I studied French at university. I started out translating film subtitles for pirated DVD sellers back in the days of the old fansub groups. I have worked in Chinese, English and French translation for more than twenty years and have published over a dozen books as a translator.
At a class reunion the other day, a classmate asked whether I use AI for translation now, or still type every word myself.
I said of course I use it, and I started earlier than most people.
The difference is that, to me, AI is not a tool. It is an employee.
Both popular stories are wrong
There are two popular stories about translators and AI.
One says translators are resisting as a group, clinging to pen and paper and refusing to surrender because machine translation has no soul and ruins language.
The other says the opposite: AI is already so powerful that you can throw in the entire original, click once, and get a finished manuscript. The translation profession will disappear sooner or later.
Both are wrong. The second is more seriously wrong, because that is exactly how many companies are using AI now—and not only for translation.
Here is a fact that may surprise you: translators did not start using machines to help with their work in the past two years. In the 1990s, computer-assisted translation software such as Trados was already using terminology databases and translation memories. A sentence you translated before is stored and brought up automatically the next time a similar sentence appears; proper names are registered in advance to keep a whole book consistent.
This system has been around for thirty years. It used to be called CAT: Computer-assisted Translation or Computer-aided Translation.
AI has not replaced that workflow. It has replaced one step inside it.
The boss does not do the work. The boss sets the rules.
I am the boss; AI is the employee. Once you understand this, the psychological hurdle disappears.
The boss does not do the work personally. The boss does two things: set the rules before work begins and inspect the result before delivery. The physical work in between was never supposed to be the boss’s job.
That sounds like slacking off. But if you actually follow this approach, you will find the boss’s workload is no smaller. It has simply changed form: from typing to judgment. And since I am both the boss and the employee, I still have to proofread the work myself!
Here is the full process I use as the boss.
Set the tone before getting to work
Step one: read the original manuscript from beginning to end.
AI cannot help with this step, and I should not let it interfere. After reading, I need to answer a few questions: Is this book written for specialists or for beginners? Is the author being precise here, or making a joke? Should the sentences be relatively long or short? Which awkward turns in the original should we keep, and which ones must be smoothed out?
If I do not settle these questions first, every sentence that follows will drift. The most common problem in a translation is not an outright mistake. It is having no consistent tone: one paragraph reads like an academic paper, the next like advertising copy. Readers may not be able to say what is wrong, but the prose feels awkward.
I set the tone. Once it is set, it becomes the first rule for my employee.
For example, if a conversation is spoken in an informal register, the translation should not sound stiff and literary. Even if the original dialogue uses subordinate clauses, that may simply reflect the nature of the source language, not the personality of the author or speaker. I would still render it in a more conversational way.
A glossary
Step two: build a glossary.
People’s names, place names, technical terms and other specialist vocabulary all need to be checked in dictionaries and reference materials before I settle on translations. Where a standard translation exists, I use it. Where none exists, I create one. The first choice is the choice I use throughout the book. I used to make this list for myself, relying on memory and Ctrl+F to find things. Now I make it for AI.
This step is tedious, but it is the foundation of the whole manuscript. If the same term is translated one way early in a book and another way later, readers may not be able to explain the problem, but an expert will immediately see that you did not pay attention.
Only after the rules are set does the employee get to work.
It writes the first draft. I own the mistakes.
Step three: AI produces a first draft according to the tone and glossary established above.
Step four: I proofread it sentence by sentence.
There is not much to say about these two steps, but one thing must be clear: AI wrote the first draft, but I own the mistakes. My name is printed in the book, so I am the one putting my work out into the world. If readers come across an awkward passage, they will blame me. That line cannot be blurred: you can hand off the work in your hands, but you cannot hand off responsibility.
Most people’s idea of an AI translation workflow ends here: the machine produces a draft, a person edits it, and that is that. Some do not even edit it.
But I have a fifth step. This is the step I consider most valuable.
Let it review me in return
Step five: I send my proofread translation back to AI together with the original. I ask it to compare the two against the acceptance criteria I have given it and point out anything worth a closer look.
Notice that this time it is not reviewing its own first draft. It is reviewing the draft I have edited.
It tends to find a few kinds of issues: I rewrote a sentence to make it read more smoothly, but its meaning shifted; I missed a pun or a callback in the original; I translated a term one way earlier and changed it later—a common human mistake when you get tired halfway through a book; or I deleted a qualifier to improve the flow even though the author put it there deliberately.
Of the ten things it lists, seven may be nonsense. The remaining three are worth it.
The key is that it does not decide for me. It only puts the things I missed on the table. I still decide whether to change them.
How is this different from the first four steps? In those steps, it follows my instructions. In the fifth, I ask it to find my mistakes.
An employee who only knows how to obey is not worth the salary.
This is where most people get stuck when using AI. They know how to make it do work, but they do not know how to ask it to find their own mistakes. Deep down, they still think of it as a machine: how could a machine tell me I am wrong?!
Put another way: if you hire someone who says only, “Sure, understood, I’ll do it right away,” from beginning to end, then either there is a problem with them or there is a problem with you. Take a good look at yourself.
No need to explain everything again each time
Step six: write the entire process down.
If I had to explain it from scratch for every new project—read the manuscript, build the glossary, translate, proofread, compare at the end—I would go mad sooner or later, and I would inevitably skip a step a few times. So I wrote the workflow into a fixed set of instructions: what to do at each stage, which rules to follow, and what it must not change on its own. AI follows them.
Once the instructions are written, I do not have to explain the process again when I start a new project. I only need to tell it the tone of this manuscript and give it the glossary. It already knows what to do with the rest.
The instructions themselves are an asset. They come with me from one project to the next, get smoother with each use, and incorporate the lessons from past mistakes.
The task in my hands can change. The workflow does not have to learn all over again.
Give it a more impressive name
So far, I have explained everything in plain language. But these six steps are not an improvised method I invented on my own. Each has a more impressive name in project management, and anyone working on AI projects should recognize them.
| Translation workflow | The impressive name |
|---|---|
| Read the original and set the tone | Requirements definition, Definition of Done |
| Build a glossary | Terminology database, constraints, context injection |
| AI produces a first draft | Generation phase, first deliverable |
| I proofread it sentence by sentence | Human-in-the-loop |
| Ask it to review me in return | LLM-as-a-judge, cross-checking |
| Write the instructions once and for all | SOP, workflow orchestration |
The terminology itself is not important. It does not make the work more sophisticated. Its purpose is to make the work transferable. If you do it alone, you can call it “the way I like to do things.” If three other people in the company need to follow it too, it needs a name.
The two terms in the middle are the ones worth remembering.
Human-in-the-loop means that a person must be connected to the critical step, not merely standing by and watching. The test is simple: does this person have veto power? If yes, they are in the loop. If all they can do is sigh afterward and say, “This is not quite right,” they are outside it. What many companies call “human review” is really the latter.
LLM-as-a-judge means using one model to evaluate another model’s output. The industry originally used it for large-scale quality checks: when there were too many drafts for people to review, machines would screen them first. I turned it around and applied it to myself: the model evaluates a translation it did not write, one I have already edited.
As for the final term, “workflow orchestration,” the word orchestration comes from orchestra. Its original meaning is arranging the instrumentation: deciding which melody goes to the strings and which to the brass.
The person on the conductor’s podium does not touch a single instrument for the entire performance.
But the performance is their responsibility.
Is anyone in your company acting as the boss?
I have been talking about translation, but what I am really describing is project work. These six steps apply almost anywhere: writing bid proposals, responding to customer inquiries, organizing sales records, writing product copy and preparing meeting notes. Any repetitive task with clear standards that still requires a person to make the final judgment has the same structure.
I have seen many companies buy AI, use it for six months, then conclude, “It’s not all that.” The problem usually is not the tool.
No one is doing step one, and no one is doing step five.
Most companies use AI like this: put the work in, take the output out, use it if it looks good, and if it does not, complain that AI is useless. No one sits down first to define what a good result should look like, and no one asks it afterward to find their own mistakes.
That is not being the boss. That is being the operator.
That person may run a company of a few dozen people. They may have taken an AI course last year, posted “Very enlightening” in the group chat, and never followed up. It is not because they did not care. No one told them what the first thing to do after they started caring should be.
And the first thing is not mysterious at all. Sit down and explain clearly what the task should look like when it is done.
It is exactly like onboarding a new employee.
In the end
So, back to the question at the dinner table.
Of course I use AI. I have used it for a long time.
I just do not hand it the manuscript. I hand it the standards.
It translates; I revise. Once I am done, I ask it to find my mistakes. After it does, I decide whether to revise.
When the book comes out, it still bears my name.