Glossary
Machine translation
Converting text from one language to another automatically, now good enough for most working use and not for everything.
In plain terms
Turning text into another language automatically. It is good, which is the problem in one specific way: when it is wrong, it is wrong fluently, so nothing about the result looks like a warning.
Why it matters
Because the old signal for a bad translation was that it read badly, and that signal has gone. Output that is confident, idiomatic and subtly wrong is now the failure mode, and it passes every check a non-speaker is able to make.
How it works
For understanding, it is comfortably good enough. Reading an incoming message, following a document, working out what a supplier is asking: all of that is served well, and treating it with suspicion costs time for nothing.
For anything that commits you, it is a draft. Contracts, safety information, medical or legal wording and public promises all turn on precise phrasing, and precise phrasing is exactly where a fluent approximation is indistinguishable from the real thing.
The failure mode is fluency rather than clumsiness. Old translation failed visibly, so a reader knew to be careful; a plausible, well-formed sentence carrying the wrong obligation gives nobody a reason to look twice.
Which makes it unverifiable by the person who ordered it. Somebody who cannot read the target language cannot tell a good translation from a confident one, so the check has to come from someone who can, or not at all.
House and technical vocabulary is where it slips first. Your product names, internal terms and settled phrases have a correct form in each language, and nothing in a general system knows what your organisation decided to call things.
Tone travels less reliably than meaning, and it is the part nobody reviews. A sentence can be accurate and land as brusque, deferential or informal in a way the original was not, which matters most in exactly the correspondence people are most likely to send unreviewed.
Back-translation is a weak check that beats none. Translating the result back and reading it catches gross errors and misses subtle ones, so it is worth doing where no reviewer is available and is not a substitute for one where the wording matters.
How translation used to fail, and how it fails now
Common misconceptions
People assume
If it reads well, it is right.
In fact
Reading well is what it does most reliably. A confident, idiomatic sentence carrying the wrong obligation is the characteristic failure, and fluency is the property that stops anybody looking again.
People assume
We can check it ourselves.
In fact
Not if you cannot read the target language. Somebody who cannot will find every output equally convincing, which is why the review has to come from a speaker or be honestly recorded as not having happened.
Questions
- When is it fine to use without review?
- For understanding: incoming messages, documents you need the sense of, anything at all where you are the reader. The risk arrives only when the output is published, sent to somebody else, or relied on as a statement of what you have agreed to.
- Why is a fluent translation harder to catch than a clumsy one?
- Because clumsiness was the warning. When bad translation read badly a reader knew to take care, whereas a well-formed sentence carrying the wrong meaning removes the signal and leaves the error, which on that one dimension is a worse position than before.
- What can we do without a speaker to hand?
- Translate the result back and read it, which catches gross errors and misses subtle ones. Keep a short list of your own terms with their agreed forms in each language, which prevents the most common and most visible category of error.
Key takeaways
- Excellent for understanding; a draft for anything that commits you.
- Fluency is the failure mode: wrong output reads perfectly.
- A non-speaker cannot check it, and will find every version convincing.
- Keep your own terms in a list; general systems do not know them.
Last checked August 2026