It’s the question every translator hears time and again: isn’t Google Translate or DeepL enough? The short answer is: it depends on what could go wrong. When it comes to understanding an email from a supplier, machine translation is an excellent tool. But when it comes to a clinical report, a contract or a machine manual, it’s not quite the same. Let’s see where the line is drawn.
What machine translation does well
Today’s neural network-based engines have improved immensely and produce fluent sentences. They are excellent for three things: grasping the gist of a text you don’t fully understand, translating low-risk internal communications, and generating a first draft that someone will then review.
The criterion is simple. If a mistake costs, at most, a misunderstanding that can be resolved with another email, the system works. If a mistake could cost someone their health, money, a legal deadline or a company’s reputation, the system is no longer fit for purpose on its own.
Where the problems begin: specialised translation
In specialist texts, the risk does not lie in the grammar – which machines usually get right – but in the meaning. Each field has its own terminology, conventions and implications:
- Medical: The abbreviation ‘MS’ can stand for multiple sclerosis, morphine sulphate or mitral stenosis. A human translator specialising in the field makes a judgement based on the clinical context; a machine chooses the most likely option, which is not always the correct one. A missed negation or an incorrectly converted dose can have serious consequences.
- Scientific: The strength of a statement is measured by the verbs used. ‘Suggests’, ‘indicates’ and ‘demonstrates’ do not carry the same weight, and swapping one for another alters what the study claims.
- Technical: The same term must always be translated in the same way. If ‘bracket’ appears as ‘support’ on one page and as ‘angle bracket’ on the next, the person carrying out the maintenance will waste time or install it incorrectly.
- Legal: Many terms belong to different legal systems and have no exact equivalent. In English law, ‘consideration’ refers to the consideration provided by each party, not ‘consideration’ in the general sense; and replacing ‘shall’ with ‘may’ transforms an obligation into a possibility.
What makes this tricky is that machine translation produces fluent text even when it is incorrect. It may omit sentences or add others, and the result reads so well that nobody suspects a thing. Only those who are fluent in the source language and familiar with the subject area can spot what has been left out.
Proofreading and post-editing: where the human eye makes all the difference
When the text is sensitive, the crucial factor is not who writes the first draft, but who reviews it. The industry calls it post-editing to the hybrid model in which a human translator uses the machine-translated text as a starting point and corrects it. The standard ISO 18587, which is dedicated to this process, argues that only qualified post-editors should be involved and that the final post-edited version should be of a quality comparable to that of a human translation.
There is also the standard ISO 17100, for translation services, which requires each translation to be reviewed by a second qualified linguist before delivery. It is easy to see why: translators tend to read what they meant to say, rather than what is actually there.
What a human proofreader checks, apart from grammar:
- correct and consistent terminology from start to finish;
- numbers, units, dates and proper names;
- sentences omitted or added without notice;
- tone, style and conventions appropriate to the reader and the target country.
And in some cases, there’s no debate at all. According to the DGES, when a certified translation is required in Portugal, it is certified by bodies such as notaries, solicitors, legal practitioners, consulates or chambers of commerce and industry, and the translator must declare to them that the text is an accurate translation. No machine can do that.
So, when should you use which one?
The higher the cost of a mistake, the more human eyes the text needs. This is my rule of thumb:
| Text type | Just machine translation? | What I recommend |
|---|---|---|
| To understand an email or a web page | Yes | Machine, no problem |
| Low-risk internal communication | Yes, carefully | Machine; review if text is to be circulated |
| Website, marketing and brand communication | No | Human translator, or post-editing by a professional |
| Technical documentation and manuals | Just as a draft | Post-editing by a technical translator, with a glossary |
| Scientific articles and reports | Just as a draft | Post-editing by a specialist in the field and proofreading |
| Medical documents | No | Specialist translator and proofreading by a second linguist |
| Contracts and legal or official documents | No | Legal translator; certified translation where required |
A good translator does not compete with the machine: they use it when it makes sense and take responsibility for the final result. It is this responsibility, and not the dictionary, that the machine is still unable to assume.