Nordic NuanceWhat I still check when machine translation sounds right
Machine translation has become good enough that many of its harder errors no longer look like errors.
The sentence is grammatical. The terminology sounds plausible. Nothing immediately interrupts the reading.
That is precisely why fluent output still needs review.
1. A fluent word can still change the meaning
One recent source referred to understanding user journeys.
The Danish translation used brugeradfærd — user behaviour.
There is nothing unnatural about brugeradfærd. The problem is that it describes a broader concept. A user journey is the sequence or path through interactions; behaviour can mean almost anything the user does.
The sentence sounded fine. The distinction had still disappeared.
This is one reason I am cautious about reviewing a translation only for fluency. A sentence can become easier to read at the same moment it becomes less precise.
2. Literal terminology can point to the wrong world
Another text mentioned advertising trading desks and supply-side and demand-side platforms.
Literal Danish produced terms resembling handelsafdelinger and udbuds- og efterspørgselsplatforme.
Both are linguistically plausible. But without the advertising context, the first can sound like a commercial or financial trading department, while the second can evoke procurement or general economics rather than programmatic advertising.
The better Danish needs to identify what these things actually do.
This is where terminology work differs from replacing one word with another. A term belongs to a domain, a product and a reader — not just to a dictionary entry.
3. Small words can change who does what
Prepositions are particularly good at hiding inside otherwise natural sentences.
In one review, the English wording around identifiers being licensed by selected partners was itself ambiguous. A literal Danish rendering with af could imply that the partners were granting the licence, while the surrounding context suggested that they were receiving the right to use the identifiers.
That was not a case for confidently “fixing” the translation. It was a case for asking what the source actually meant.
The useful skill is sometimes not finding a better translation.
It is recognising that the available information does not justify one yet.
4. A single preposition can change an instruction
In a German-to-Danish review, an instruction had become:
Tilsæt sukker efter smag under saftkogningen.
Add sugar to taste during the steaming process.
But the surrounding instructions showed that the sugar was supposed to be added before the heating began.
The revised Danish therefore used før saftkogningen. One small word changed the sequence of the action.
This is a good example of why sentence-by-sentence fluency checking is not enough. Sometimes the error only becomes visible when the reviewer reads what happens before and after the sentence.
5. Numbers and formatting are language work too
Some errors are less glamorous but potentially more consequential.
In a recent technical review, two load values — 200 kg and 227 kg — had been transposed during the handoff into the translation file. The Danish faithfully reproduced what it had been given; the reference artwork showed that the translation file itself was wrong.
The same project included the usual smaller details: keeping units with their values, using consistent dimensions, decimal conventions and terminology across several related documents.
A translation can contain excellent sentences and still fail because a number, unit, identifier or component name is wrong.
And sometimes the correct review decision is to flag the source or workflow rather than blame the translation.
Fluency is the beginning of the check
Modern machine translation is very good at producing text that looks finished.
That is a real improvement. It is also why review has changed.
The obvious errors are easier. The harder ones now often involve scope, context, terminology, actors, sequence and tiny differences in meaning that a reader can glide straight past.
For me, the useful question is no longer simply:
Does this sound like good Danish?
It is:
Does this say exactly what it needs to say, in this context, for this reader?
That is a much slower question.
But it is where linguistic review still earns its place.