AI Translation Tools Are Quietly Hallucinating False Sources Into Wikipedia

A non-profit organization that pays contributors to translate Wikipedia articles using AI has been caught injecting fabricated citations and hallucinated sources into the world's largest encyclopedia. The Open Knowledge Association, or OKA, used large language models to automate most of its translation work, and the results are exactly as messy as you'd expect.

How the Hallucinations Got Into Wikipedia

OKA's model is straightforward on paper. They pay monthly stipends to contributors who translate existing Wikipedia articles from one language into another. The catch? They rely on AI to do most of the heavy lifting. Contributors run articles through large language models, then theoretically review the output before publishing.

The problem is that LLMs don't just translate. They improvise. When a model encounters a citation in the source article, it doesn't simply convert the reference from French to English. Sometimes it swaps in a different source entirely. Sometimes it invents a page number that doesn't exist. And sometimes it generates entirely new claims that weren't in the original article at all.

Wikipedia editor Ilyas Lebleu (known as Chaotic Enby on the platform) was among the first to flag the issue. While checking a translated article about the French royal La Bourdonnaye family, Lebleu found that a cited book and page number didn't actually discuss the family at all. The AI had hallucinated the citation.

The Scope of the Problem

This wasn't an isolated incident. When Wikipedia editors started doing spot checks across OKA's translations, they found errors scattered throughout. Some articles had swapped sources. Others contained unsourced sentences that appeared out of nowhere. One article about the 1879 French Senate election included entire paragraphs sourced from completely unrelated material.

The editors also discovered that OKA was primarily using contractors from the Global South who were paid low rates. Many of these contractors had poor English skills and weren't catching the AI-generated errors during review. In some cases, the copy-paste nature of the work broke article formatting entirely.

This raises a question that every AI company should be thinking about: when you automate knowledge work with AI, who's actually checking the output? OKA had good intentions -- expanding Wikipedia's reach across languages is a worthy goal. But the execution was sloppy, and the consequences for information quality are real.

Why AI Translation Hallucinations Are Different

Regular AI hallucinations are bad enough. When ChatGPT makes up a legal case or invents a scientific study, that's a problem between one user and one chatbot. But when hallucinations get baked into Wikipedia, they become part of the information ecosystem that billions of people rely on.

Wikipedia isn't just a website people read. It's a source that Google's Knowledge Graph pulls from. It feeds into AI training data. Students cite it. Journalists reference it. When a fabricated citation enters Wikipedia, it can propagate across the entire information landscape before anyone catches it.

The specific type of error here is particularly insidious. It's not that the AI wrote obviously wrong content -- it produced plausible-looking text with real-seeming citations. The book cited in the La Bourdonnaye article actually exists. The page number looks legitimate. You'd have to physically check the source to realize the AI made the connection up.

This is why understanding how different AI models handle accuracy matters so much. Not all language models hallucinate at the same rate, and some are better at preserving source fidelity during translation tasks. Picking the right tool for a given task isn't just a technical decision -- it's an editorial one.

Wikipedia's Defense Mechanisms Actually Worked

Here's the silver lining in this story: Wikipedia's open governance model caught the problem. Volunteer editors noticed suspicious translations, investigated, ran spot checks, and implemented restrictions on OKA contributors. Some translators have been blocked outright.

This is exactly how Wikipedia is supposed to work. It's not a perfect system, but its distributed editing model creates a kind of immune response to bad content. When AI-generated errors entered the system, human editors identified and neutralized them.

The lesson here isn't that AI shouldn't be used for Wikipedia translation. It's that AI output requires meaningful human review, and "meaningful" is doing a lot of work in that sentence. Having a low-paid contractor skim AI output and click publish isn't review. It's rubber-stamping.

The Bigger Picture for AI and Knowledge

This incident is a microcosm of a tension that's playing out across the entire internet. AI tools can dramatically increase the volume of content production. But volume without quality control just creates more noise.

Wikipedia has dealt with this challenge before. They fought spam, vandalism, and promotional editing for years before AI entered the picture. But AI scales the problem in ways that manual abuse never could. One person with an LLM can produce more translated articles in a day than a human translator could in a month. If even a small percentage of those contain errors, the cumulative damage adds up fast.

The Machine Brief glossary defines hallucination as "when an AI model generates content that sounds plausible but is factually incorrect." What the Wikipedia case shows is that hallucinations aren't just a chatbot problem -- they're an information integrity problem that affects every platform where AI-generated content can be published.

What Comes Next

OKA hasn't shut down, but Wikipedia editors are now watching their output closely. New policies require more rigorous spot-checking of AI-assisted translations. Some editors have proposed requiring disclosure of AI use in all Wikipedia translations, similar to how paid editing must be disclosed.

The broader AI translation space will also need to reckon with this. Companies like Google and Meta have invested heavily in AI translation for low-resource languages, arguing that it's a net positive for global information access. That may be true, but only if the quality bar is high enough. Translating a Wikipedia article from French to English and accidentally inventing citations doesn't expand knowledge -- it corrupts it.

For anyone interested in how AI is changing content creation across the web, Machine Brief's learning resources cover the intersection of AI quality, content integrity, and the tools being built to manage both.

FAQ

How many Wikipedia articles were affected by OKA's AI translations?

The exact number hasn't been publicly disclosed, but OKA operates across multiple languages and has been active for an extended period. Wikipedia editors have been conducting spot checks to identify affected articles, and the investigation is ongoing.

Are other Wikipedia translation projects using AI too?

Yes, several Wikipedia translation initiatives use AI tools to some degree. However, the OKA case is notable because AI was the primary translation method rather than a supplementary tool. Wikipedia's community is now discussing broader policies around AI-assisted translation.

Can the hallucinated sources be easily identified?

Not easily, which is what makes this particularly concerning. The AI-generated citations often reference real books and publications but with incorrect page numbers or fabricated connections. Verifying them requires manually checking each source, which is time-consuming and resource-intensive.

Has Wikipedia banned AI-assisted translation?

Wikipedia hasn't implemented a blanket ban on AI translation. Instead, editors are placing restrictions on specific contributors whose work has been found to contain errors. The community is discussing new policies that would require disclosure of AI use and more rigorous review processes for AI-assisted content.