AI translation errors can raise a drug dose a thousandfold — human checking is still essential

In French, when it rains very heavily, people say "il pleut des cordes". Translated word for word, it means "ropes are falling from the sky". But its real meaning is "it is raining cats and dogs". Can artificial intelligence, or AI, understand the correct meaning of such a phrase? It can, but not always. And that is exactly where humans are still needed in translation.

As technology improves, AI is being used more and more in translation. With services like Google Translate, DeepL, ChatGPT and Gemini, a text in one language can be translated into another in a moment. According to Google, about one trillion words are translated every month using its translation service. Such an enormous volume of work is impossible for humans alone.

So will translators one day become unnecessary? Translating with AI saves both time and money. The Swiss non-profit "Open Knowledge Association" translates Wikipedia articles into different languages. Its founder, John Zimmerman, says that without AI, translating 10,000 articles would have cost millions of dollars.

In their method, AI first produces a draft translation. Then a human checks every sentence. A few years ago, AI translations had many mistakes, so translators had to rewrite quite a few sentences. Now, in many cases, the translation is so fluent that no separate correction is needed.

Yet not all texts are the same. In a product description or an ordinary internal company document, a small mistake may matter little. But in an encyclopaedia, a medical text or a court document, a single mistake can cause great harm.

Ariela Balter translates court documents from English into French. Recently, a man had some of his documents translated by AI and sent them to her for checking. Right at the start, she found two major mistakes.

The French word "le demandeur" means the plaintiff in a case. But the AI had translated it as "the requester". That does not convey the correct meaning in a legal document. Again, "Madame X, épouse Y" means "Mrs X, wife of Y". The AI had translated it as if a woman named X were marrying Y. In other words, the machine used a word that describes an identity to describe the act of marriage.

Such mistakes in legal documents can have serious consequences. In the United States, there have been cases where detained people’s release was delayed because machine translation of their asylum applications contained errors.

In medicine, a wrong translation is even more dangerous. Medical translator Carol Berriuter once saw a translation in which the Greek letter "μ" (mu), meaning "micro", had been mistakenly written as "milli". The two are not the same at all. One milligram is a thousand times one microgram. So if such a mistake is made while stating a drug dose, a patient could receive a thousand times more medicine than prescribed. That puts life at risk.

Another problem is that AI translations are now much more fluent. Earlier, a wrong translation had grammar problems or sounded unnatural. Now many wrong sentences read perfectly naturally. So even more attention is needed to catch the mistakes.

To translate a sentence correctly, its context must be understood. The same word can mean different things in different situations. In football commentary, "taking a corner" means taking a corner kick. But without understanding the sporting context, the translation can go wrong. The French phrase "il pleut des cordes" shows the same thing. Translating the words literally is not enough; you must understand what the phrase actually means.

Research has also been done on this. Funded by Wikimedia, the Open Knowledge Association analysed 119 articles, 20 languages and more than a thousand hours of translators’ work. The study found that 27 percent of the AI-generated text had to be corrected before publication. Three-quarters of these corrections were mistakes made by the AI. And 5 to 6 percent of the mistakes changed the meaning of the original text entirely.

According to François Yvon, a researcher in automatic translation at France’s National Centre for Scientific Research, it is hard to imagine AI translation systems becoming one hundred percent reliable. That is because AI often chooses words and sentences on the basis of probability. Just because a translation is likely to be correct does not mean it certainly is.

Researchers hope, however, that future translation systems will be better. A system could be built in which one AI translates, another finds the mistakes in that translation, and a third produces a corrected version. In this way, several AIs working together could raise the quality of translation.

But in Yvon’s view, the greatest progress will come when AI does not just stop at translating, but also tells humans where it is unsure. So will humans no longer be needed in translation? For ordinary texts, human help may gradually be needed less. But where a single wrong word can cause great harm, human checking remains extremely important. Because language is not just the meaning of words; it also involves context, culture, feelings and human experience.