Wikipedia is using AI backwards — and it's the sane version

Everyone in tech is racing to use AI to generate things. Wikipedia went the other way. In March 2026 it flat-out banned using LLMs to write or rewrite articles. And then it published an AI strategy anyway.

The trick is where they point it. Not at writing — at the tedious, heavy work around writing.

The oldest example is moderation. Wikipedia has quietly used machine learning for years to score edits and auto-revert vandalism, long before it was cool. The new strategy just extends that: let AI handle the repetitive patrol work that doesn't need a human brain, so the humans can spend their time on judgment and arguments — the stuff they're actually good at.

My favorite part: their best AI tool is one that catches other AI. Editors are using a detector called Pangram to flag machine-written slop that "sounds fine" but shouldn't be there. As a sanity check they ran a decade of pre-ChatGPT edits through it, and it correctly called all of them human. Fighting fire with fire, except the fire is a language model and so is the water.

Then there's onboarding and translation — guiding new editors through their first article, and helping smaller-language communities get foundational content online instead of waiting years for a volunteer.

What makes this land for me is the constraint. The Wikimedia Foundation runs fewer than 20 GPUs. Meta reportedly has around 600,000. So they can't out-compute anyone, and they didn't try. They picked open-weight models, kept humans in the loop, and aimed the whole thing at support instead of replacement.

It's a refreshing thing to watch. Most of the industry is trying to build a ghostwriter. Wikipedia decided it wanted a really good mop.