Two Kinds of AI TranslationAnd why the difference decides what your content is safe for
People say “AI translation” as though it were one thing. It is two, built on different principles, and they fail in different ways.
You do not need a technical background to follow this page. The distinction matters because it determines how much human checking your content needs before it is safe to publish.
The short version
The specialist
Built for one job. It learned by studying enormous numbers of sentences that had already been translated by people, and it does nothing else. Google Translate, DeepL and Microsoft Translator all began this way.
The generalist
Built to work with language in general - writing, summarising, answering, and translating among other things. It learned from vast amounts of text of every kind. Claude and GPT are the familiar examples.
A useful comparison: the specialist is a career translator who has worked in one field for twenty years. The generalist is a gifted linguist who reads widely, writes beautifully, and will occasionally improve on your original when you did not ask them to. Both are valuable. They need different supervision.
How each one actually works
Neural machine translation: trained on translations
An NMT engine is trained on parallel text - millions of sentence pairs where a human has already translated the same content between two languages. Legislative records, technical manuals and published translations are typical sources. The system learns the patterns of how meaning moves between those two languages, and that is the entirety of what it knows.
Traditionally it worked one sentence at a time. That is efficient and predictable, but it means the engine cannot see that “the patient” three sentences earlier was a woman, so in a language that marks gender it has to guess.
Large language models: trained on everything
An LLM was not built for translation. It was trained to continue text plausibly, across every subject and style it encountered, and translation turned out to be one of the things that ability produces. It reads far more of the surrounding document at once, so it resolves the gender question above without difficulty.
The same design produces its characteristic weakness. A system trained to produce plausible text will produce plausible text even where the source was unclear, damaged or ambiguous - filling a gap rather than flagging it.
Where they differ in practice
The failure modes matter more than the strengths, because the strengths are obvious and the failures are not.
A reviewer catches an awkward sentence easily. A fluent error can pass straight through, which is why LLM output needs a reviewer comparing against the source rather than simply reading the translation.
Yes. Run it twice, get the same result.
Not necessarily. The same page can come back slightly differently on a second run.
Select any row marked with an arrow to see why the difference matters in practice.
The single most important difference
If you take one thing from this page, take this. When a traditional engine gets something wrong, it usually looks wrong. When a language model gets something wrong, it looks right.
Consider a consent form stating that a side effect occurs in a small number of patients. A traditional engine might render it awkwardly - a reviewer notices and fixes it.
A language model might produce a sentence that is clearer and better written than the original, and states a slightly different frequency. Nothing about the sentence draws attention to itself. Only someone comparing it line by line against the source would catch it.
This is why our review stages compare against the source document rather than simply reading the translation for fluency, and why our highest tier removes AI from the workflow altogether. Reading well is not evidence of correctness.
How translation quality gets measured
Worth knowing, because published quality comparisons often rest on measures that suit one architecture better than the other.
Word-overlap scores (BLEU and similar)
Compares the machine output against a human translation and counts how much of the wording matches.
The catch: It rewards matching the reference wording rather than conveying the meaning. A language model that translates a passage correctly in different words scores badly - so these measures systematically understate LLM quality.
Meaning-based scores (COMET and similar)
Uses a trained model to judge whether the meaning survived, rather than whether the words match.
The catch: Fairer across both architectures, and now the standard in serious evaluation. Still a proxy, not a final verdict.
Structured human review
Qualified linguists classify each error by type and severity, so a dropped clause is not counted the same as a comma.
The catch: Slower and more expensive than any automated score, and the only method that reliably surfaces the fluent-but-wrong error.
What this means for our service tiers
The architecture is not a detail we keep to ourselves - it determines which tier suits your content.
Predictable, fast and repeatable. With no human reviewing the output, predictability matters more than flair.
A light review checks fluency rather than comparing every line against the source, so the architecture less likely to invent or omit content is the safer starting point.
A subject-matter translator edits against the source line by line, which is exactly the check that catches a fluent error. That makes it safe to use a language model where nuance and tone genuinely benefit from one.
No AI at any stage. For regulatory, legal and medical content the risk of a fluent, invisible error is not one worth managing - it is one worth removing.
A note on how fast this is moving
The two categories are converging, and anyone presenting them as permanently opposed is describing a snapshot rather than a rule.
- ›The major engines now offer LLM-based modes alongside their traditional ones, so the choice increasingly happens inside a single provider rather than between providers.
- ›Newer neural systems read beyond the single sentence, narrowing the context gap that was the clearest distinction between the two.
- ›Language models are being tuned specifically for translation, which improves their tag handling and reduces the tendency to embellish.
What does not change is the underlying question: how much human checking does this particular content need before it can be relied on? We review our engine choices on a fixed schedule for exactly this reason.
Not sure which your content needs?
Tell us the language pair, the subject and what the document is for. We will recommend a tier and explain the reasoning.