Machine translation was once easy to spot. Stilted sentences, obvious grammatical errors and strange word choices gave the game away. Today's neural systems, trained on vast collections of texts, produce output that is often fluent enough to pass a quick read. That fluency is a genuine achievement, but it also creates a new risk: people trust the text because it reads well, even when it is subtly wrong.

How neural machine translation works

Unlike older systems that translated word by word using fixed rules, neural models learn statistical patterns across millions of sentence pairs. They encode the meaning of a source sentence and generate a target sentence that is statistically likely to be appropriate. Large language models go further still, predicting text from enormously broad training data.

The strength of this approach is also its limitation. The models do not "know" things in the human sense; they predict what a plausible translation looks like. When the plausible and the correct coincide, the output is excellent. When they diverge, the models can still write with complete confidence.

What NMT does genuinely well

  • High-volume, low-risk content. Internal communications, support tickets, community posts and similar material can be translated fast enough for practical use.
  • Gisting. When the goal is simply to understand what a document says, machine translation gives readers rapid access in dozens of languages.
  • Repetitive material with established terminology. Engine customisation, glossaries and previous high-quality translations substantially improve consistency.
  • Languages with abundant data. Major languages paired with English are served far better than low-resource languages with limited corpora.

What it cannot reliably do

The failures are not random; they cluster in precisely the areas where stakes are highest.

Omission and invention. Neural output can silently drop a negation, skip a clause or add information that was never there. Because the resulting sentence reads naturally, such errors are hard to notice without comparing carefully against the source.

Terminological precision. A model may choose an everyday word where a regulated term is required, or pick the wrong sense of a technical term. In clinical and legal texts, a single word can alter obligations or medical meaning.

Numbers, names and identifiers. Dates, measurements, dosages, product codes and proper nouns can be mistranscribed, and these errors are among the most dangerous in professional content.

Context beyond the sentence. Long-range references, document-wide consistency and decisions that depend on understanding the purpose of a text remain weak points.

The role of post-editing

This is why professional post-editing has become a distinct discipline. Full post-editing aims for text indistinguishable from human translation; light post-editing delivers accuracy and readability without perfect polish. Either way, the editor must command both languages, understand the subject and compare against the source systematically rather than merely smoothing the target text.

The process is governed by standards such as ISO 18587, which sets requirements for the post-editing of machine output and the competences of the people doing it.

Custom engines and language portfolios

One development that genuinely changes outcomes is customisation. Rather than relying on a generic engine, professional teams can build or adapt systems using a client's own approved translations, glossaries and style decisions, producing output that respects established terminology from the first draft. For ongoing programmes, technical documentation, software strings and support content, this narrows the gap between machine draft and finished text considerably.

Even so, customisation is only as good as the data supplied. Feeding an engine inconsistent or unreviewed material simply teaches it bad habits at scale, which is another reason quality assets have to be curated. A well-customised engine also remains weak at the edges of its experience: novel wording, unusual document types and language pairs outside its training. The disciplined approach is to match engine to task and keep expert review proportionate to the stakes.

Choosing wisely

The mature position on neural machine translation is neither enthusiasm nor dismissal but judgement. It is an excellent tool for the right content, and an unsuitable one for high-stakes text without professional review. The decision should start with three questions: What happens if the translation is wrong? Is a qualified linguist reviewing the output? Does the text involve regulated terminology, numbers or cultural nuance?

Used honestly, NMT makes cross-language communication cheaper and faster than at any point in history. Treated as a finished product for consequential content, it remains a risk that no responsible organisation should take. The technology has come of age; the need for human expertise has not disappeared, but changed form.