Can AI write Wikipedia? The battle over AI content

Wikipedia has a problem. Actually, Wikipedia has always had problems. Vandalism, bias, edit wars over whether a hot dog is a sandwich. But the current problem is different in kind: AI language models can now generate text that reads like a decent Wikipedia article, and a growing number of people are using them to do exactly that. Some of those contributions are helpful. Many are not. And Wikipedia's volunteer editors are caught in the middle, trying to figure out where to draw lines that the technology keeps moving.

What Wikipedia actually says about AI

In February 2024, Wikipedia's English-language community formally adopted guidelines on the use of large language models. The policy, developed through months of community discussion, doesn't ban AI outright. Instead, it puts the responsibility on the human editor.

The key word is "responsible." Wikipedia already requires that every factual claim be supported by a reliable, verifiable source. An AI can generate a paragraph about the history of Latvian folk music, but if none of the statements are backed by real citations, it doesn't matter how well-written the text is. It violates Wikipedia's core policies.

This is the part most people miss when they talk about AI on Wikipedia. The challenge isn't whether AI can write encyclopedia-quality prose. It can, mostly. The challenge is that Wikipedia isn't just prose. It's a system of verifiable claims linked to real-world sources. And AI models have a well-documented habit of fabricating sources, inventing plausible-sounding citations that point to journal articles, news stories, or books that don't exist.

The Scots Wikipedia disaster (a cautionary tale before AI)

To understand why Wikipedia's community is so nervous about AI content, it helps to know about the Scots Wikipedia incident. This happened before the current wave of AI tools, but it foreshadowed the same problems.

In August 2020, a Reddit user discovered that the Scots Wikipedia (the version written in Scots, a Germanic language spoken in Scotland) had been largely written by a single American teenager from North Carolina. Over seven years, this person had created or edited roughly 27,000 articles, writing in what amounted to English with a few Scots-sounding words sprinkled in. He didn't speak Scots. He was essentially making it up.

The scale of the problem was staggering. By the time it was discovered, the Scots Wikipedia contained about 60,000 articles, and roughly half of them had been created or significantly edited by this one person. Native Scots speakers who found the Wikipedia described it as insulting, like a parody of their language. The entire project's credibility was destroyed.

What makes this relevant to the AI debate is the failure mode. One person, working persistently over years, generated enormous quantities of content that looked superficially legitimate but was fundamentally flawed. Nobody caught it because few people were paying attention to the Scots Wikipedia, and the content looked "close enough" to pass casual inspection.

AI tools enable the same failure mode at much greater speed.

AI hoaxes and catches: 2023-2025

Since ChatGPT's launch in late 2022, Wikipedia editors have caught a steady stream of AI-generated content. Some cases made headlines.

In 2023, editors identified a series of articles about obscure historical topics that contained fabricated events, people, and sources. The articles were well-written and internally consistent. They just weren't true. One described a supposed medieval battle in Portugal that never happened, complete with invented commanders, casualty figures, and a fake citation to an academic journal.

In early 2024, Wikipedia editor "Headbomb" (a well-known volunteer who specializes in detecting problematic content) flagged over 100 articles suspected of being AI-generated. Many contained what editors call "hallucinated references" (citations that look legitimate but point to nonexistent publications). Some had been on Wikipedia for months before anyone noticed.

By mid-2024, the pattern was clear enough that Wikipedia created dedicated workflows for handling suspected AI content. Editors developed checklists: check every citation, look for characteristic AI phrasing, verify that the subject of the article actually exists through independent sources.

The chart above shows Wikipedia's steady growth over two decades. The question now is how much of future growth will be AI-assisted, and whether that assistance will help or hurt quality. The lighter bars represent the pre-LLM era, the darker bars the post-ChatGPT period.

The arguments for AI on Wikipedia

There are legitimate reasons to think AI could benefit Wikipedia. The encyclopedia has significant coverage gaps. Topics related to the Global South, women in history, non-English-language cultures, and many scientific fields are underrepresented. Wikipedia has about 6.8 million English articles, but many topics that have sources and deserve articles simply don't have anyone willing to write them.

AI could, in theory, help by drafting initial versions of articles on underserved topics that human editors then verify and improve. Some editors already use AI this way. They'll ask a model to generate a rough draft, then spend their time on the parts humans do best: checking sources, ensuring neutrality, adding context that requires genuine understanding.

There's also a translation argument. Wikipedia exists in over 300 languages, but the vast majority of articles are only available in a handful. AI translation tools, while imperfect, could make knowledge accessible to millions of people who currently can't read the English version. The Wikimedia Foundation has acknowledged this potential, though carefully.

And some maintenance tasks benefit from AI assistance. Fixing grammar, standardizing formatting, identifying articles that need updating, these are tedious jobs that AI can handle well with human oversight.

The arguments against

The case against AI on Wikipedia is strong, and most experienced editors I've talked to lean this way.

The first problem is sourcing. Wikipedia's entire model depends on verifiable citations. If I write that a bridge was built in 1903, there needs to be a source saying so. AI models don't check sources. They generate text that sounds like it should have sources. When pressed to provide citations, they fabricate them with frightening specificity: real-sounding journal names, plausible author names, correct formatting, and DOIs that lead nowhere. An editor who doesn't check every single citation will miss these.

The fundamental issue is that our readers trust us because every claim can be verified. AI-generated text that looks right but isn't verifiable is worse than no text at all, because it has the appearance of reliability without the substance.

The second problem is scale. One person writing bad Scots articles was a disaster, but at least it was one person, eventually caught. AI enables thousands of people to generate thousands of plausible-looking articles simultaneously. Wikipedia's editing community, roughly 40,000 active editors for the English version, is already stretched thin. A flood of AI content that requires human verification to catch errors could overwhelm the system.

The third problem is more philosophical: original research. Wikipedia has a strict policy against original research, meaning articles should synthesize existing published sources, not generate new analysis. But AI doesn't synthesize sources. It generates text based on patterns in its training data. The output often looks like synthesis but is actually closer to original composition. It's creating something new that happens to sound authoritative, which is precisely what Wikipedia's policies were designed to prevent.

How Wikipedia detects AI content

Wikipedia editors use a combination of methods to identify AI-generated articles, and they're honest about the limitations.

Automated tools like GPTZero and similar detectors have high false positive rates and are considered unreliable for individual articles. Wikipedia doesn't use them as definitive evidence. Instead, editors look for patterns that signal AI involvement.

Behavioral signals matter most. An editor who suddenly starts producing five long, well-formatted articles per day on topics they've never edited before raises flags. An account that creates multiple articles in a short period, all with similar structure and phrasing patterns, gets noticed. Wikipedia has sophisticated tools for tracking editing patterns, and unusual productivity is often the first sign.

Content-level signals include: citations that don't check out, text that's suspiciously well-organized for a first draft, information that's accurate at a general level but wrong on specifics (a classic AI pattern), and phrasing that matches known AI tendencies. Experienced editors develop an eye for prose that's "too smooth," similar to the skill Bluffpedia players develop for spotting our AI-generated summaries.

Wikipedia also relies on its community structure. Articles go through various review processes (new page patrol, recent changes patrol, good article review), and at each stage, human eyes check the content. The system isn't perfect. AI-generated articles have survived for months. But the multi-layered review process catches most problems eventually.

What this means for the future of knowledge

I find this debate interesting because it's not really about AI. It's about trust.

Wikipedia works because millions of people trust it to be roughly accurate. That trust was built over 25 years by a community of volunteers who check each other's work. It's imperfect trust. Wikipedia has errors, biases, and blind spots. But it's trust grounded in a transparent process: anyone can check the sources, anyone can see the edit history, anyone can participate.

AI threatens that trust not by being wrong (humans are wrong plenty), but by being wrong in ways that are hard to detect. A human editor who makes a mistake usually makes a specific, catchable error. An AI that generates a plausible-sounding article with fabricated sources creates something that looks trustworthy but isn't, and catching the problem requires checking every claim individually.

The optimistic view is that Wikipedia's community will adapt, the way it has adapted to every previous challenge. Vandalism, paid editing, content disputes, the community has developed processes and tools for all of these. AI is a new challenge, but the underlying approach (verifiability, transparency, community review) still works.

The pessimistic view is that AI content generation scales faster than human content review, and eventually the volume of plausible-but-unverified text will outpace the community's ability to check it. This hasn't happened yet. But the tools keep getting better, and the barrier to generating convincing text keeps dropping.

Where Bluffpedia fits in

We think about this a lot, for obvious reasons.

Bluffpedia is basically a miniature version of what Wikipedia editors do every day. You're presented with text that might be real or might be AI-generated, and you have to figure out which is which. The skills are the same: checking for specificity, looking for verifiable details, noticing when something sounds too polished or too vague, distrusting your initial feeling of "that sounds right."

The difference is that we turned it into a game. Wikipedia editors do it as unpaid labor, which is both admirable and slightly insane. We give you points and achievements.

But the underlying cognitive skill (telling real information from convincing fakes) is the same one that will matter more and more as AI-generated text becomes a larger part of the information we all consume. Every round of Bluffpedia is practice for a world where you can't take text at face value, no matter how authoritative it sounds.

Wikipedia's editors have been doing this work for 25 years, mostly without recognition. They deserve more credit than they get. And if playing our game makes even a few people better at thinking critically about the text they encounter online, we'll consider that a win alongside the fun.