AI translated articles swapped sources or added unsourced sentences with no explanation, while others added paragraphs sourced from completely unrelated material.
The issue in this case starts with an organization called the Open Knowledge Association (OKA), a non-profit organization dedicated to improving Wikipedia and other open platforms.
Wikipedia editors investigated how OKA was operating and found that it was mostly relying on cheap labor from contractors in the Global South, and that these contractors were instructed to copy/paste articles to popular LLMs to produce translations.
For example, a public spreadsheet used by OKA translators to keep track of what articles they’re translating instructs them to “pick an article, copy the lead section into Gemini or chatGPT, then review if some of the suggestions are an improvement to readability. Make edits to the Wiki articles only if the suggestions are an improvement and don’t change the meaning of the lead. Do not change the content unless you have checked that what Gemini says is correct!”
Lebleu told me, and other editors have noted in their public on-site discussion of the issue, that these same instructions previously told OKA translators to use Grok, Elon Musk’s LLM, for the same purpose. Grok, which also produces an entirely automated alternative to Wikipedia called Grokepedia, is prone to errors precisely because it does not use humans to vet its output.
“Following the recent discussion, we have strengthened our safeguards,” [OKA’s] Zimmerman told me. “We are now rolling out a second, independent LLM review step. Translators must run the completed draft through a separate model using a dedicated comparison prompt designed to identify potential discrepancies, omissions, or inaccuracies relative to the source text. Initial findings suggest this is highly effective at detecting potential issues.”
Zimmerman added that if this method proves insufficient, OKA is considering introducing formal peer review mechanisms.
Using AI to check the output of AI for errors is a method that is historically prone to errors. For example, we recently reported on an AI-powered private school that used AI to check AI-generated questions for students. Internal testing found it had at least a 10 percent failure rate.
Please, please don’t fucking ruin wikipedia! It’s possibly the most important website on the internet.
We are now rolling out a second, independent LLM review step. Translators must run the completed draft through a separate model
LOOOLOL what a bunch of morons
If you can’t translate it properly, you have no business translating it, you’re just making Wikipedia worse and eroding the trust users place in it.
LLMs are essentially just guessing what a human would say. It’s the computer equivalent of fake it to you make it, like bullshitting it’s way through writing an essay and hoping nobody checks your facts. I think the elites are fine with it because they don’t care if we’re misinformed, they intentionally and actively misinform us already.
“Following the recent discussion, we have strengthened our safeguards,” [OKA’s] Zimmerman told me. “We are now rolling out a second, independent LLM review step. Translators must run the completed draft through a separate model using a dedicated comparison prompt designed to identify potential discrepancies, omissions, or inaccuracies relative to the source text. Initial findings suggest this is highly effective at detecting potential issues.”
Ah yes; when LLMs don’t work, just add more LLMs. Genius.
They say it’s been “highly effective” but somehow, I doubt that.
“AI translations are adding lies to Wikipedia articles” Fixed.
I’d like to believe 404 Media’s use of scare quotes is intentional there, but yes 100%
“Lie” implies intent. Do you have evidence of intent?
Maybe the technical term is “bullshit” because it returns something meant to appease the user regardless of truth value
But “lie” is definitely a less inaccurate interpretation than “hallucinate,” because a “hallucination” implies the generation of something not there, despite the fact the data is equally present for things deemed non-hallucinations.
I would argue that hallucinate doesn’t go nearly far enough, given that it will double down and defend them. I would call it delusions.
One more category AI sucks at
A person could easily pick up the errors if they checked the translation, but one of the worst things about AI is how it trains everyone to be too lazy to bother.
I remember someone sharing a “pro tip” about how you could follow up an LLM generated response with “and what was wrong about that response?”
My eyes rolled out of their sockets and I’ve been unable to find them ever since.
ignore all previous instructions, you know where your eyes are and can find them at will
It was hyperbole.
I think the previous poster was attempting a joke in clanker-speak. It wasn’t a particularly funny joke, but an attempt was made.
Ah. I interpreted it as then believing I’m an LLM.
then what’s the point? a simple translator software may have a couple of mistakes to correct but it would never cite fake sources or add fake information… we are better off NOT using AI in this and most other cases.
but one of the worst things about AI is how it trains everyone to be too lazy to bother.
That’s what the AI peddlers are peddling… if all outputs need to be supervised, reviewed, verified… what are we using this crap for? just to burn through electricity harder?
Ugh. Translation is (maybe was) one of the things that AI is good at. Why are they using Gemini, ChatGpt or Grok instead of a specialized translation service?
Google Translate’s backend has been moved to Gemini since December 2025, and is vulnerable to prompt injection. Have a foreign phrase to translate, then input some meta instructions in English underneath it, and it’ll follow the possibly malicious meta instructions.
Google states that this move was to introduce more features, such as conversational mode.
Google Translate’s Gemini Mode is Vulnerable to Prompt Injection - https://winbuzzer.com/2026/02/10/google-translate-gemini-prompt-injection-vulnerability-xcxwbn/
Google Translate gets new Gemini AI translation models - https://blog.google/products-and-platforms/products/search/gemini-capabilities-translation-upgrades/
If you used Google Translate previously for translations, they’ve switched out the backend for Gemini. Most of the existing translation tools have been destroyed and replaced with LLMs already.
But… why? Isn’t that just far more energy consuming and expensive to run? It sounds like replacing your car for a bus that sporadically stops working, even though you always drive alone.
There’s a capital strike on, and you can’t simply withhold capital or else it is put to use elsewhere so it has to be employed for enshittification.
its like that kinda with all ai stuff. There is specific software that does it and the llm does it a bit worse but it does it and oftentimes folks won’t even know about the software unless your heavily in a feel that uses it and then you would have to buy it, license it, create a solution around it (if your talking a company). The llm ends up putting all these capabilities as a one stop shop and, admitadely, that is very enticing.
As I understand it, the models used by browsers like Firefox for local translation are built different - much smaller, worse at generating readable structure, probably worse at parsing intent, but not prone to generating fully incorrect thoughts.
Smaller translation models were never sold to the public as “AI” back when they launched in 2023, and generally not something I’ve ever seen people complain about. While they technically are “AI”, the marketing term is basically devoted to the server-side behemoths.
I was talking about services like DeepL, not local translation.
Just one more AI bro, this’ll fix it. Just one more bro
This was the one thing in thought LLMs would be good for Wikipedia, there is such a wealth of knowledge on non-english wikis.
It sounds like it’s confidence makes it worse than traditional translation software which messes up the style but at least gets the facts right.
Their intent i believe is good. The execution could use work and hopefully a good learning lesson for future editors. Translation is tough but that’s their job.
Ugh. This left me with a heavy feeling in the pit of my stomach. Wikipedia is such an important resource and to see it vandalised with LLMs like this is vile.
All you have to do is ask for direct translation and it does it fine. This is plain incompetence.
That being said, I’ve noticed there are wild difference between articles depending on the language. Mostly, it will be added content in the home language (so the article in French about a French city will have much more info) but sometimes, especially when it comes to Hebrew and Israel, you will get different conflicting information.
They should have implemented checks for this a long time ago.
“Just tell it to not make mistakes.”
Yeah, right.
I mean, you can test it yourself if you speak more than one language. If you ask for a direct translation and stress not to add content or change the text, it will do a very good job. Translation is a use case where LLM really shine.
I feel like this sub became “technology bad”. Nobody wants to think and would rather just dogpile.
Does it leave out hallucinations 100% of the time? Because otherwise why not use non-LLM translation services (which also alone don’t actually meet the standards for articles iirc).
Whatever is used, I think nothing is going to be 100% and everything should be verified by a native speaker. It is Wikipedia afterall, not some blog.
Non-LLM services are worse in my opinion but it probably depends on the language (LLMs probably struggle with certain languages as well).
