Translation you can check

Machine translation is the most-used language feature in the world and the one people check least, for an obvious reason: if you could evaluate the output, you would not have needed it.

This lesson is about checking anyway.

The fluency trap

E7·L1’s argument, arriving in its most acute form.

Older machine translation was visibly clumsy. That clumsiness was information — it told you to be careful, get a human, not send it.

Modern translation is fluent. It reads as though a competent speaker wrote it, including when the meaning is wrong. The warning signal has been removed while the error rate has fallen but not vanished, which is a worse combination than it sounds.

You now need a deliberate check where you used to get a free one.

Back-translation

The single most useful technique, and it costs a minute.

Translate your text into the target language. Then, in a fresh conversation, translate the result back.

The fresh conversation matters. In the same thread the model has your original in context and will reproduce it rather than genuinely re-translating — you get a comforting match that tested nothing.

Compare the round trip to your original:

Meaning preserved, wording different — normal, and fine.

A fact, number, name or negation changed — a real error. Fix it.

Something disappeared — usually a qualifier, and usually the important part. E4·L3 and E5·L7 again: conditions go first.

Something appeared — the translation added a claim.

What back-translation cannot catch

Be clear about the limit, because this technique inspires false confidence.

Register and politeness. A translation that is grammatically perfect and insultingly informal back-translates perfectly. In languages with grammatical honorifics this is the most common serious failure, and the round trip is blind to it.

Cultural fit. An idiom rendered literally, a form of address that is wrong for the relationship, a phrasing that is technically right and lands badly.

Terminology conventions. Your industry’s accepted term versus a correct but unused synonym.

All three need a speaker. Back-translation checks meaning, not appropriateness — and for anything going to a customer, an official or a public audience, appropriateness is most of the job.

What to translate, and what not to

Fine to machine translate: internal understanding, gist of an incoming message, first drafts, low-stakes correspondence, your own notes.

Get a human: contracts and anything legally binding, medical information, safety instructions, marketing (where tone is the product), anything published under your organisation’s name, and anything where being wrong is expensive or embarrassing.

The middle ground — machine translate, then have a speaker review — is frequently the right answer and much cheaper than full human translation. Ask for a review rather than a translation and the cost drops a long way.

Give it the context it lacks

Translation quality improves substantially with information the text does not contain:

Translate into [language]. This is a message from a supplier to a long-standing customer. Formal but not stiff. Keep “SKU” and the product codes in English. The audience is in [place].

Audience, relationship, register and do-not-translate terms. Those four lines prevent most of what back-translation cannot catch — which is the practical answer to the previous section.

Try it now (7 minutes)

Take a short message you have actually sent. Translate it, then back-translate in a new conversation.

Compare against the original and note which category any difference falls into. Then re-translate with audience, relationship and register specified, and see how much moves.

Check your understanding

1. Back-translation must be done in a fresh conversation because: It is faster Otherwise the model has your original in context and reproduces it rather than re-translating The context window would overflow

2. Which error does back-translation reliably fail to catch? A changed number A politeness level that is grammatically perfect and insultingly informal A dropped qualifier

3. Modern fluent translation is more dangerous than older clumsy translation because: It is wrong more often Clumsiness used to be a free warning signal, and fluency removed it It is harder to back-translate

Recap

Fluent translation removed the clumsiness that used to warn you, so check deliberately. Back-translate in a fresh conversation and look for changed facts, dropped qualifiers and added claims — while remembering it is blind to register, culture and terminology. Supply audience, relationship, register and do-not-translate terms up front, and get a human for anything binding, medical, safety-related or published.