When Kenneth Wehr started managing the Greenlandic-language version of Wikipedia four years ago, his first act was to delete almost everything. It had to go, he thought, if it had any chance of surviving.
Wehr, who’s 26, isn’t from Greenland—he grew up in Germany—but he had become obsessed with the island, an autonomous Danish territory, after visiting as a teenager. He’d spent years writing obscure Wikipedia articles in his native tongue on virtually everything to do with it. He even ended up moving to Copenhagen to study Greenlandic, a language spoken by some 57,000 mostly Indigenous Inuit people scattered across dozens of far-flung Arctic villages.
The Greenlandic-language edition was added to Wikipedia around 2003, just a few years after the site launched in English. By the time Wehr took its helm nearly 20 years later, hundreds of Wikipedians had contributed to it and had collectively written some 1,500 articles totaling over tens of thousands of words. It seemed to be an impressive vindication of the crowdsourcing approach that has made Wikipedia the go-to source for information online, demonstrating that it could work even in the unlikeliest places.
There was only one problem: The Greenlandic Wikipedia was a mirage.
Related Story
Virtually every single article had been published by people who did not actually speak the language. Wehr, who now teaches Greenlandic in Denmark, speculates that perhaps only one or two Greenlanders had ever contributed. But what worried him most was something else: Over time, he had noticed that a growing number of articles appeared to be copy-pasted into Wikipedia by people using machine translators. They were riddled with elementary mistakes—from grammatical blunders to meaningless words to more significant inaccuracies, like an entry that claimed Canada had only 41 inhabitants. Other pages sometimes contained random strings of letters spat out by machines that were unable to find suitable Greenlandic words to express themselves.
“It might have looked Greenlandic to [the authors], but they had no way of knowing,” complains Wehr.
“Sentences wouldn’t make sense at all, or they would have obvious errors,” he adds. “AI translators are really bad at Greenlandic.”
What Wehr describes is not unique to the Greenlandic edition.
Wikipedia is the most ambitious multilingual project after the Bible: There are editions in over 340 languages, and a further 400 even more obscure ones are being developed and tested. Many of these smaller editions have been swamped with automatically translated content as AI has become increasingly accessible. Volunteers working on four African languages, for instance, estimated to MIT Technology Review that between 40% and 60% of articles in their Wikipedia editions were uncorrected machine translations. And after auditing the Wikipedia edition in Inuktitut, an Indigenous language close to Greenlandic that’s spoken in Canada, MIT Technology Review estimates that more than two-thirds of pages containing more than several sentences feature portions created this way.
This is beginning to cause a wicked problem. AI systems, from Google Translate to ChatGPT, learn to “speak” new languages by scraping huge quantities of text from the internet. Wikipedia is sometimes the largest source of online linguistic data for languages with few speakers—so any errors on those pages, grammatical or otherwise, can poison the wells that AI is expected to draw from. That can make the models’ translation of these languages particularly error-prone, which creates a sort of linguistic doom loop as people continue to add more and more poorly translated Wikipedia pages using those tools, and AI models continue to train from poorly translated pages. It’s a complicated problem, but it boils down to a simple concept: Garbage in, garbage out.
“These models are built on raw data,” says Kevin Scannell, a former professor of computer science at Saint Louis University who now builds computer software tailored for endangered languages. “They will try and learn everything about a language from scratch. There is no other input. There are no grammar books. There are no dictionaries. There is nothing other than the text that is inputted.”
There isn’t perfect data on the scale of this problem, particularly because a lot of AI training data is kept confidential and the field continues to evolve rapidly. But back in 2020, Wikipedia was estimated to make up more than half the training data that was fed into AI models translating some languages spoken by millions across Africa, including Malagasy, Yoruba, and Shona. In 2022, a research team from Germany that looked into what data could be obtained by online scraping even found that Wikipedia was the sole easily accessible source of online linguistic data for 27 under-resourced languages.
This could have significant repercussions in cases where Wikipedia is poorly written—potentially pushing the most vulnerable languages on Earth toward the precipice as future generations begin to turn away from them.
“Wikipedia will be reflected in the AI models for these languages,” says Trond Trosterud, a computational linguist at the University of Tromsø in Norway, who has been raising the alarm about the potentially harmful outcomes of badly run Wikipedia editions for years. “I find it hard to imagine it will not have consequences. And, of course, the more dominant position that Wikipedia has, the worse it will be.”
Use responsibly
Automation has been built into Wikipedia since the very earliest days. Bots keep the platform operational: They repair broken links, fix bad formatting, and even correct spelling mistakes. These repetitive and mundane tasks can be automated away with little problem. There is even an army of bots that scurry around generating short articles about rivers, cities, or animals by slotting their names into formulaic phrases. They have generally made the platform better.
But AI is different. Anybody can use it to cause massive damage with a few clicks.
Wikipedia has managed the onset of the AI era better than many other websites. It has not been flooded with AI bots or disinformation, as social media has been. It largely retains the innocence that characterized the earlier internet age. Wikipedia is open and free for anyone to use, edit, and pull from, and it’s run by the very same community it serves. It is transparent and easy to use. But community-run platforms live and die on the size of their communities. English has triumphed, while Greenlandic has sunk.
“We need good Wikipedians. This is something that people take for granted. It is not magic,” says Amir Aharoni, a member of the volunteer Language Committee, which oversees requests to open or close Wikipedia editions. “If you use machine translation responsibly, it can be efficient and useful. Unfortunately, you cannot trust all people to use it responsibly.”
Trosterud has studied the behavior of users on small Wikipedia editions and says AI has empowered a subset that he terms “Wikipedia hijackers.” These users can range widely—from naive teenagers creating pages about their hometowns or their favorite YouTubers to well-meaning Wikipedians who think that by creating articles in minority languages they are in some way “helping” those communities.
“The problem with them nowadays is that they are armed with Google Translate,” Trosterud says, adding that this is allowing them to produce much longer and more plausible-looking content than they ever could before: “Earlier they were armed only with dictionaries.”
This has effectively industrialized the acts of destruction—which affect vulnerable languages most, since AI translations are typically far less reliable for them. There can be lots of different reasons for this, but a meaningful part of the issue is the relatively small amount of source text that is available online. And sometimes models struggle to identify a language because it is similar to others, or because some, including Greenlandic and most Native American languages, have structures that make them badly suited to the way most machine translation systems work. (Wehr notes that in Greenlandic most words are agglutinative, meaning they are built by attaching prefixes and suffixes to stems. As a result, many words are extremely context specific and can express ideas that in other languages would take a full sentence.)
Research produced by Google before a major expansion of Google Translate rolled out three years ago found that translation systems for lower-resourced languages were generally of a lower quality than those for better-resourced ones. Researchers found, for example, that their model would often mistranslate basic nouns across languages, including the names of animals and colors. (In a statement to MIT Technology Review, Google wrote that it is “committed to meeting a high standard of quality for all 249 languages” it supports “by rigorously testing and improving [its] systems, particularly for languages that may have limited public text resources on the web.”)
Wikipedia itself offers a built-in editing tool called Content Translate, which allows users to automatically translate articles from one language to another—the idea being that this will save time by preserving the references and fiddly formatting of the originals. But it piggybacks on external machine translation systems, so it’s largely plagued by the same weaknesses as other machine translators—a problem that the Wikimedia Foundation says is hard to solve. It’s up to each edition’s community to decide whether this tool is allowed, and some have decided against it. (Notably, English-language Wikipedia has largely banned its use, claiming that some 95% of articles created using Content Translate failed to meet an acceptable standard without significant additional work.) But it’s at least easy to tell when the program has been used; Content Translate adds a tag on the Wikipedia back end.
Other AI programs can be harder to monitor. Still, many Wikipedia editors I spoke with said that once their languages were added to major online translation tools, they noticed a corresponding spike in the frequency with which poor, likely machine-translated pages were created.
Some Wikipedians using AI to translate content do occasionally admit that they do not speak the target languages. They may see themselves as providing smaller communities with rough-cut articles that speakers can then fix—essentially following the same model that has worked well for more active Wikipedia editions.
You must be logged in to post a comment.