By Joachim Lépine, M. Ed., C. Tr.
State of play
Today’s neural machine translation engines are impressive, especially at first glance, and they have unquestionably improved in recent years.
I personally had to change several NMT revision exercises in my workshops within the last year because the NMT engine had improved and managed to sidestep several errors produced by the same engine just a few years earlier!
Limitations
That being said, they lack the sensitivity and nuanced understanding that only a human translator can bring.
These engines do not « understand » the text in any meaningful way—for all their mystique and sophistication, they are blind to the subtleties of the situation, the intent behind the message, and the cultural sensitivities and regionalisms that color language. Not to mention the subtext, i.e., the text « between the lines » that writers routinely expect a skilled translator to spot, grasp, and convey.
In spite of the constant barrage of « tech bro » marketing announcing that the « problem » of language and translation has been solved, this prognostication just hasn’t materialized. NMT results are frequently laughable (and even farcical).
The fact is, NMT makes mistakes, so translators have to carefully review every word of the output. An NMT translation can never be fully trusted.
Moreover, machines often « hallucinate » to appear more idiomatic, bypassing and skipping elements they do not understand, and they never ask questions to make sure they’ve understood the intent behind the text. This goes against their fundamental programming: NMT engines are unable to ask questions and LLMs are programmed to sound authoritative even when they have no idea what they’re talking about.
It’s also worth mentioning that poring over a machine-made mess isn’t what the vast majority of translators signed up for. It can be demotivating and even outright depressing. Some translators have left the profession over NMT, and at least anecdotally, many more seem to be jumping ship today in the era of GenAI.
Impact on idiom
One of the most significant challenges of machine translation is idiomaticity. Natural, fluid expression requires deep contextual sensitivity, professional and stylistic judgment, and cultural mastery—qualities that today’s neural machine translation engines lack entirely.
Instead, NMT tends to reinforce non-idiomatic patterns it has frequently encountered, especially those from low-quality materials. Garbage in, garbage out. This can lead to a loss of tone and style; for example, a slogan or motto may be translated with the right words but miss the rhythm, style, or impact of the original.
So, is NMT worth using?
There’s no one-size-fits-all answer; it depends.
In the best of cases, for more routine types of texts, NMT can alleviate the drudgery of starting from scratch. This means less typing and mental exhaustion for the translator. And it turns out that many translators prefer working this way. I personally would not go back to typing out every letter of every word from scratch for most of my translations—especially things like policies and administrative forms.
Incidentally, though, like many translators, I’m receiving fewer of those kinds of texts to translate these days. AI is increasingly obviating the need for more general and boilerplate (« mid-market ») translations.
In the worst of cases, NMT should be avoided altogether since it will not only make the translator’s work longer and more arduous, but yield a poorer text, influenced by a bland, cookie-cutter NMT translation that will leave readers cold. This is especially true at the sentence level, since NMT will almost never significantly change the sentence structure, especially when used inside a translation memory environment (where it will only suggest one translation—the « most likely » one).
The role of translators in the near future
Given these limitations, our role is more crucial than ever. Tomorrow’s translators will need to keep bridging the gap between languages and cultures in order to achieve measurable outcomes and avoid reputational, physical, and financial damage. And they will need to emphasize these things in their marketing so clients understand their value (in a world of “one-click everything”). Interested readers can consult my book, AI Resilient, for down-to-earth tips to help them in this regard.
We will need to work alongside technology, using it as a tool rather than as a crutch, to enhance our work and make sure that the richness of language never gets lost.
The same could be said about writers, by the way.
And here we come up against a limitation of language: When someone says they are translating or writing « with » AI, what does that even mean today? Are they revising a machine draft, or are they consulting AI for inspiration or for tweaks? Those are completely different use cases.
We are going to need a new lexicon in the age of AI. And we are going to need new guideposts and frameworks. Perhaps, as language professionals, we are exactly the right people to tackle this challenge.
Looking ahead
In conclusion, while machine translation continues to evolve, there is zero indication that it can replace the human touch (although it may be able to take care of some routine texts with minimal editing).
As we look to the future, we probably need to accept that in most cases, translation, like writing, is predominantly going to be done « with » AI.
It’s going to be important in this brave new world to explore what « with » means. We may need to create some permutations of “with.”
It is a truism that AI makes a great servant but a terrible master. Tools and workflows that support language professionals are key to handling ever-growing translation volumes. Tools and workflows that try to replace humans or make them an afterthought will only harm clients’ interests, and should be roundly condemned.
Le blogue 