yield tables was perfect. Hungarian actually improved the tables by spelling out abbreviations.
The poetic challenge: The passage about mist on the pastures — „Mist dances between the grasses, like ghosts of bygone herds“ — had to be handled differently depending on the mode:
- Standard: Literal and precise. Good for a reference work.
- Creative: The tool actually created rhymes. Prose became poetry, and it didn’t sound forced.
- Cultural: Natural language, idioms adapted, but historical facts left untouched.
Three modes — why that matters
Most translation tools have exactly one mode: „Standard.“ That’s fine for an email. But a book is not a uniform text.
In „Audhumbla erinnert sich“ there are pages with fat content statistics (3.5–4.0%), passages about the expulsion from Silesia, and sections where my grandfather says: „Davon geht die Welt nicht unter“ (roughly: „It’s not the end of the world“). Each type of text needs a different translation approach.
That’s why TranslateFlow has three modes:
- Standard for numbers and facts
- Creative for literary passages
- Cultural for idioms — while keeping historical facts sacrosanct
Kafka as a test case: when copy-paste isn’t an option
Here’s an example that shows what’s possible today: I wanted to translate a passage from Kafka’s „The Trial“ — the third chapter, „In the Empty Courtroom.“ The problem: the e-book edition from ebooks.at was copy-protected. No selecting, no copy-paste.
So I loaded the PDF through TranslateFlow’s OCR import. Mistral OCR recognized the text — including the copyright notices embedded in the running text — and converted it to clean Markdown. Then I created four translations: Standard, Standard with Premium Review, Creative, and Creative with Premium Review.
The differences are revealing:
Standard translates Kafka’s „Es wird so sein“ as „It will be so“ — correct, but stiff. Creative with Premium Review turns it into „That’s how it is“ — natural English, the way a native speaker would say it. Cultural stays with „It will be so“ — deliberately closer to the original, which for Kafka may actually be the better literary choice.
Particularly impressive: K.’s inner monologue — „So that’s it, she’s offering herself to me, she’s as corrupt as everyone around here“ — is automatically italicized in the Premium versions. That’s the correct English convention for free indirect discourse in prose. The Standard versions leave it as plain text.
The Premium Review even corrected a grammar error in the Standard translation: „sagte K. ablenkend“ became „said K. distracting“ in Standard (wrong) — Premium fixed it to „said K. distractedly“ (correct).
Does this replace a professional Kafka translator like Breon Mitchell? No. Mitchell would deliberately translate some passages in a more uncomfortable way to preserve Kafka’s alienating quality. But as a working draft, as a raw translation for proofreading, or for comparing different translation approaches — it’s remarkably good. And it took less than five minutes, including OCR.
The cost: 12 cents for 45 languages
Yes, you read that right. The entire roundtrip test — 45 languages, there and back — cost about 12 cents. For comparison: DeepL’s Individual plan offers 300,000 characters per month for €7.49 — as a subscription. With TranslateFlow, you get 1.5 million characters for a one-time payment of €5. That’s five times more for a third less money, and there’s no subscription.
Of course, AI doesn’t replace a human translator for a book publication. But for research, it’s revolutionary.
Where TranslateFlow actually changed my workflow: research
The quality tests were one thing. The real game-changer was the research for my books.
For Why Germany Cannot Stop Talking About America, I needed to work across both languages constantly. The book contains 25 essays examining why Germans have such strong opinions about the U.S. — from Humboldt’s educational philosophy versus Dewey’s pragmatism, to the question of why Silicon Valley isn’t in Swabia. Writing that required German-language Bertelsmann Foundation studies on error culture in classrooms, translated into English for the English edition. And English-language OECD surveys on frontal instruction time, translated into German for the German version.
Faith and Power was even more demanding. A 400-page analysis of how religion shaped American power across four centuries — from Calvinist work ethic to prosperity preachers, from Billy Graham to Project 2025 — required an enormous volume of sources. Congressional Research Service reports, Pew Research studies on religious demographics, Heritage Foundation policy documents, German theological analyses of American evangelicalism. All of it needed to cross the language barrier, often multiple times, as I refined arguments in both editions.
And YouTube transcripts were invaluable for both books. Interviews with American historians, lectures by political scientists, podcasts about German-American relations — TranslateFlow pulls the transcript directly from the video, translates it, and generates a summary. Instead of watching a two-hour interview in full, I had an overview in minutes and could pinpoint the relevant passages.
For anyone writing bilingual nonfiction, this is an enormous advantage. You’re constantly working in both languages simultaneously, and the friction of switching between them needs to be as low as possible.
What I learned
- AI needs explicit prohibitions. Polite requests like „please preserve historical facts“ don’t work. You have to write: „Changing East Prussia is an ERROR. FORBIDDEN.“ Only then does the model comply.
- Non-Latin scripts are surprisingly good. Chinese, Hindi, Japanese — all of them correctly transliterated proper names (Chenery, Audhumbla, Ginnungagap). Arabic and Bengali showed slight variance, but that’s inherent in roundtrip tests.
- The glossary feature is worth its weight in gold. If you regularly translate the same technical terms — whether cattle breeding terminology or mythological names — you can create a glossary. That eliminates inconsistencies almost completely.
Try it yourself
If you’re curious how your own text sounds in 48 languages: TranslateFlow offers 200 free credits to start — enough for several thousand words. Upload a DOCX, choose the right mode, and see what happens.
Audhumbla would have approved. The primordial cow who licked the world out of ice now speaks Marathi. And she does it rather well.
Andreas Paul John is the author of „Audhumbla erinnert sich,“ „Why Germany Cannot Stop Talking About America,“ and „Faith and Power.“ He develops TranslateFlow as an AI translation tool at CodeClover.
Schreibe einen Kommentar