How Wikipedia’s Editors Keep the Online Encyclopedia Human-Written

We have all seen it. You land on a niche Wikipedia page looking for a quick fact, and the writing feels... off. It is too balanced, too repetitive, and uses phrases like "stands as a testament" or "rich cultural heritage" way too often. It reads like a high school essay written by a student who didn't read the book. In 2023, this was the calling card of the first wave of LLM-generated spam. Today, it is a full-blown war for the soul of the encyclopedia.

The Slop Detectors

Ellia Lebleu, a founding member of WikiProject AI Cleanup, started noticing these patterns late in 2023. What began as a few editors chatting on Discord has turned into a systematic effort to prevent Wikipedia from becoming an AI-generated echo chamber. Lebleu, who is also an AI researcher, saw the signs early: bizarre conclusion sections and "why this matters" headers that have no place in a neutral encyclopedia.

But the tell-tale signs are changing. The obvious clichés are being replaced by what Lebleu calls the "rule of three" and negative parallelisms. The AI is getting better at hiding, but the volume is still staggering. At one point, detection tools flagged nearly 20% of new drafts as machine-generated. Even if that number is dipping slightly now, we don't know if that is because editors are winning or because the models are just getting harder to catch.

The Hallucination Trap

For builders, the real lesson here isn't about the text; it is about the citations. We often talk about Retrieval-Augmented Generation (RAG) as the cure for AI hallucinations. If the model can look at a source, it won't lie, right? Wrong. The cleanup crew is finding that AI will often cite a real, legitimate source, but then completely fabricate what that source actually says.

This is a major problem for the "builder-first" mindset. If you are building tools that rely on automated data entry or content summarization, you cannot trust the RAG process to be self-correcting. Wikipedia editors are seeing cases where the citation looks perfect on the surface, but the underlying data is a total ghost. This requires a human-in-the-loop, or at the very least, a secondary AI verifier specifically designed to cross-check claims against their sources.

The Philosophical Pivot

Wikipedia has officially banned the use of LLMs for generating articles. On the surface, this was a practical move to stop the flood of garbage. But as Lebleu points out, it has become something bigger: an identity crisis. Wikipedia is now marketing itself as the "last good place on the internet" because it is human-written.

From a founder's perspective, this is a fascinating branding pivot. In a world where every website is being cannibalized by AI-generated content for SEO, being "human-only" is a high-value differentiator. Even the AI companies need this to be true. They need clean, human-produced datasets to train the next generation of models. If Wikipedia becomes 50% AI-generated, the models start training on their own output, leading to model collapse. Wikipedia's survival as a human project is, ironically, essential for the survival of AI.

The "Google Zero" Threat

There is a bigger storm on the horizon that every builder should be watching: Google Zero. This is the projected future where users never leave the search engine because an AI overview gives them the answer directly. If a user gets a Wikipedia fact via a chatbot, they don't click the link. If they don't click the link, they don't see the "Edit" button.

This breaks the feedback loop that keeps Wikipedia alive. The Wikimedia Foundation is trying to fight this by pushing for better attribution frameworks. They want the AI to tell the user when the article was last updated or if there are disputes about the facts. They are even looking at ways users could submit edits through the AI itself. This is the future of the interface: the platform becomes invisible, and the contribution happens through the layer above it.

What This Means for Builders

If you are building in the AI or crypto space, you need to understand that "slop" is the new technical debt. Building a system that generates 10,000 articles a day is easy. Building a system that ensures those 10,000 articles are accurate, well-cited, and don't degrade the quality of the broader internet is the actual challenge.

• Detection is a moving target: Don't rely on static markers. Structural analysis of speech is the new frontline.
• Citations aren't truth: Verify the relationship between the claim and the source, not just the existence of the source.
• Human-only is a feature: In an automated world, being the "human-written" option is a premium position.

AI was something to clean up yesterday. Today, it is a tool for the cleanup. Tomorrow, it will be the air we breathe.

The goal shouldn't be to keep AI out of everything. Even the Wikipedia editors are using AI tools to verify sources and flag pasted text. The goal is to ensure that the AI is the assistant, not the author. As founders, we need to be careful that we aren't just contributing to the noise. If your product doesn't add a layer of human verification or unique insight, you aren't building a tool; you're just building a faster way to make the internet worse.

Read the original at IEEE Spectrum →