Quick Answer: Wikipedia handles AI-generated content mainly through human review. Volunteers in WikiProject AI Cleanup look for fluent but unsupported prose, promotional wording, and citations that don't check out. Their core test is verifiability: if a claim can't be traced to a real, relevant source, it gets removed, whatever wrote it.
Wikipedia has roughly two decades of practice keeping anonymous contributors honest, and that experience now faces a new problem. Large language models produce text that reads like a competent encyclopedia entry. The grammar is clean, the tone is measured, and the structure looks right. Yet the facts may be invented, and the sources may not exist. As IEEE Spectrum reported in an interview with WikiProject AI Cleanup cofounder Ellia Lebleu, the editors who do this work have had to learn a different kind of attention. Here's how it works and why it's harder than it sounds.
Why AI Text Is Harder to Spot Than It Looks
Early AI-generated text was easy to flag. It was repetitive, awkward, or obviously off-topic. Current models rarely make those mistakes. The sentences flow, the paragraphs are balanced, and the article reads like a thoughtful summary.
That's the problem. Most editors were trained to notice sloppy writing as a warning sign, and polished prose now lowers their guard. A paragraph can be grammatically perfect and still assert something no source supports.
Here's where most people go wrong: they look for bad writing. The better question is whether each claim is anchored to something checkable. A smooth paragraph with no traceable source deserves more scrutiny than a clumsy one with a solid reference.
The Tells Editors Actually Look For
Lebleu's group works from accumulated observations rather than a single detection rule. The patterns below reflect common features of LLM-generated articles that editors have documented. None is proof on its own. Human writers produce all of them too.
Editors typically check for:
- Citations that don't match the claim. The reference exists, but the cited page says something else, or says nothing about the topic.
- Sources that don't exist. Titles, authors, journal names, and DOIs that look plausible but return nothing when searched.
- Stock phrasing. Stacked adjectives, vague significance claims ("a pivotal role in shaping..."), and summary sentences that repeat the paragraph's point without adding information.
- Promotional drift. Language that praises a subject, company, or person in the way marketing copy does, without the neutral tone Wikipedia requires.
- Confident gaps. Specific-sounding dates, figures, or names with no citation, presented as settled fact.
The third and fourth signs are the easiest to see and the weakest evidence. Plenty of careful human editors write in a formulaic style. The first two signs carry more weight because they can be tested directly.
Fake Citations: The Failure Mode That Matters Most
If you take one thing from this topic, make it this: the most damaging AI error on an encyclopedia isn't a clumsy sentence. It's a confident claim with a fabricated or misapplied source.
Here's a concrete case. An article states that a regional water authority adopted a particular treatment standard in a given year, and cites a trade journal article. The journal exists. The volume and page number look right. But the article is about a different country and never mentions the standard. To a casual reader, the citation adds credibility. To an editor who opens it, the claim is unsupported.
The reason this is so hard to catch is that checking every citation takes real effort. A single article can carry dozens of references, and a volunteer usually has limited time. A tool that generates plausible references faster than humans can verify them creates a bottleneck that didn't exist before.
This is why Wikipedia's verifiability policy matters so much here. The rule doesn't ask whether the text is accurate. It asks whether a reader can confirm the claim from a reliable source. That's a more useful standard for AI-era review, because it shifts the question from "does this sound right?" to "where does this come from?"
Detection Tools Aren't the Answer (Yet)
Many people expect a detector to solve this, and the appeal is obvious. Paste in an article, get a probability score, move on. In practice, the approach has serious limits.
Detectors measure statistical features of text, and those features overlap heavily with careful human writing, especially in technical or formal genres. A high score doesn't tell you who wrote the text or whether the facts are correct. A low score doesn't clear anything. An article can pass a detector and still contain invented citations.
Compare the main approaches:
| Approach | What it catches | Main weakness | Best use |
|---|---|---|---|
| Automated AI-text detectors | Statistical patterns typical of model output | False positives on formal human prose; can't check facts | Triage only, to prioritize review |
| Stylistic checklists | Stock phrasing, promotional tone, filler | Many humans write this way too | Supporting evidence, never sole grounds |
| Citation verification | Fake references, misapplied sources | Time-consuming for long articles | Strongest direct evidence |
| Edit-history review | Sudden large additions from new accounts | Doesn't prove AI use alone | Context for deciding where to look |
The practical pattern is to use detectors and style checks to decide where to look, then rely on citation checks to decide what to do. That ordering matters. Acting on a detector score alone would wrongly penalize good-faith contributors whose writing happens to look machine-like.
Here's where it gets interesting: the strongest evidence is the one that takes the most time. That's a real structural tension, and it's why volunteer groups focus on verification rather than classification.
What Happens When AI Text Is Found
Finding AI-generated content doesn't automatically mean deletion. Wikipedia's response depends on the problem. Some text can be fixed by removing unsupported claims and keeping the sourced portions. Other articles can't be rescued because the subject itself lacks reliable coverage.
Editors generally work through a sequence:
- Check whether the core subject is notable and has reliable independent sources.
- Verify the citations, starting with the ones supporting the most specific claims.
- Remove or rewrite unsupported statements, with an edit summary explaining why.
- Escalate to deletion discussions only if the article can't be salvaged.
That last step is where practitioners disagree. Some argue that any undisclosed AI-generated article should be removed quickly, since the cost of checking every sentence is too high. Others argue that removing text because of its style punishes contributors who used AI honestly to tidy their prose, and that the verifiability test should decide outcomes, not authorship. The source doesn't settle this, and reasonable editors land in different places. I'd lean toward the verifiability approach because it's the only one that doesn't depend on guessing how text was produced. Still, the bulk-removal camp has a fair point about volume.
Where Practitioners Disagree
A few tensions show up repeatedly in discussions of this work, and they're worth naming honestly.
First, stylistic evidence is unreliable on its own. Calling an article "obviously AI" can be a convenient shortcut that misfires on skilled human writers. Editors who rely only on feel make mistakes, and accusations of AI use can damage good-faith contributors.
Second, disclosure rules are hard to enforce. If a contributor says they used AI and cites real sources, the question becomes whether the sources hold up. If they don't disclose, the same problem exists, but there's less context. Either way, the work of checking falls on volunteers.
Third, the technology keeps changing. Patterns that identified AI text two years ago may not identify it now. That's why a checklist-based approach ages badly, and why the citation test has held up better over time.
Frequently Asked Questions
What is WikiProject AI Cleanup?
WikiProject AI Cleanup is a volunteer effort on Wikipedia that works to identify and fix AI-generated content that doesn't meet the encyclopedia's standards. Its cofounder Ellia Lebleu described the group's focus on verification and editing in an interview with IEEE Spectrum. Its members share observations about common patterns and work through articles that show signs of unsupported AI text.
How do editors spot AI-generated content on Wikipedia?
Editors look for claims with citations that don't support them, references that can't be found, and stock phrasing that adds no information. Style alone isn't enough, because skilled human writers produce similar text. The most reliable method is opening the cited sources and checking whether they say what the article claims.
Why does AI-written text get past editor review?
AI text is fluent and well structured, which lowers the instinct to scrutinize it. Most editors were trained to treat awkward writing as a warning sign, but current models rarely produce awkward prose. Checking every citation in a long article is also slow, so unsupported claims can survive longer than they should.
Can you use AI to write a Wikipedia article?
Wikipedia's policies focus on the content, not the tool. Every claim must be verifiable from reliable sources, and text must be written in a neutral tone. Using AI to draft text is risky because models can invent facts and references. If you do use AI, you need to verify every claim and citation yourself before publishing.
Wikipedia AI-generated content is a verification problem more than a detection problem. Fluent prose isn't evidence of accuracy, and polished tone isn't evidence of human authorship. Try this the next time you read an article that seems too clean: pick the two most specific claims and open their citations. If the sources don't say what the article claims, you've found the problem, no detector required. Pass this to anyone who edits or relies on Wikipedia and is wondering how to spot AI-generated text without a detector.