Artificial intelligence in translation: why human review remains essential

Artificial intelligence has profoundly changed the way we produce and translate text.

Machine translation tools and generative AI models can now translate thousands of words in seconds and produce results that, at first glance, look natural and convincing.

This is a significant technological advance.

But it also introduces a less obvious problem: the more convincing a translation looks, the easier it can be to trust an error that does not look like an error.

There is a fundamental difference between producing linguistically plausible text and ensuring that the text accurately conveys the content of the source.

That difference is where human review remains essential.

The most dangerous error is the one that looks correct

A spelling mistake or a poorly constructed sentence is easy to identify.

A perfectly written sentence that conveys a different idea is far more problematic.

An artificial intelligence system may select the wrong sense of a word, omit a restriction, alter a logical relationship, resolve an ambiguity or implicitly add an explanation that is not in the source.

The result may still sound entirely natural.

This is particularly relevant because generative models are designed to produce coherent, plausible language. The quality of the wording is not, by itself, evidence of the quality of the translation.

Translation is not word substitution

Professional translation does not consist of finding an equivalent word in another language for every word in the source.

Languages divide up reality differently, and concepts do not always coincide.

This is especially clear in legal translation.

The English term consideration, for example, has a specific meaning in the contract law of common-law systems. Automatically translating the word into Portuguese as the cognate «consideração» may produce a linguistically possible sentence without conveying the legal concept at issue.

The problem cannot be solved simply by looking for a dictionary correspondence. The concept must first be understood, and only then can a decision be made about how to represent it in the target language.

The same principle applies to financial, business, administrative, scientific and technical terminology.

The relevant context may be several pages earlier

Another limitation of a purely sentence-by-sentence approach is that meaning rarely ends with the sentence.

A pronoun may refer back to an entity identified earlier. A term may have been defined at the beginning of a contract. An abbreviation may acquire a specific meaning within the document. A terminology decision made on page 2 must remain consistent on page 40.

Translation should therefore be analysed as a document, not merely as a sequence of independent segments.

This global reading makes it possible to assess not only whether a sentence is possible, but whether it is the right sentence in that particular text.

Machine translation and generative AI are not exactly the same thing

It is also important to distinguish between different technologies that are often grouped together under the label “artificial intelligence”.

Machine translation systems are developed specifically to transform content from one language into another.

The large language models used in generative AI tools have a much broader function. They can translate, summarise, explain, reformulate, develop arguments and produce new content.

That versatility can be useful, but it creates an important distinction.

In a summary or explanation, making an implicit idea clearer may be desirable.

In a faithful translation, it may amount to changing the source.

The translator has to distinguish between what the text says, what it probably means and what could be added to make it clearer.

They are not always the same thing.

So should artificial intelligence not be used for translation?

That conclusion would be excessive.

Artificial intelligence can be a valuable tool in language work. It can support research, make it possible to compare solutions, help detect inconsistencies or produce a first version suitable for subsequent post-editing.

The European Commission's Directorate-General for Translation itself uses this kind of technology. The Commission developed eTranslation, and its translators can consult machine translation output as an optional resource.

This example matters precisely because it shows that technology and human intervention do not have to be alternatives.

The use of a tool does not, by itself, determine the quality of the result.

The process into which that tool is integrated does.

How does the European Commission assess translation quality?

The Directorate-General for Translation's approach is particularly instructive.

The Commission describes a fit-for-purpose translation as one that conveys the same meaning and effect as the source while also working naturally in the target language. Its quality system takes account of context, target audience and the risk associated with each document.

Not all errors have the same consequences.

An inaccuracy in a short-lived informational text does not necessarily carry the same risk as an error in a legal act, contract or document intended to produce administrative effects.

That is why higher-risk texts justify more experienced professionals and more extensive quality controls.

The same logic applies outside the European institutions: the translation method should be appropriate to the purpose and risk of the document.

Machine translation itself comes with a warning that it can be wrong

There is a particularly revealing detail.

On its automatically translated Europa pages, the European Commission explains that the quality and accuracy of machine translation may vary considerably between texts and language pairs, and that it does not guarantee the accuracy of machine-translated output. For EU legislation, it directs users to the authentic versions available on EUR-Lex.

This does not mean that machine translation is useless.

It means that machine-translated text and validated translation are not the same thing.

What is post-editing?

When machine translation provides a sufficiently good basis to work from, it can be post-edited.

Professional post-editing does not simply mean making the text more elegant.

It involves checking the translation against the source, correcting errors of meaning and terminology, removing omissions or unjustified additions, ensuring consistency and producing a text that is suitable for the defined purpose.

The importance of this activity is sufficiently recognised for there to be a specific international standard.

ISO 18587:2017 sets requirements for full human post-editing of machine translation output and for the competences of the professionals responsible for that work. The published standard is currently under revision.

Why is this particularly important in legal translation?

A legal document is not just language.

It is language embedded in a conceptual system.

Apparently similar terms may represent different legal concepts in different legal systems. Ordinary words may acquire technical meanings. A modal expression may alter the scope of an obligation. An overly explanatory translation may remove an ambiguity deliberately present in the source.

In such cases, it is not enough to ask whether the chosen solution is linguistically likely.

It is necessary to determine whether it is legally appropriate in that context.

This may require consultation of legislation, case law, institutional terminology databases, official documentation or other specialist sources.

Professional review adds precisely that layer of research and validation.

There is also a confidentiality issue

Before entering a contract, pleading, report, identity document or other client content into an external AI tool, there is a question that comes before translation quality itself:

Whenever personal data are involved, processing remains subject to the General Data Protection Regulation.

Article 5 of the GDPR sets out, among others, the principles of lawfulness, fairness and transparency, purpose limitation, data minimisation and accuracy.

The European Data Protection Supervisor has also published specific guidance on generative AI and data protection, drawing attention, among other things, to purpose, the responsibilities of the entities involved and data-subject rights. The guidance is addressed directly to EU institutions, bodies, offices and agencies, but it is a useful institutional reference for understanding the data-protection issues raised by generative AI.

What about the EU Artificial Intelligence Act?

The European Union also has a specific legal framework for artificial intelligence.

Regulation (EU) 2024/1689 — the Artificial Intelligence Act lays down harmonised rules on AI in the European Union and adopts a differentiated approach according to the risks associated with systems and their uses.

The Artificial Intelligence Act is not, however, a translation-quality standard, and the existence of the Regulation does not mean that a translation produced by a particular system has been linguistically validated.

The quality of the translation still has to be assessed independently.

The translator's real role is changing

For a long time, the traditional image of a translator was someone who personally produced every sentence of the translation.

Today's tools make that description increasingly incomplete.

Translation memories, terminology databases, automated research, machine translation and artificial intelligence can all participate, to different degrees, in the working process.

Professional value is increasingly concentrated in what technology does not solve automatically: selecting, interpreting, checking, contextualising and taking responsibility for the decisions made in the text.

This makes human expertise less visible at some stages of the process, but not less important.

For certain documents, it makes it even more important.

The right question

Simply asking “was this translation produced with artificial intelligence?” now says relatively little about the quality of the result.

There are better questions:

Which tool was used?

For what purpose?

Was the document suitable for that process?

Was the output compared in full with the source?

Was the terminology validated?

Were the context and purpose of the document taken into account?

Was there a professional responsible for the final result?

Whatever technology is used, the essential question ultimately remains very simple: