Translator: How It Works
Machine translation has become genuinely good at conveying meaning and remains unreliable at conveying nuance, register and cultural context. Knowing which tasks fall on which side of that line is what makes it useful rather than embarrassing.
Where it works and where it does not
| Reliable for | Unreliable for |
|---|---|
| Understanding the gist of foreign text | Legal contracts and official documents |
| Travel phrases and signage | Medical instructions |
| Informal messages between individuals | Marketing copy and slogans |
| Technical documentation with plain sentences | Poetry, humour, wordplay |
| Getting a first draft to be edited by a speaker | Anything published under your name |
The pattern: it handles denotation well and connotation badly. If the sentence means one thing plainly, translation is usually good. If it depends on tone, implication, register or a cultural reference, it is not.
Why quality varies by language pair
Systems are trained on parallel texts, so quality tracks how much translated material exists between two languages. English paired with Spanish, French, German or Chinese has enormous training data and produces strong results. Lower-resource languages have far less, and quality drops accordingly.
Translating between two lower-resource languages is often worse still, because the system may route through English internally — and errors from each step compound. Where quality matters, translating each language to and from English separately can produce better results than a direct pair.
Formality is the recurring failure
Many languages encode social relationship in grammar. German has du and Sie, French tu and vous, Japanese has an elaborate honorific system, and Korean and Tamil have their own registers. English has none of this, so an English source carries no signal about which to use, and the system guesses.
The consequence is real: a business email translated from English into Japanese or Korean may arrive at an inappropriate level of formality, which reads as rudeness rather than as a translation artefact. This is among the most common ways machine-translated correspondence causes offence.
Idioms
Idioms translate literally and become nonsense. 'Break a leg', 'under the weather' and 'the ball is in your court' all produce confusion or alarm. Rewriting the source into plain literal language before translating is the single most effective improvement available — write what you mean rather than how you would say it.
Getting better results
- Write short, simple sentences in the source language.
- Avoid idioms, slang and humour.
- Use explicit subjects — languages that drop pronouns cause ambiguity.
- Avoid ambiguous words. 'Bank', 'light' and 'right' each have several meanings.
- Back-translate to check. Translate the result back and see whether it still says what you meant.
- Have a speaker review anything consequential.
What to never machine-translate
Legal agreements, medical instructions, safety information, official filings, and any published text representing an organisation. Errors in these categories carry real consequences, and certified human translation exists precisely because of them. Machine translation is an excellent way to understand a document and a poor way to produce one that someone will rely on.