Start here Start here 4 min read
A familiar scene

AI translates smoothly and then gets the same term wrong everywhere; fixing each spot by hand treats the symptom — handing it a glossary once treats the cause

A fluent translation isn't a correct one — it just doesn't know your field's vocabulary yet

Cluster11 · Start here
Topicscontext · beginners
TL;DR

AI translates fluently because it reaches for the most common rendering — but the most common rendering is rarely the term your field uses. Worse, it's wrong consistently: the same word wrong everywhere, so it looks deliberate and you stop scanning for it. Don't hand-fix each spot — give it a short glossary once ("this word must be X, not Y"), and every translation after that holds the line.

The translation reads smoothly, like an insider wrote it. You nod through the first two paragraphs. On the third line you stop short: that domain term — it's wrong. Not grammatically, not in plain meaning. Wrong in that your field has a fixed way of saying it, and it used a different one. You scroll down and your stomach drops: the same mistake sits in a dozen places below, perfectly even.

01Fluent isn't correct — they're two different things

AI translates smoothly because it learned from an ocean of text and always reaches for the most common phrasing. And for a specialist term, the most common phrasing across that whole ocean is rarely the one your narrow field uses. It doesn't know your field's "internal dictionary" — the words a whole industry quietly agrees on, or the ones your company deliberately keeps in English, or renders a very particular way for legal reasons.

Plain translation, no glossary

It picks the rendering that's most common for everyone
Grammatical, fluent — and off your field's standard
You hand-fix each term, in each document, forever

Translation with a glossary attached

You give it upfront: "this word means X, not Y"
It holds that list; the rest still reads naturally
On-standard from the very first draft

Same AI, same source text. The difference isn't that it "got better at the language" — it's that you told it the one thing it could never guess: what your field calls that thing.

The core point: fluency and term-accuracy are two separate abilities. The first it has by default. The second can only come from you — because it lives in your head and in your field's conventions, not in the ocean of text it learned from.

02Wrong but even — so you stop scanning

This is the dangerous part. A random error is easy to catch — your eye trips over it. But AI is wrong very evenly: get one term wrong and it gets that term wrong identically everywhere, no wobble. That consistency fools your eye. By the fifth time you read past the wrong word, your brain has stopped raising the alarm — it's grown used to the word, and the familiar gets mistaken for the correct.

So the wrong translation slips through: not because it hid well, but because it's wrong in a way too tidy to look like anything but a deliberate choice. The real cost isn't the few minutes of fixing each word — it's the times you don't fix it, because by the tenth instance you'd stopped looking.

03Hand it the glossary, once

The fix isn't telling it to "translate more carefully" — that gives it nothing new to hold onto. The fix is handing it the exact thing it's missing: a short glossary, made once, reused forever.

1
List the "sensitive" words

Gather the ten or fifteen domain terms that get mistranslated most. One per line: "source term → render as X (NOT Y)."

2
Name what stays as-is

"These names stay in English, don't translate them." Machines are very eager to translate the things you didn't want translated.

3
Paste it on top each time

"Translate using the glossary below; for any term not on it that you're unsure about, flag it — don't decide on your own."

This list isn't extra work — it's where the time you keep spending on hand-fixes goes to live. Write it once, and the next translation is just a paste away.

Notice the last line: "if you hit an unfamiliar term you're unsure about, flag it — don't decide." This is what turns a blind translator into an assistant that raises its hand. Instead of silently guessing at a new term and burying that guess inside a fluent sentence, it stops and points you at exactly the spot that needs a human eye. You no longer comb the whole document for errors — it tells you where to look.

And the payoff is bigger than "less fixing this time": the glossary compounds. Next translation, you don't start over — you add the few new terms you hit, and it matches your field's voice more closely each round. The wrong-but-even translation disappears, and so does the strain of scanning every line. This is really a larger habit of the craft: setting it up once so you skip the explaining forever, and choosing what to put in for the agent to anchor on. That mistranslated term was never the AI's fault. It was just the glossary you hadn't handed over.

Other pains

Get new pieces by email

Field notes on working with AI agents — occasional, no spam.