Wikipedia's volunteer editors are doing less of the unglamorous work that keeps the site running, and generative AI hasn't changed that.
A new paper, "The Evolution of Coordination in a Collective Intelligence System" (arXiv:2609.16856), tracks nearly 25 years of English Wikipedia edits across five namespaces covering articles, talk pages, and governance discussions. The researchers find that participation in coordination and governance spaces has fallen relative to article writing, with a shrinking core of editors handling a growing share of that work. Using Markov-based session modeling, they also show editors moving between namespaces less often than before, meaning people who write articles increasingly don't also weigh in on policy or disputes. The paper then checked whether the arrival of large language models changed any of this and found only short-term blips, not a break in the long-term trajectory.
Wikipedia's governance model depends on enough editors sticking around to write style guides, mediate disputes, and vote on policy, not just add text. If that pool keeps shrinking while content keeps growing, the site runs on a thinner layer of institutional memory every year. That's a structural risk that predates the current AI panic and won't be solved by it either.
Blaming chatbots for Wikipedia's coordination problem, per this paper, misses two decades of the real story.