“Now I do believe that machines can affect my profession.”

PalmJust fifteen years ago, translators weren't worried that tools like Google Translate could replace them. "It was so bad there was nothing to fear," recalls Jean-François Cuennet, a professional with over three decades of experience. "Now I do believe that machines can affect my profession," he confesses. The novelty isn't so much the technology itself as the quality of the translations: less visible errors, more fluid texts, and a level of professional rigor that is now essential. As Antoni Oliver, an expert in machine and computer-assisted translation and a professor at the UOC (Open University of Catalonia), points out, "since the 1950s, the idea that machine translation will eventually replace translators has been repeated." Every major technological advance has been accompanied by the same prediction, but now, Oliver warns, there's a difference: "The improvement is such that the error is less perceptible." This makes the expert's human experience indispensable, searching for nuances in a law, in medical research, and, of course, in literary translation.

Machine translation didn't begin with ChatGPT or the major language models. In the 1950s, the first systems functioned as mechanical extensions of the dictionary: lists of equivalences and grammatical rules that attempted, in a rudimentary way, to imitate the workings of a language. Rigid and limited, they were only useful in very specific contexts and were incapable of handling ambiguity and style. Those early promises quickly clashed with reality: language cannot be reduced to simple rules.

In the 1990s, with machine translation, machines stopped obeying rules and began to learn from millions of translated texts. The system didn't understand what it was saying, but it was often correct through sheer repetition. Quality improved, although errors were still evident, especially in complex texts.

The real leap forward came in 2014 with neural machine translation. For the first time, systems began to take context into account, not just isolated phrases. This resulted in more fluid and coherent texts, especially between closely related languages. Google Translate and DeepL went from being a novelty to becoming professional tools.

Natural, but deceptive

The current phase goes a step further. With generative models, the machine no longer just translates, but produces text, reformulates sentences, and adapts styles. The result is more natural but also more deceptive: the error is no longer visible, and only an expert reader can detect it.

The professional consequences are already palpable. Surveys published by The Guardian and The Brussels Times indicate that more than a third of translators have lost commissions or income due to the growth of AI. At the same time, most acknowledge using these tools regularly. "The market changes, but it doesn't disappear," Oliver clarifies. Simple, repetitive, or low-risk texts have stopped coming in; now, complex, technical, or sensitive commissions are piling up, in which human judgment is still essential.

Cuennet confirms this trend in the Swiss market, where he translates from German to French for government agencies, NGOs, and companies. "The texts we receive now are difficult; the easy ones are already being machine translated," he explains. Although the workload has decreased, the intellectual demands are greater. In this new scenario, post-editing has become a central task. The translator no longer always starts from a blank page: they review, correct, and validate texts generated by a machine. The profession demands extreme attention. "Finding a machine's errors is not easy," Oliver points out. Detecting subtle flaws in meaning, nuance, and context can be more complex than translating from scratch and is often less well-paid, under the logic that "the machine has already done almost everything." The translator's value lies not in the visible changes, but in detecting what isn't working and assuming final responsibility for the text. And there is still one aspect that AI doesn't master: cultural adaptation. “Artificial intelligence doesn’t know how to acculturate a text,” says Cuennet. Translation isn’t just about transferring words from one language to another, but about writing for a specific reader, with a specific mental and cultural framework. “Sometimes you have to explain a concept; other times, adapt it or even omit it,” he adds. These decisions don’t appear in the data the machine is trained on: they require judgment, experience, and sensitivity.

Ethical Dilemmas

AI also raises ethical dilemmas. Vast amounts of text, many copyrighted, have been used to train these systems. "The internet doesn't mean everything can be used freely," Oliver reminds us. Transparency with the client is also at stake: it's necessary to disclose if the text has been machine-translated and subsequently reviewed. In any case, the ultimate responsibility lies with the expert: "If a translator reviews an AI-generated translation, the responsibility is theirs," Oliver emphasizes, even if the error is the machine's. Neither Oliver nor Cuennet believe the profession will disappear. They agree that only those who adapt and maintain high standards will survive. "There's work for humans, without a doubt," Oliver asserts, "but it has to be done well." Cuennet sums it up this way: "If my clients appreciate well-written texts, I'll have work. If they only think about money, I won't." Between technological enthusiasm and the fear of being replaced, translation continues to depend on the same thing it did decades ago: someone willing to read carefully and be accountable for every word.