The translator cliff
What the decline of professional translation tells us about designing AI translation for books.
AI translation is changing the economics of professional work—and exposing a crucial design choice about where human attention belongs.
A recent Financial Times article by Sarah O’Connor and John Burn-Murdoch has a rather uncomfortable title: “How AI has de-skilled translation.”
It is uncomfortable partly because the numbers are difficult to argue with.
In the United States, translator employment began losing momentum around the time Google Translate became widely used, and in the past five years the number of translators has fallen substantially relative to total employment. Typical translator wages, meanwhile, have slipped below the economy-wide average, while the number of new translation projects advertised on major online freelance marketplaces fell by almost 50 per cent in the two years after ChatGPT appeared.
But the most interesting part of the article, at least from where we sit at TranswrAIte, is not that machines are becoming capable of translation. We have built an entire product on the assumption that they are.
It is what happened to the humans once the machines arrived.
Welcome to the post-editing factory
The translation industry did not wake up one morning and replace every translator with an LLM. The transition has been more mundane.
Many agencies adopted what is called Machine Translation Post-Editing, or MTPE. Instead of giving a translator an original text and asking them to translate it, the machine produces a first version and the human translator receives the result with instructions to check it, correct it and polish it.
On a spreadsheet, this looks wonderfully efficient. The expensive human no longer has to translate every sentence. The machine has already done the heavy lifting, so the translator merely needs to tidy things up.
The FT gives a striking example. Petr Čermoch, who translates television subtitles from English into Czech, says that one agency which previously paid $5 per minute of video reduced its rate to $1.50 after moving to MTPE. Another translator working across legal, financial and academic texts reported rates being cut by roughly half.
And the work did not necessarily become easier.
Mark Rawson, an English-Chinese translator interviewed by O’Connor, describes the cognitive burden of simultaneously interrogating the source and the machine-generated translation. His verdict is wonderfully concise: “It’s two or three times harder than pure translation work.”
That sentence caught our attention, because there is something peculiar about the whole arrangement.
We have taken one of the most interesting parts of translation, deciding what a sentence means and how it should live in another language, and given it to the machine. Then we have taken one of the least interesting parts, hunting for errors, inconsistencies and awkward formulations, and given it to the human.
And somehow we call this human in the loop. But which loop?
Translating a novel is not simply a sequence of sentences that can be processed independently. A word on page 17 may acquire another meaning on page 280. A character speaks differently from another character. A metaphor returns. A joke depends on something that happened three chapters earlier. The narrator has habits, rhythms, prejudices and favourite constructions. Sometimes an awkward sentence is accidentally awkward. Sometimes it is awkward because the author wanted it that way.
A translator working from scratch gradually builds a mental model of all this. The book becomes a system.
Traditional MTPE breaks that system into little pieces and asks the translator to inspect the output of another intelligence. The translator is no longer primarily asking, How should I translate this? The question becomes, Is what the machine produced acceptable?
The second creates an anchor. Once a plausible sentence is sitting in front of you, it becomes surprisingly difficult not to inherit its choices. Vocabulary, syntax, rhythm and interpretation have already been proposed. The human becomes a critic of a translation rather than its author.
For technical material, that trade-off can make perfect sense. For literature, it can quietly sand away precisely the things we care about.
This is why our pipeline does not begin with post-editing
When we started building TranswrAIte, one tempting architecture was obvious: generate a translation with an LLM and provide tools for a human to correct it.
It is simple. It is cheap. It resembles the workflow the translation industry already knows.
But we increasingly think that treating AI translation as machine output plus human correction misunderstands what these models are good at.
Modern LLMs can do more than replace words across languages. They can analyse context, compare alternative interpretations, reason about character voice, follow stylistic instructions and revisit earlier decisions. That does not make them infallible translators. Far from it. But it means the interesting engineering problem is no longer merely how to generate the first translated sentence.
It is how to surround generation with enough context, analysis, memory and verification that the model can make better decisions in the first place.
That is why we think in terms of a pipeline, rather than a single translation call.
The system needs to understand the book before and while it translates it, maintain information across chapters, preserve terminology and names, identify passages where literal translation fails, review its own choices and flag uncertainty rather than quietly burying it beneath fluent prose.
There is a lesson here beyond translation
O’Connor and Burn-Murdoch make an interesting comparison with software development.
So far, AI appears to be affecting experienced programmers differently. Coding agents can remove some of the repetitive work of producing code, leaving developers more time for architecture, coordination, problem-solving and judgment. Translation has often experienced the reverse: the technology takes over the creative act of producing the translation while leaving the human with repetitive inspection.
That difference may tell us something important about AI products in general.
Again, the useful question is: Which percentage are we automating?
Automate the tedious part and technology can amplify expertise; automate the interesting part and leave humans to clean up the debris, and you may technically have a human in the loop while simultaneously removing much of the reason a skilled human wanted to be there.
That seems to be what happened to a significant part of professional translation.
Translation is not disappearing.
None of this means we believe the old translation market is coming back.
The FT data suggests quite the opposite. Automated translation first weakened demand, then improved, then became dramatically more accessible with large language models. Companies are increasingly willing to use machine translation without any human post-editing at all.
And there is another side to that story which is easy to overlook.
Cheaper translation also means more things can be translated.
For books, that matters.
A traditionally published bestseller can justify several thousand euros for a professional translation because the publisher expects enough sales to recover the investment. An independent novelist with a 120,000-word manuscript usually cannot make the same calculation. Nor can a small publisher afford to test five languages simply to discover whether readers exist there.
Let's be honest here: those books were not being beautifully translated by humans before AI arrived.
They were usually not being translated at all!
This is the part of the economics we find particularly interesting. AI is certainly reducing the cost of work that previously belonged to professional translators, and the consequences for that profession are real. But it is also pushing the marginal cost of translation low enough to create a market where previously there was none.
The question therefore becomes less interesting when framed as human versus machine.
For us, the better question is what kind of translation system we can build once machine translation becomes cheap enough to be treated not as the final product, but as one component inside a much larger process.
Perhaps the strange mistake of the first AI translation era was believing that the machine should translate and the human should clean up.
We are trying the opposite philosophy: make the machine translate, criticise, reconsider, remember, compare and clean up as much as possible, then leave the human where human attention is actually valuable.
That feels less like de-skilling translation and more like redesigning it.