Feed 19 repetitions of the word “dog” to Google Translate and ask it for a Maori conversion and you get this: “Doomsday Clock is three minutes at twelve. We are experiencing characters and a dramatic developments in the world, which indicate that we are increasingly approaching the end times and Jesus’ return.”
There’s lots more where this came from, especially when translating into languages with small corpuses like Maori or Somali, and no one’s exactly sure why. The smart money is on “neural machine translation,” which is producing the literary equivalents of the furry animals that appear in images that are repeatedly processed through Deep Dreaming.
Sean Colbath, a senior scientist at BBN Technologies who works on machine translation, agreed that strange outputs are probably due to Google Translate’s algorithm looking for order in chaos. He also pointed out that the languages that generate the strangest results—Somali, Hawaiian and Maori—have smaller bodies of translated text than more widely spoken languages like English or Chinese. As a result, he said, it’s possible that Google used religious texts like the Bible, which has been translated into many languages, to train its model in those languages, resulting in the religious content.“If they tried to build a model out of that stuff, it may be that the model simply throws a hail-mary pass (pun semi-intended) and barfs out a piece of its training,” said Colbath, who emphasized that he was speaking for himself rather than for his employer, in a chat.
Rush agreed that if Google is using the Bible to train its model, it could explain some of the strange outputs. Indeed, several of the bizarre translations in Somali resemble specific passages from the Old Testament. Exodus 27:18 references a hundred cubits, and several verses including Numbers 3:18 discuss the sons of Gershon.
Why Is Google Translate Spitting Out Sinister Religious Prophecies? [Jon Christian/Motherboard]