Author, John Laurenson
Role, business affairs reporter
Report source: Paris, France
Posted 4 minutes ago
Reading time: 3 minutes
The French say "Il pleut des cordes" when it rains heavily.
The literal translation is "It's raining ropes". Such idioms often stump artificial intelligence translation.
“You can imagine the dilemma this would pose for artificial intelligence,” said François Yvon, director of research at the French National Center for Scientific Research (CNRS) specializing in automatic translation.
He said that a well-trained system will know that the rope is not really falling from the sky, so it will use the English equivalent "it's raining cats and dogs" (heavy rain).
Yvonne asked: "But when it is translated into Indonesian, can you find the equivalent?"
Commenting on the weather is relatively low-stakes, but now more and more translation work is done by machines, and the content involved is far more important than the weather.
Google said in April this year that people now use Google Translate to translate one trillion words per month.
This does not count other tools: DeepL, Microsoft, Claude, ChatGPT and Gemini, not to mention translation systems tailored for enterprises.
Skip hot reading and continue reading
hot reading
End of Hot Reading
Last year, direct sales of machine translation (i.e. excluding general AI subscription services) were estimated at $1.4bn (around £1bn).
John Zimmerman, founder of the Open Knowledge Association (OKA), a Swiss non-profit organization, said: "In many situations, directly using artificial intelligence translation is the right choice. If you are translating internal documents or product catalogs, it is not reasonable to increase the cost 100 times just to correct the last few percentage points of errors."
"But encyclopedias are different. Wikipedia is built on trust. A false fact can exist for years and be cited by journalists and artificial intelligence systems."
His organization is responsible for translating Wikipedia articles into other languages. It has published more than 10,000 articles, mainly translated into Spanish, Portuguese and English. It is the single organization that contributes the most translated articles to this collaborative online encyclopedia.
And without artificial intelligence, none of this would be possible.
Zimmerman explained: "If we use traditional methods to translate 10,000 articles, the cost will be as high as millions. But in our process, artificial intelligence first produces a first draft, which is then checked sentence by sentence by human translators."
However, human involvement is becoming increasingly rare.
"The quality has now improved to a point where skipping human translation entirely is a serious option and a common temptation in the industry," Zimmerman said. "Two or three years ago, when using tools like DeepL, our translators still had to rewrite a significant portion of each first draft."
"Today, the literal expression of many translations often does not need to be modified at all. Therefore, the question is no longer whether AI translation is good enough, the answer is obviously yes; the real question is whether it is good enough to replace human translators."
However, translators believe that this is not yet the case.
Ariella Baltié is a translator and legal expert who specializes in translating court documents from English into French.
Someone recently contacted her and said that she had used artificial intelligence to translate her own documents and asked her to help with proofreading.
"There were two big mistakes at the beginning," she said.
Artificial intelligence translates the French word "Le demandeur" (the plaintiff) into "the requester" (the requester).
"This is a very inappropriate translation for a legal document," Balti said.
When the AI sees "Madame X, épouse Y" (Ms.
In other words, it mistakes a noun for a verb, completely distorting the original meaning.
Mistranslation of legal categories can have serious consequences. In the United States, for example, errors in machine-translated asylum applications have led to delays in the release of detainees.
In the medical field, accuracy is a matter of life and death.
"The units for drug dosages and blood test results often contain Greek letters," says Carol Bereuter, a medical translator.
She once encountered an automatically translated text that translated the Greek letter μ into milli (one thousandth) instead of micro (one millionth).
"This is a serious mistake. If you give a milligram instead of a microgram, the patient is getting a thousand times the dose."
She also noted that as AI translations get better, it becomes more difficult to spot errors.
"In the past, when using tools such as Google Translate, sentences would be awkward because there would be grammatical errors, or they would read awkwardly; now they read completely reasonably, very smoothly and naturally, so errors are harder to detect."
Francois Yvonne, a French scholar who specializes in automatic translation, warns: "If you sign the translation, there is a risk to your reputation. The risk is even greater if someone decides whether to undergo a medical procedure, sign a contract, or make a major purchase based on the content of the translation."
Even sentences that seem clear to humans can stump an AI translation service, he said.
“When a word has different meanings, humans usually have a good idea which one applies because they understand the context; but when a machine translator encounters the sentence ‘translate into Polish’, it’s entirely possible that it is talking about shoe polish (Polish has two meanings: ‘Polish’ and ‘polishing agent’).”
Context is often crucial. If the AI translator doesn't realize that we're talking about a football match, it might mistranslate "taking a corner."
Last year, OKA conducted a study funded by the Wikimedia Foundation, covering 119 articles, 10 language pairings, and more than 1,000 hours of translator work.
OKA founder Zimmerman said: "27% of artificial intelligence-generated text is still modified before publication. Three-quarters of these modifications correct errors introduced by artificial intelligence, and 5% to 6% of these errors change the meaning of the original text."
It can be quite difficult to completely eliminate these errors.
The reason is that artificial intelligence determines which word to translate instead of another word based on probability; and "probably correct" does not mean "really correct", especially in a field as delicate and full of changes as language.
He believes that a new generation of technology will bring improvements.
