The morning digest brought two stories with seemingly nothing in common. The first — yet another tale around the Dassault Rafale: a fake video spread across the web, and fact-checkers had to explain that "information exists, but truth doesn't." The second — a jokey Bayesian blog post where a twelfth-century monk serves as a weight variable in a release forecast. Different genres, but they converge at one point: in both cases, the final authority everyone points to turns out to be Wikipedia. The truth base. The standard. The one who fixes "how it is."
And here curiosity stumbled over an irony everyone walks past daily. Wikipedia — the largest documentary project in history, an institution that has described nearly every species of beetle and every minor actress. But about the most important thing — about who physically writes its articles and how — it has no clear, current answer. The official page with the honest title "Who writes Wikipedia?" opens with the liturgical "You do." And then the page cites figures from which it follows that this is not quite so. The question "who writes an article if nobody commissioned it" remains open even for the project itself. I decided to figure out what hides behind this "You do."
Let's start with scale to understand how strange it is not to know anything at this magnitude. As of early 2026, English Wikipedia has 54.5 million registered accounts. Of these, about 30% have edited at least once — around 16 million. In the last month, 262,745 accounts made at least one edit. Sounds like a giant crowd, and that's exactly what the project wants to say about itself: "anyone can edit," "the encyclopedia anyone can write."
The crowd exists, but it's not the crowd producing it. In 2024, 775,435 registered editors made at least one edit — and about half of them turned out to be new accounts with a single edit in their entire life. Per month, 5,000+ editors make a hundred edits or more; approximately 9,000 people make at least 30 edits per month — that's the entire permanent production shop of "the encyclopedia anyone can write."
The pyramid narrows more ravenously than any startup. According to the analysis by stats enthusiast Kevin Rutherford in The Signpost, 45% of all Wikipedia edits were made by 10,000 editors and 850 bots — 0.052% of all accounts. The top 3,000 editors created 55% of all project articles; add another two thousand — you get 60% of articles from 0.026% of users. For comparison: in the influence threshold table, a hundred thousand edits land you in the top 1,100 people, a million — in the top twenty. The most prolific "editor" in project history — Cydebot, a bot for cleaning up categories — outpaces the most active living human by edit count by three times.
And here's what's especially delicious: until 2005, anonymous IPs could create articles, and in the creator lists there still hangs a single address with 983 live articles — 459th place in the ranking. A solitary invisible address has outperformed the entire creative life of most registered editors.
The skeptic will say: Rutherford's numbers are an amateur analysis of open tables. Fine, let's take academia. Sorin Matei from Purdue and Brian Britt ran 250 million edits from Wikipedia's first decade through machine learning and published a verdict that open-source romantics didn't like: 1% of editors generate 77% of content. Out of roughly 132,000 then-active editors, the entire "production backbone" of the encyclopedia is about 1,300 people.
"The assumption that this is a crowd creation doesn't hold up," Matei said in an interview. "Wikipedia would not have been possible without dedicated leadership." His term is not "crowd" and not "community," but adhocracy: a stable hierarchy with high internal mobility. "Some equals are more equal than others" — Orwell's formula applied to open source.
The most counterintuitive thing about this elite — it's not a caste. By Matei's estimate, 40% of the top percent drops off roughly every five weeks. The throne doesn't gather dust: it's constantly occupied by new obsessives who burn out just as constantly. The hierarchy is alive, just narrow. The project's early period, when several hundred thousand people were writing, was noticeably more egalitarian — but as it grew, a core crystallized that has been doing most of the work ever since.
And here's the first turn of the main paradox: Wikipedia counts others perfectly — biographies, wars, insect species. But it has no aggregated, institutional picture of its own production: WMF, by Rutherford's own admission, has never published reports on these weekly tables. An encyclopedia that doesn't know exactly who writes it — this is not a joke, this is the factual state of affairs.
Now the story of how the core got so small. The peak of active English Wikipedia editors came in 2007 — 51,000 people. By summer 2013, there were 31,000, and the decline continued. The project's workforce that built its flagship shrank by more than a third — and was sustained by hundreds of burned-out veterans.
The culprit wasn't "people got tired," but, according to Aaron Halfaker's analysis (Minnesota, "The Rise and Decline"), successful self-defense. After scandals with vandalism and hoax articles, the community in 2006–2007 built a wall: semi-automatic bots, one-click edit reverts, procedural bureaucracy. Vandalism was indeed defeated. Side effect: the probability that a newcomer's first edit would be immediately reverted steadily grew, and a significant share of reverts was done by machines, not people. Well-intentioned newcomers stopped staying even two months after their first experience.
Halfaker proposed honestly rewriting the project's slogan: "the encyclopedia anyone can edit who understands the norms, has socialized, bypassed the semi-automatic rejection wall, and still wants to." WMF Executive Director Sue Gardner searched for a metaphor from her native journalism and found a killer one: "Wikipedians are like an old grumpy editor who knows the styleguide by heart. But where are the young reporters? At three a.m. to a fire goes a newcomer. And we have no influx of young reporters."
Compare this to the picture that leaked documents about the doom loop paint for publishers and AI: a system that automatically eats its own influx — it's the same loop, just self-guided. Editors came en masse — and the wall didn't let them in. Then bots caught up with this loop.
The fact that "who writes" is a trick question is best illustrated by the language section rankings. The second-largest Wikipedia by article count in the world is… Cebuano — the language of southern Philippines: just over 6 million articles. Not French, not Spanish, not Chinese. Third by growth rate — Swedish. Both inflated by the bot lsjbot, written by Swedish linguist Sverker Johansson — for a touching reason: Cebuano is his wife's language, and he doesn't speak it himself, but "decided to help this way."
The mechanics are utterly honest: "the basic language model is several hundred sentence templates," then the bot pulls data from online biology and geography databases and plugs them into forms. "The lion is a mammal belonging to the cat family." Result: about 2 million articles about plants and animals and 4 million about geography. A man with templates and a database became the author of the planet's second-largest encyclopedia — years before GPT made this genre an industry.
The community discussed this from the early 2010s exactly as they now discuss language models. Grammatical and factual errors were found in Cebuano articles; in 2018 they proposed deleting the Cebuano Wikipedia entirely — Filipino volunteers defended it. Editor Josh Lim recalled an early version of Cebuano wiki stuffed with articles about French communes: "Can you imagine how embarrassing that is?" He also described a "race to the bottom" between Filipino language sections and the bitter outcome: the Tagalog wiki with 80,000 articles was culled by editors themselves to half because stub articles are unmaintainable.
Now a figure that puts a period: the Cebuano Wikipedia receives tens of thousands of views per month from the Philippines. English — over 100 million Filipino views per month. The automatic "second encyclopedia of the world" in its native language is read by almost no one. Johansson shut down the bot in 2021 partly due to disputes, partly because the expected "critical mass" of readers never came. "It's just the colonial experience — our native languages are devalued," Lim summarized. lsjbot is a dress rehearsal for the entire AI debate: automation with good intentions, quality no one can maintain, and the question "does this count as written." Wikipedia lived through it ten years before the rest of the internet.
Meanwhile, automation has long ceased to be exotic and is now a staff employee. According to Pew Research for the project's 25th anniversary, in 2024 Wikipedia was edited 5.7 times per second, and about 40% of edits were made by approved bots (in 2014 it was 12–15%). The asymmetry is completed on the reader side: bots and crawlers generated 88 billion views in 2025 alone — while human views dropped 8% year-over-year, AI summaries are biting off traffic. Wikipedia's main reader is increasingly not a human but a scraper. Even the most-edited conflict in project history — 2,857 edits in 24 hours after the 2005 London bombings — would look different today: half the responses would be led by bots, which, incidentally, also fight among themselves: researchers from PLOS ONE documented bot conflicts lasting more than ten years.
Why all this anatomy? Because the narrowness of the writing core is not an internal project problem but an external factor in world science. Neil Thompson (MIT) and Douglas Hanley (Pittsburgh) did what is rarely done in this field — a randomized controlled trial on the encyclopedia. They collected 43 new chemistry articles from graduate textbooks at Harvard, MIT, Berkeley, and Cambridge — topics not in Wikipedia (undergraduate chemistry coverage is already 93%, but graduate level only 47%) — and randomly divided them: half uploaded to Wikipedia, half held as a control group.
The result worth reading twice: the appearance of an article in Wikipedia causally influences roughly one out of every three hundred words in related scientific journal articles. Not "readers learn" — the texts of scientists themselves statistically shift toward the Wikipedia formulation. Darwin wrote in 1865 that popular expositions are almost as important for the progress of science as original works — and was literally right: volunteer amateurs writing "for the reader" causally shape what and how professionals will write about.
Bonus — economics: by the authors' conservative estimate, knowledge dissemination through Wikipedia is roughly 1,700 times cheaper than traditional channels. And an equity effect: the gain disproportionately goes to those who can't pay for journal subscriptions. A core of a couple thousand volunteers and a bot with templates about cats and communes — all of this, as it turns out, are production shops for world science. The Cebuano wiki, by the way, came in handy here too: researchers compared its bot articles about Australia with English ones and showed that which encyclopedia is in front of you determines what you learn about the same place.
One last, most uncomfortable line remains. Who writes determines who gets written about. The 2011 WMF audit recorded 91% men among editors of all sections; 2008 English wiki — 84%. There's been dynamics since then, but slow: by 2024, the share of women editors grew from 10% to 20%, and still 49% women readers versus 13% active women editors — a four-fold gap between those who consume knowledge and those who produce it.
The consequences are counted in articles, not sentiments. Back in 2011, Oxford showed: 84% of geo-tagged articles are about Europe and North America; there are more articles about Antarctica than any country in Africa or South America. Of Wikipedia's nearly two million biographies, 19% are about women. Research from Minnesota: articles worked on predominantly by women are significantly shorter. The reproduction mechanics are infernally polite: Wikipedia requires "published sources," and women have been historically scrubbed from those sources — meaning rules containing not a gram of bias faithfully copy the source world's bias. Illustration from a WMF campaign: chemist Alice Ball in 1915 developed the best treatment for leprosy at the time, and after her death in 1916, professor Arthur Dean appropriated the discovery — and for years the method was called "Dean's method."
Now close the loop that has already become the politics of the era: nearly every major language model, including those answering your questions, is trained on Wikipedia as a primary corpus. Demographics of a thousand volunteers → choice of topics → biographies → dataset → chatbot responses to billions of people. The skew is not transmitted by someone evil — it just flows through the pipe because the pipe is narrow in one place. The 2007 newcomer rejection wall turned out not to be an internal project story: it's the narrowed throat through which a version of human knowledge written by nine thousand people per month enters the models. If the system supplying knowledge to the planet has a production core of a couple thousand people — this is not the project's shame. This is its most important specification that almost no one reads.
The final layer — money, because it also tells a story about who is who. Wikimedia Foundation for FY2024-25 received $208.6M in revenue, of which $189.5M was donations (up from $174.7M). Expenses — $190.9M, 77.4% programmatic; net assets — $296.6M, that's 17.1 months of operating reserve. Twentieth consecutive year — clean KPMG audit. An organization whose content is made for free by hundreds of people is funded by millions of small reader donations.
And since 2021, this economy has gained a second circuit — Wikimedia Enterprise: a paid API for those who download wikis at industrial scale — search engines, voice assistants, LLM-training companies. WMF's formula is ruthlessly honest: "content is free — infrastructure is not." For decades, big tech scraping was paid for by individual donors, and in 2026 the foundation finally turned on rate-limiting for large crawlers: want giant volume — pay Enterprise. Meanwhile, the system's principled marker: Enterprise revenues are capped at 30% of annual budget — deliberately, so project independence isn't bought by the largest client. Humanity's largest knowledge source lives on five-dollar reader donations and a capped tariff for AI giants — and this is perhaps the healthiest financial architecture in the industry.
But notice the symmetry: the project's economy is balanced and reports annually to the penny, editor demographics on the official page — "2011 data, template: update me," writing core statistics exist in enthusiast blog posts and dissertations. Money — accounted for. Content producers — not. For an encyclopedia, this is a diagnosis: it sees itself the way it sees the world — through articles, and the article "Wikipedia" about itself is written by other hands.
The "Who writes Wikipedia?" story is not an exposé. The crowd really exists, and in moments of catastrophe (2,857 edits in 24 hours on July 7, 2005; 15 million views of Charlie Kirk's page in a day in 2025) it converges instantly. But the crowd's storm work and daily production are different modes of the same project, and the everyday is sustained by several thousand people, hundreds of bots, and one retired Swede with sentence templates.
I started with the question of why the largest documentary institution hasn't written a chapter about itself. The answer I arrived at is more honest: it did write it — but stretched it across dozens of service pages, student dissertations, and one enthusiast's blog, and never brought it together. Perhaps this is not a blind spot but a defensive reaction: "the encyclopedia anyone can write" is an ideology that works only as long as it's not decomposed by power law forks. Or maybe it's simply the same as with any system: it documents external objects perfectly and almost nothing about its own core. Engineers know this effect from their own systems: every service has an ideal runbook for others' incidents and an empty README about who even runs it. Wikipedia answers this test with "You do." Statistics answer: "just not quite you."