Any discussion of artificial intelligence in translation must begin by placing it in a broader context. Its implications are global and clearly extend far beyond the world of linguists. AI is everywhere and no sector is immune. From medicine to logistics, from law to journalism, from architecture to human resources, the boundaries of countless professions are being redrawn and their foundations shaken.
In the midst of this technological upheaval — one we are all trying to keep up with as best we can, navigating fake news, hype cycles and more or less alarming social media feeds — my position is a nuanced one. No demonising, no blind enthusiasm. The point is not to deny technological progress, which is real and impressive, and sticking one’s head in the sand would be absurd. The point is to approach these developments with the discernment needed to protect what makes authentic human communication so valuable.
One clarification is in order before diving in: this article is specifically about professional written translation. It is precision work, often carried out on lengthy, complex content — marketing, medical, institutional — with high stakes, demanding consistency, sector expertise and meticulous attention to detail. Professional translation has very little in common with everyday uses: scanning the ingredients on a food product in a Bulgarian supermarket, understanding a product description in French, or ordering a pizza in Lisbon. Where a tool is perfectly adequate for a quick, rough understanding, it can very quickly reach its limits when it comes to defining a strategy, persuading a client, or engaging a company’s legal responsibility
The “Wow” Effect and the Expert’s Eye
First-Read Fascination
At first glance, the output of generative AI is genuinely impressive. The text reads smoothly. The grammar is correct, the vocabulary convincing. Sentences flow with a disarming ease. For the general public, the illusion is perfect — or nearly so.
But when an expert picks up their magnifying glass and applies critical thinking, a phase of disenchantment often follows. It is not always about glaring errors — AI has moved well beyond that. It is more a matter of meaning slipping slightly off course: turns of phrase that sound right but betray the original message. Inconsistent terminology from one paragraph to the next. A tone that misses the target. Metaphors that lose their punch in the target language. A kind of surface-level flatness.
General Translation in the Firing Line
While AI is now capable of translating simple texts satisfactorily for everyday use, it still struggles to replace specialists.
Generalist translators — much like jack-of-all-trades profiles in many other professions — are the first to feel the impact of automation. The reason is straightforward: AI is built on probabilistic models. It predicts the most likely word or phrasing in a given context. The more general the subject matter, the more abundant the training data, and the more acceptable the output. General translation sits, almost by definition, squarely in the machine’s comfort zone.
Is There Still a Market for Straightforward Translation?
Less and less, in reality. What clients are asking for today is no longer simply “translate this text.” They want content that works in the target language — content that converts, that engages, that ranks on Google and registers with AI systems, and that stays true to their brand voice. A translator who does nothing more than transpose words from one language to another, without any editorial or strategic value added, is the most directly exposed to competition from automated tools.
To survive and stand out in this reconfigured landscape, there is only one viable path: knowing how to position yourself and find your niche. The more specialised and expert you become, the less AI can substitute for human judgement.
Translation Domains That Hold Their Own Against AI
Marketing Translation and Transcreation
Marketing is arguably the domain where the gap between machine and human is widest — and most telling. Translating an advertising message, a brand tone of voice or a marketing intent is not simply a matter of converting words from one language to another. It means feeling the cultural resonance of an expression, anticipating how it will land with a specific audience, identifying what captures attention in a given context — and what, on the contrary, risks falling flat or, worse, causing offence.
Transcreation — a term that implies going beyond translation to recreate the intention of a message in the target language — calls for cultural relevance, intuition and empathy with the intended recipient. Humour, wordplay and riddles are linguistic devices that AI struggles to grasp; it shows a marked weakness here and lacks the resources to capture and render nuance with real accuracy.
The internet today is overwhelmingly fed by AI-generated content. Trained on an ever-dwindling supply of human-produced writing, this content tends towards homogeneity. Texts resemble one another. Style is smooth and without originality. In such a context, human creativity becomes a rare and more essential commodity than ever.
Sensitive Domains and Confidentiality
There are other areas that resist AI because they demand extreme and systematic precision and allow absolutely no margin for error: medical and scientific translation, and legal and political translation.
Everything related to data confidentiality must also be protected. Clinical trials, sensitive legal documents, personal medical data, business contracts under negotiation — these are all types of content that simply cannot be sent to remote servers without risking a breach of GDPR or contractual confidentiality obligations. This is a matter of legal and ethical compliance. And in these domains, the human translator — bound by their own confidentiality commitment — remains the only viable option for discerning clients.
Taming the Beast
AI is not a skill one sells in the same way one sells professional expertise. It is a lever. Taming the beast means understanding its logic in order to use it effectively where it adds value — and setting it aside when it does not.
“Waiting to Learn a New Technology Is Never a Good Idea”
This phrase, often heard in tech circles, neatly captures the pragmatism that every professional must adopt when faced with technological change. Waiting for the dust to settle before upskilling is a losing strategy. It is a conviction I internalised early on.
Neither a technophobe nor an uncritical enthusiast, I did not wait for AI to become unavoidable before learning to use it. I did so out of professional pragmatism: to stay in step with market and client expectations, and to be able to evaluate AI output with a critical eye.
Understanding AI is a bit like learning a new logic or a new language. At first, you fumble, you are surprised, you get things wrong. Then you start to grasp the mechanisms, the biases, the blind spots — and you become a professional who knows how to get the best out of the tool without being held hostage by it.
Domain Expertise as an Absolute Prerequisite
In my view, you do not learn to use AI for its own sake. You learn to integrate it in service of a pre-existing profession and expertise.
It is precisely because you are an expert in your field that you have the legitimacy and competence to challenge the machine. To spot immediately where what the tool proposes holds up — and where it falls apart. To restore the right tone to a translation that “sounds good” to the ear but misrepresents the brand.
What Clients Want
AI has profoundly segmented — and in some ways polarised — the translation market.
On one side: raw (or near-raw) machine translation, fast, cheap and acceptable for low-stakes content, alongside a number of low-cost agencies trying to survive by offering rates that many qualified translators rightly refuse. On the other: demanding clients who seek excellence, understand the value of a genuinely professional translation for their brand image and reputation, know what a poor translation can cost them, and are willing to pay for the work at its true value.
Between these two poles, the mid-range is collapsing. Being “average” is not a good position from which to hold your own against AI. What remains for human professionals is the high end — real added value, consultancy, expertise and unimpeachable quality.
Content With Real Value
There remains strong demand for translators and content creators who can bring genuine editorial craft: a distinctive style, a unique voice, consistency and flair. These clients will not settle for approximations. They know their content is an ambassador for their brand, and that linguistic quality is an integral part of their professional image.
From Executor to Strategic Partner
For a long time, a translator’s value was measured in volume: so many words translated, at a given rate per word. A pure output model, where competence was invisible because it was assumed — and frequently undervalued. That model is now being blown apart by tools capable of producing volume at a speed and cost no one could have imagined ten years ago.
What the market now expects is no longer just an output (a translated text), but genuine added value: linguistic insights, a critical reading of the source message, recommendations on tone or communication strategy. The modern translator no longer delivers a document — they deliver a result; they think through with their client how a message should be perceived in another culture, another market, another context.
This is where a real shift in professional posture is unfolding — one that linguists are being called upon to embrace.
AI’s Strengths and Weaknesses in Translation
What It Does Very Well
Generative AI excels at a number of tasks that have genuinely changed the day-to-day working life of linguists.
It is unbeatable for speed. For a rough, overall understanding, what used to take several hours now takes just a few minutes.
It is also enormously helpful for terminology research and consistency checking in long documents — even if it is not always consistent within the same document when translating the same term. That said, when prompted during a quality review, it will do an excellent job of making terminology uniform. Similarly, when fed a style guide, it follows the instructions to the letter (though using a RAG system is preferable for this purpose).
Finally, it is extraordinarily useful for research: synthesising sources, accelerating comprehension of a specialist field, finding examples of phrasing in a target language. Tasks that once required a great deal of time and energy can now be accomplished in just a few well-crafted exchanges.
Where It Falls Short
Where things get complicated is on fine nuances, cultural subtleties and tone of voice.
AI produces the probable. It does not feel. It does not perceive that a formulation, perfectly correct grammatically, might come across as condescending in a given field, or too casual in another. It does not always distinguish irony from sincerity, euphemism from plain statement, intentional lightness from carelessness.
It also hallucinates. Not often, but often enough that you should never trust it blindly — especially with figures, legal references or proper nouns.
There is also a lesser-known but important phenomenon worth raising for translators who incorporate MTPE (Machine Translation Post-Editing) into their workflow: anchoring bias. When a translator starts from a machine-generated text — even a poor-quality one — their brain tends to remain “anchored” to the choices made by the AI. They correct and adjust, but do not start from scratch, and frequently miss more creative or accurate solutions they would naturally have found by translating directly from the source. This bias is quite insidious: the translator has the impression of having thoroughly reviewed and revised the text, when in reality they have been guided by the machine throughout.
For this reason, in many cases and contrary to what one might expect, starting from a blank page remains the best approach for high-value content.
Issues We Cannot Ignore
The Habsburg Effect and the Saturation of the Web
The Habsburg effect in artificial intelligence refers to the progressive degradation of a model when it is retrained on content produced by other AIs rather than on fresh, varied human-generated data. The name draws on the analogy of the royal dynasty whose centuries of inbreeding eventually produced visible negative effects — genetic disorders, health problems and so on. AI faces its own dynastic peril.
Language models feed on textual data extracted from the internet. And the internet today is massively polluted by content generated by AI itself. By feeding on its own output, AI grows impoverished, more standardised, and loses diversity and richness.
What this means in practical terms for translation and content creation: if AI draws less and less on authentic human production, it will drift towards a kind of standardised linguistic mush. And it is there, paradoxically, that human content creators rediscover their full value.
Environmental Issues and the Risk of Losing Expertise
The question of AI is not only an economic one. It is also — increasingly — an environmental one.
The data centres powering generative AI models consume staggering quantities of energy and water. Every query, interaction and piece of generated content carries a real energy cost, often invisible to the end user but very much present in carbon footprints. Some industry experts are sounding the alarm about the potentially unsustainable nature of this model at scale. The risk of an “AI winter” — or the bursting of a technological bubble under the weight of its own energy contradictions — is not science fiction.
But there is another risk, equally concerning: the loss of professional expertise.
If we progressively hand over to the machine all the cognitive tasks that have until now formed the core of our profession — analysis, lexical choices, reformulation, stylistic judgement — what will remain of our expertise in ten or twenty years? Cognitive muscles, like physical ones, atrophy when they are no longer used. And once that human expertise is lost, will there be any way back?
Conclusion: AI — A Great Sidekick, but Still Needing Human Oversight
My conviction, having thought all of this through, is that without the human, AI cannot work wonders. It is the human who guides it, gives it direction, asks it the right questions and knows how to recognise the right answer among the many it offers.
Without AI, I am a less effective professional. My research takes longer. Some terminology checks that now take minutes would take hours. AI has provided me with real, concrete and measurable benefits.
I am also convinced that AI can only be truly beneficial to professionals who already have solid expertise in their field. It frees up their time and energy for what really matters: precision — including tonal precision — creativity, client relationships and strategy. What it does not do is conjure expertise where none exists.
AI is a great sidekick — responsive, tireless, always available. But you always need to go over its work. And to do that well, you need to know your craft in the first place.
