Facebook says it automatically translates about 4.5 billion posts each day. With new research and advancements, the technology keeps improving, but it’s not perfect—and its flaws can have serious consequences.
Facebook $META says it automatically translates about 4.5 billion posts each day. With new research and advancements, the technology keeps improving, but it’s not perfect—and its flaws can have serious consequences.
Last week, a Palestinian man working at a construction site in an Israeli settlement in the West Bank, posted on Facebook a photo of himself in front of a bulldozer, with the words “Good Morning” in Arabic. The platform’s automatic translation mistook the words for “attack them” in Hebrew, and “hurt them” in English.
Join 500,000+ readers who start their day with Quartz.
By subscribing, you agree to our Terms of Service and Privacy Policy.
Here’s an explanation from the Israeli daily newspaper Haaretz, which first reported the incident, about the nature of the mistake:
Arabic speakers explained that English transliteration used by Facebook is not an actual word in Arabic but could look like the verb “to hurt” – even though any Arabic speaker could clearly see the transliteration did not match the translation.
Google $GOOGL Translate also mistranslates the phrase, turning it into “become them,” likely because the literal translation of “morning” in Arabic is “day becoming.”
The Jerusalem Post reports that Israeli police got multiple complaints from civilians about the post, and that they acted upon it because of previous terrorist attacks using a bulldozer. Reportedly, no officer who spoke Arabic saw the post before the arrest was made. The police released the man from custody after several hours, realizing the mistake.
Facebook introduced its translation tool in 2011. Then, it used Microsoft $MSFT’s Bing to translate. In late 2015 the platform completed a shift to its own AI translation technology. In 2016, users started being able to have their statuses automatically translated, visible to everyone in the language they post in. In August 2017, the platform fully transitioned from phrase-based translation to the more accurate neural network system. The company is developing a more sophisticated neural technology, which it says is nine times faster than its competitors.
Recently, another case of mistranslation by Facebook resulted in “Mexicanitos,” meant to be a term of endearment, being posted as the slur “wetbacks.” The mother from El Paso who posted the photo of her kids, calling them “Mexicanitos,” complained publicly.
