How we do it
How a book is translated, and one book told whole
The five steps
A book is translated in five steps:
- Before translation begins, a professional linguist who knows the book's subject and both languages writes a detailed brief.
More ▾
The brief sets out the book's specific terminology, its names, the phrases that must stay as they are, the author's voice and tone, and the traps between the two languages. - AI translates the book according to the brief.
- AI then revises the whole translation against the original and the brief, and marks every passage it can't confirm is right.
- The linguist revises the marked passages against the original, together with a random selection of the rest, and corrects what needs correcting.
- You receive the translated book with an Assurance Report, signed by the linguist, that shows what was revised and what was changed.
Did a person read every word?
No. The linguist revises the passages that the AI marks as uncertain, and a random selection of the rest. The Assurance Report states how many passages the linguist revised and how many were changed. Revising only where it's needed is what lowers the cost and the time. If you want a linguist to revise the whole book as well, we can add that.
How do you choose the linguist?
We look for professional linguists qualified to translate the whole book on their own, the traditional way, who know its subject and both languages. Before working with us, each one sits a blind sample: a passage of a real book, revised under the conditions of the work. We score it against the original, read their reasons, and rank the candidates. The linguist placed on a book is paid by the hour, on a timer that becomes part of the record.
How much should a person revise?
Three, and they differ in how much of the book a linguist revises, which is most of the cost. AI-checked: the AI translates to the brief and revises its own work in several independent passes; no linguist revises the text, and the report says so. It's for samples, reading copies and deciding which books to translate. Assured, our standard: the linguist revises every passage the AI marks as uncertain, and a random selection of the rest, and signs the Assurance Report. Fully revised: the linguist also revises the whole book against the original. At every level, a professional linguist who knows the subject writes the brief before translation, and the report says exactly what was done. The price rises with the linguist's share.
AI-checked
- What a person does
- The AI translates according to the brief, then revises its own work in several independent passes. No linguist revises the text, and the report says so.
- What the report says
- "No linguist revised this translation", and the checks it went through.
- Who it's for
- For samples, reading copies, and deciding which books to translate fully.
Assured
- What a person does
- The linguist revises every passage the AI marks as uncertain, and a random selection of the rest, and signs the Assurance Report.
- What the report says
- The counts: marked, revised (with the sample), changed. "The linguist didn't revise every word."
- Who it's for
- Our standard, for most non-fiction.
Fully revised
- What a person does
- On top of that, a linguist revises the whole book against the original.
- What the report says
- "A linguist revised the whole translation."
- Who it's for
- For scripture, law, medicine, and any book a publisher's name rests on.
A report on a translation you already have
- What a person does
- A linguist revises a random selection and the passages our AI marks as uncertain, and signs a report on whether it can be trusted.
- What the report says
- Whether it can be trusted, and where not.
- Who it's for
- For partners and publishers holding a translation made elsewhere.
Price follows the level. Ask us for a quote.
Our first book, in seven chapters
Chapter 1
The book
- French into English
- about 20,000 words
- seven units
Our first book was a work of theology, French into English, about 20,000 words. Translated the usual way, the English would have cost more than it was ever likely to earn. So, like most books, it would have stayed in French.
Chapter 2
Before a word
Three entries from the brief
- A term
- témoin → "witness" for a person; "textual witness" for a manuscript
- A name
- Pères → "the Church Fathers"
- A trap
- annonce → "proclamation", not "announcement", when it is the Gospel being proclaimed
A linguist who knows the subject and both languages sat with us before anything was translated. She settled 89 terms, the names, the Bible the author quotes, and the traps between French and English. That's the brief. The AI translates according to it.
Chapter 3
The first reader
Seven units, translated 27 Sep to 3 Oct
53 passages marked as uncertain in units three to seven
Unit one: read in full by the linguist
Why a passage is marked
The AI couldn't confirm its rendering against the French and the brief, so it says so instead of guessing.
The AI translated all seven units in a week. Then it read its own work against the French and the brief, and marked every passage it couldn't confirm: in units three to seven, 53 of them. It doesn't guess. It says where it isn't sure.
Chapter 5
The stranger
The clay birds
- What the English said: That the child Jesus killed birds and brought them back to life.
- What the old text says: He made birds of clay and gave them life.
- What happened: The author agreed, and the correction is on the list for his French edition.
- Eight quotations from the Church Fathers and Josephus, retranslated from the French, restored to their published English.
- Bible verses that ran past, or stopped short of, where the author had cut them.
Then we did something new for this book. Two AI models from two other developers, OpenAI's GPT and Google's Gemini, read the whole book fresh, once against the French and once as English alone. They were not shown what our AI had marked, or what the linguist had changed. They found 116 places worth a look. Most were small. One was not.
Chapter 6
The person decides
Nothing changes because a machine said so. The linguist ruled on what touched the meaning: she typed changes in 12 of the 85 paragraphs she was shown in six of the seven units, and the second reader's 12 questions for this book went to her page. The author ruled on what touched his book. The rest, we ruled on or declined, and the record says why.
Chapter 7
The record
- Brief written before translation: 89 terms, and the names
- Passages the AI marked as uncertain: 53 in units three to seven; the author answered 52
- Revised by the linguist, with a random selection: Unit one in full and a random selection: about 44% of the book
- Changed by the linguist: 12 of 85 paragraphs
- A second reader, for this book: 116 places
- Signed by the linguist at delivery
This book's counts. The book is not yet delivered, so its report is not yet signed.
All of it goes into the Assurance Report: the brief, how many passages were marked, how many the linguist revised and how many she changed, what the second reader found for this book, and her signature. It says plainly that she didn't revise every word. When a translation is questioned, the report is the answer.
Who caught what
This book was read five times. No reader caught everything. Here is who caught what. Every book's report lists its readers, with its own counts.
The brief
What it did: Written with the linguist before translation: 89 terms, the names, the Bible the author quotes.
What it caught in this book: The terminology and the names were settled before a word was translated.
The AI's own revision
What it did: Read the whole translation against the French and the brief.
What it caught in this book: Marked 53 passages it couldn't confirm, in units three to seven.
The author
What it did: Answered the passages the AI marked as uncertain.
What it caught in this book: 52 of 53, the same evening.
The linguist
What it did: Read the first unit in full and a random selection of the rest, at least one passage in every unit, including a random share of the passages the author answered; 13.76 hours.
What it caught in this book: Typed changes in 12 of 85 paragraphs. Most of her corrections were punctuation, word choice and grammar: of 13 of that kind, the two outside models found 2.
The second reader (Two AI models from other developers, for this book)
What it did: Read the whole book twice, fresh, without being shown what our AI had marked or what the linguist had changed.
What it caught in this book: The clay birds; 8 quotations restored to their published English; Bible verses cut where the author cut them; and, unprompted, all 11 places we'd changed on purpose.
The machines read everything. The person reads where judgement is needed. The report says which readers your book had.
Ten papers into Brazilian Portuguese
Our ten papers into Brazilian Portuguese, signed by their linguist. A professional linguist settled the terminology with us before translation, then revised a random selection of passages without being told which ones the AI had marked as uncertain, and one paper in full.
See a real Assurance ReportThe second reader
It ran once, for that book. It isn't yet a standing step, and every book's report says which readers it had.
The clay birds, in chapter 5Why we do this
Translation is abundant now. Trust isn't.
Human time is what guarantees the quality, and it's also most of the cost. As the models improve, we reduce the human share where the evidence allows, book by book and use case by use case, but a linguist always writes the brief, and the report says exactly what was done. We pass those savings on, to partners and to publishers. We think cheaper, trustworthy translation will mean far more books translated, not fewer jobs: abundant translation, with trust.
Will this take work away from professional linguists?
We expect the opposite. When a valuable service becomes much cheaper, people use much more of it; economists call this the Jevons paradox. Today most books are never translated, because translation costs more than the book is expected to earn. When the cost falls to a fraction, many more books can be translated, and each one needs a professional linguist who knows its subject. Our expectation is that linguists with subject expertise will have more work, not less.