Wikipedia at 25: A Wake-Up Call II - Data strikes back
This piece was first published in the author's userspace at m:User:Schiste/what-now-2-data-strikes-back on August 26, 2026."
By Christophe Henner - schiste · August 2026
Former Chair of the Board of Trustees, Wikimedia Foundation
20-year Wikimedian
In January, for Wikipedia's 25th birthday, I wrote that we had a two-year window.[1] I argued that our readership was diverging from the growth of the internet, that our contributor pipeline was collapsing underneath us, and that we had until roughly Wikipedia's 27th birthday to decide what we want to be in for the next 10/25 and to execute on that decision.
And now I'm back, but with new, additional and different data points. I don't think the trend needs to be proven any more, but I feel like we, as global communities, are yet to grasp the urgency we're facing. So I analysed the pace of readership decline and started to work on community health and editorship decline.
So, on the pace. In January, and in emails since, I have been sharing a faster decline, especially since I stopped looking at desktop traffic. I will get into the detailed rationale for why we should stop looking at that data later on, but for now the -8% you had in mind from Foundation communication has now reached -17.5% year over year.
The decline pace doubled in six months.
The underlying trend rate of mobile web human readership, fitted over the nineteen months to July 2026.
On the same method, the published all-access total gives -9.0% a year.
Log-linear fit on seasonally adjusted monthly volumes, all Wikimedia projects, bot-filtered. R² 0.86 on mobile web.
With identical parameters for both figures.[2]
Mobile web is our least contaminated high-volume signal; two thirds of desktop traffic is automated.
Because the number you pick is a choice, here is every measure of the same decline, so you can use the one you prefer.
| Measure | Mobile web | All access, as published |
|---|---|---|
| July 2026 against July 2025, single month | -12.9% | -3.6% |
| Twelve months to July 2026 against the twelve before | -14.8% | -8.4% |
| January to June 2026 against the same months in 2025 | -18.9% | -11.0% |
| Underlying trend, fitted, annualised | -17.5% | -9.0% |
I use the fitted mobile web trend because it is the only one of the four that does not depend on which month you happen to stand in. A single month can be distorted by anything, a trailing year lags by construction, and a half-year comparison inherits whatever was odd about the half-year before it. The trend uses every month in the window and tells you the rate the series is actually moving at. But whichever row you prefer, the gap between the two columns stays between six and eight points, and that gap is the part of this essay that matters.
In the next few sections I will detail the rationale of why I chose that one in detail, especially why I avoid desktop traffic as a baseline.
My take, short term::
Longer term: Big Discussions on what matters and what is out of bounds, Big Projects to replace what wikipedia.org used to give us externally and make it the best experience for the contributing communities, and a thousand experiments to start moving fast again, which first requires making failure socially acceptable again.
Let's start our data dive with the first half of the year, all Wikimedia projects, human traffic only, in billions of page views.[3]
| Channel | S1 2023 | S1 2024 | S1 2025 | S1 2026 |
|---|---|---|---|---|
| All access | 94.9 | 96.8 (+2.0%) | 90.1 (-6.9%) | 80.2 (-11.0%) |
| Mobile web | 61.9 | 63.9 (+3.3%) | 58.9 (-7.9%) | 47.8 (-18.9%) |
| Desktop | 31.2 | 31.0 (-0.8%) | 29.4 (-5.2%) | 30.8 (+4.8%) |
| Mobile app | 1.8 | 1.9 (+4.6%) | 1.8 (-2.6%) | 1.6 (-9.1%) |
Two years ago the first half of the year traffic was growing. it's now falling at 11% annually in aggregate and at 19% on the channel that carries three quarters of the volume of traffic.
There is a traffic oddity in those numbers: desktop page views. Everyone working on traffic data knows how complex it's to clean up desktop traffic from automated traffic. it's a very complex uphill battle, because we used to see a lot of scrapers and crawlers running all year long on a regular schedule, but now scraping campaigns are short lived. By the time you have seen one, and you start deploying metrics fixes, the campaign is over. They rarely last more than three months.
Desktop went up 4.8% in S1, great, and month by month it looks spectacular: +24.9% year on year in June 2026, +16.0% in July. That single channel is the only reason the all-access total for June 2026 reads -0.8% instead of something closer to -11%.
If you are starting to rejoice, dont. This is not recovery, but a lot of scraping campaigns running on our contents.
A small disclaimer here: I'm going to explain and rationalise why I don't use desktop traffic numbers to assess the trends. All I can provide are correlations; it's absolutely impossible to prove causation. So here are the facts, in ascending order of how hard it's to explain away.
First, the composition of desktop traffic.
| Year | Desktop | Mobile web | App |
|---|---|---|---|
| 2019 | 64.2% | 93.3% | 100.0% |
| 2022 | 45.1% | 82.3% | 99.8% |
| 2024 | 44.5% | 82.8% | 99.8% |
| 2025 | 36.8% | 79.5% | 99.5% |
| 2026 | 34.1% | 76.7% | 98.5% |
Two thirds of everything arriving at desktop is now automated, against one third in 2019. In absolute terms, desktop absorbed 10.0 billion bot page views in July 2026 against 2.5 billion for mobile web.[4] From what we detect, roughly 80% of all the automation hitting our servers lands in one bucket: desktop traffic.
The reason is quite simple: the people scraping don't care. They don't care about user agents and use library defaults, usually a desktop browser. Nobody writes a scraper that pretends to be an Android phone unless they're deliberately trying to evade detection, and they don't need to evade any detection to on our platforms.
So we know that if scraping campaigns go undetected, they're most probably running with a desktop user agent.
Second, the two channels move in opposite directions inside the same country. In itself that is not proof of anything, except when the trend from our numbers goes against the trend in mobile web usage and is the same across countries and regions.
Here is the first half of 2026 against the first half of 2025, in millions of human page views.
| Country | Desktop 2025 | Desktop 2026 | Change | Mobile 2025 | Mobile 2026 | Change |
|---|---|---|---|---|---|---|
| United States | 8,201 | 9,706 | +18.3% | 15,134 | 12,767 | -15.6% |
| India | 855 | 1,158 | +35.4% | 3,816 | 3,143 | -17.6% |
| United Kingdom | 1,689 | 1,776 | +5.2% | 3,987 | 3,680 | -7.7% |
| France | 1,400 | 1,504 | +7.4% | 2,663 | 2,397 | -10.0% |
| Spain | 604 | 677 | +12.1% | 1,362 | 1,150 | -15.6% |
| Mexico | 314 | 392 | +24.8% | 916 | 661 | -27.9% |
| Argentina | 220 | 317 | +44.0% | 517 | 416 | -19.6% |
| Indonesia | 279 | 317 | +13.5% | 933 | 629 | -32.5% |
| South Africa | 91 | 142 | +55.8% | 259 | 177 | -31.6% |
| Chile | 106 | 146 | +38.0% | 279 | 224 | -19.7% |
| Colombia | 122 | 162 | +32.2% | 315 | 230 | -26.9% |
| Malaysia | 138 | 177 | +28.0% | 439 | 354 | -19.3% |
| Thailand | 160 | 183 | +14.7% | 371 | 269 | -27.5% |
| Ireland | 119 | 152 | +27.2% | 314 | 309 | -1.7% |
| Morocco | 63 | 132 | +108.8% | 133 | 102 | -22.9% |
In India, the same population, over the same seven months, opened Wikimedia 673 million fewer times on a phone and 303 million more times on a desktop computer. In Mexico, mobile fell 28% while desktop rose 25%. In Spain, mobile fell 16% while desktop rose 12%. In France, mobile fell 10% while desktop rose 7%.
I have not found any account of the evolution of reading habits under which the same people, in the same year, abandon their phones and return to desktop computers, and do it in France and Indonesia and Argentina and Ireland simultaneously, in a decade during which every measurable trend in every one of those countries has run the other way.
And I don't know of any web property that can witness +35% growth on one traffic source while going through -18% on the other. If it were across different products serving differently different devices, it could happen, but on the same content with the same product, I have yet to find another example.
And this is not a handful of anomalies, it's a global trend we're experiencing:
| Pattern | Countries | Share of 2026 volume | Desktop | Mobile web |
|---|---|---|---|---|
| Desktop up, mobile down | 128 | 60.7% | +2.94B | -5.91B |
| Both down | 59 | 37.3% | -1.81B | -3.72B |
| Both up | 14 | 2.0% | +0.06B | +0.03B |
| Desktop down, mobile up | 0 | 0% | – | – |
Mobile web readership fell in 187 of 201 countries. In 128 of those, which is 68%, desktop readership rose over exactly the same period. The desktop gains in those countries total 2.94 billion page views, offsetting half of the 5.91 billion lost on mobile in the very same places. That offset is the entire reason our published headline for the first half of 2026 reads -11% instead of something closer to -19%.
The go-to analysis you would do then is a market one. We know different markets have different trends. The US, Canada and Western Europe are historically desktop strongholds. Perhaps that could explain things here too. Well...
The desktop-up mobile-down column contains the United States, France, the United Kingdom, Spain and Ireland alongside India, Morocco, Indonesia and South Africa. The both-down column contains Germany, Japan, Italy, Canada and Australia alongside Brazil, Taiwan, South Korea, Ukraine and Singapore. The differentiator is not geography, language, or device usage trends.
I stress this because I made the opposite mistake myself when I first looked at these numbers. I started by looking at it from a share-of-traffic perspective, and it looked overwhelmingly like a Global South phenomenon: desktop share rose forty points in Morocco and three points in Spain. But diving into the ratios, it was giving absolutely absurd results, that's why I changed the approach and went back to nominal numbers.
So the traffic dynamics don't track what I consider rational patterns on web properties, especially across so many regions of the world that have very different contexts.
Third, the magical months, when going granular increases the oddity.
| Country | Baseline | Spike | Then |
|---|---|---|---|
| Singapore | 55 to 125 (2024) | 250 (Mar 2025) | 38 (Mar 2026) |
| Austria | 32 to 38 | 157 (Jan 2025) | 27 (Jul 2026) |
| Germany | 300 to 330 | 517 (Jan 2025) | 249 (Jun 2025) |
| India | 115 to 135 | 309 (Dec 2025) | 150 (Jan 2026) |
| France | 172 to 250 | 330 (Jun 2026) | 167 (Jul 2026) |
Singapore was serving 250 million desktop page views in a single month, more than its mobile web traffic, and then 87% of it vanished. France doubled year on year in June 2026 and lost it all July (when we went on holiday perhaps it will be back scrapping in September :D). Austria quintupled and returned to baseline. Such surges on a 25-year-old web property have very little chance of reflecting human traffic.
The clearest single case, to me, is Chinese Wikipedia. Its mobile web traffic has fallen between 21% and 24% year on year in every single month of 2026, worse than any other major project. Meanwhile its desktop traffic swung from -45% in February to +58% in May. Broken down by country, the United States is now the largest source of desktop traffic to Chinese Wikipedia, at 77.6 million monthly page views against Taiwan's 38.6 million, having risen 178% in a year, while United States mobile web traffic to the same project fell by more than a quarter over the same period. American readers of Chinese Wikipedia did not triple on desktop and shrink on phones. American data centres did.[5]
The same French anomaly, incidentally, appears on Chinese Wikipedia at the same moment: French desktop views of zh.wikipedia went from 0.9 million to 7.1 million in May and June 2026, then straight back to 1.0 million in July. One campaign from a French organisation hitting multiple projects simultaneously.
So now we know that our projects, across different languages and countries, see very odd surges. I did not expect France to be a "smoking gun" in that research, but well, here it's.
Disclaimer: traffic data is a pain to clean and comb through. This essay is in no way me challenging the work done by the tech team at the Foundation. Doing an analysis as an essay like this is much, much easier than implementing cleaning strategies that work at scale.
As I said earlier, my goal here is not to do an update of my January essay, but to go a bit further. I think at this point the decline is not in doubt any more. At least I hope so! But now the question is what the impact is and by when. The first step is to rationalise the pace of the decline, and I totally arbitrarily picked the big turning point at half our current traffic.
To do that we need to compute a current decline rate, a pace, and project it into the future. To do that, second decision, I decided to use a trailing twelve-month rate: the last twelve months of readership compared against the twelve before them. That way is not good at picking up weak signals, but it's great at cancelling seasonality and special events, as both sides of the comparison contain every month of the year exactly once. A single anomalous month is diluted twelvefold, which makes comparison and projection a bit more grounded.
| Month | Rate | Month | Rate | |
|---|---|---|---|---|
| Jan 2024 | +9.77% | Nov 2025 | -7.88% | |
| Apr 2024 | +9.76% | Dec 2025 | -8.76% | |
| Jul 2024 | +4.02% | Jan 2026 | -10.06% | |
| Oct 2024 | +0.36% | Feb 2026 | -10.74% | |
| Jan 2025 | -3.01% | Mar 2026 | -12.23% | |
| Apr 2025 | -6.82% | Apr 2026 | -12.85% | |
| Jul 2025 | -7.17% | May 2026 | -13.83% | |
| Sep 2025 | -7.34% | Jun 2026 | -14.23% | |
| Oct 2025 | -7.41% | Jul 2026 | -14.77% |
Thirty months ago mobile readership was growing at nearly 10% a year. Today it's shrinking at close to 15%. That is a swing of twenty-five percentage points, and the rate has got worse in 28 of the last 30 months, including every single month of 2026.
Endpoint: https://wikimedia.org/api/rest_v1/metrics/pageviews/aggregate/all-projects/{access}/user/monthly/2021010100/2026080100, called with access = mobile-web, desktop, mobile-app and all-access. No authentication required. Full API documentation: Analytics API page views reference. The same series is browsable at stats.wikimedia.org.
Each row is the sum of the twelve months ending in that month, compared with the twelve months before it. "Corrected total" replaces desktop from January 2025 onward with mobile web multiplied by 0.4787, the 2023 to 2024 global desktop-to-mobile-web ratio held flat.
| Month | Mobile web volume (B) | Mobile web | Corrected total | Published total |
|---|---|---|---|---|
| 2024-01 | 127.5 | +9.77% | +4.25% | +4.25% |
| 2024-02 | 128.4 | +10.16% | +4.90% | +4.90% |
| 2024-03 | 128.8 | +9.69% | +4.77% | +4.77% |
| 2024-04 | 129.6 | +9.76% | +5.08% | +5.08% |
| 2024-05 | 129.2 | +7.96% | +3.97% | +3.97% |
| 2024-06 | 128.5 | +5.93% | +2.64% | +2.64% |
| 2024-07 | 127.8 | +4.02% | +1.38% | +1.38% |
| 2024-08 | 127.2 | +2.48% | +0.40% | +0.40% |
| 2024-09 | 126.6 | +1.80% | +0.17% | +0.17% |
| 2024-10 | 125.6 | +0.36% | -0.56% | -0.56% |
| 2024-11 | 125.0 | -0.34% | -0.65% | -0.65% |
| 2024-12 | 124.3 | -1.68% | -1.23% | -1.23% |
| 2025-01 | 123.6 | -3.01% | -2.52% | -2.47% |
| 2025-02 | 122.6 | -4.47% | -3.72% | -3.44% |
| 2025-03 | 122.0 | -5.32% | -4.35% | -3.78% |
| 2025-04 | 120.8 | -6.82% | -5.66% | -5.02% |
| 2025-05 | 120.1 | -7.07% | -6.04% | -5.41% |
| 2025-06 | 119.3 | -7.15% | -6.27% | -5.65% |
| 2025-07 | 118.6 | -7.17% | -6.33% | -5.73% |
| 2025-08 | 117.9 | -7.29% | -6.39% | -5.82% |
| 2025-09 | 117.3 | -7.34% | -6.65% | -5.87% |
| 2025-10 | 116.3 | -7.41% | -7.08% | -5.97% |
| 2025-11 | 115.1 | -7.88% | -7.80% | -5.88% |
| 2025-12 | 113.4 | -8.76% | -8.85% | -6.54% |
| 2026-01 | 111.2 | -10.06% | -10.06% | -7.35% |
| 2026-02 | 109.5 | -10.74% | -10.77% | -7.85% |
| 2026-03 | 107.1 | -12.23% | -12.24% | -9.23% |
| 2026-04 | 105.2 | -12.85% | -12.88% | -9.29% |
| 2026-05 | 103.5 | -13.83% | -13.76% | -9.14% |
| 2026-06 | 102.3 | -14.23% | -14.09% | -8.55% |
| 2026-07 | 101.1 | -14.77% | -14.70% | -8.38% |
Fit a line to the rate itself and it comes out at -0.84 percentage points per month, which is roughly ten points of annual rate lost per year, with an R² of 0.95. The same computation on the total corrected for desktop contamination gives -0.67 points per month at R² of 0.985, and on the published total, -0.50 points per month.
| Series | Rate, Jan 2024 | Rate, Jul 2026 | Deterioration | Fit (R²) |
|---|---|---|---|---|
| Mobile web | +9.77% | -14.77% | -10.1 pp per year | 0.954 |
| Total, corrected for desktop | +4.25% | -14.70% | -8.0 pp per year | 0.985 |
| Total, as published | +4.25% | -8.38% | -6.1 pp per year | 0.947 |
If we were at -8% it would mean we're in a somewhat slow-paced decline. But at almost -15% on a twelve-month trailing basis, this is quite fast acceleration.
The point of "no return", which I decided upon absolutely arbitrarily, is when we would have lost half our traffic. Half of what traffic? Well, for the purpose of these calculations, the current one.
Same source as above. Each scenario starts from the July 2026 trailing twelve-month rate and compounds monthly. In the deteriorating scenarios the annual rate worsens each month by the slope fitted to the thirty-one observations in the previous table. In the frozen scenario the rate never changes. Values are indexed to July 2026 = 100. The grey reference line sits at 64.1, which is half of the April 2024 peak of 129.6 billion trailing twelve-month mobile web page views, expressed against the July 2026 level of 101.1 billion.
| Years from Jul 2026 | P1 mobile web | P2 corrected total | P3 published total | P4 mobile web, frozen |
|---|---|---|---|---|
| 1 | 80.6 | 81.6 | 88.8 | 85.2 |
| 2 | 56.8 | 60.0 | 73.5 | 72.6 |
| 3 | 34.3 | 39.3 | 56.4 | 61.9 |
| 4 | 17.2 | 22.6 | 39.9 | 52.8 |
| 5 | 6.9 | 11.2 | 25.8 | 45.0 |
| Code | Scenario | Half of July 2026 | Half of April 2024 peak |
|---|---|---|---|
| P1 | Mobile web, deteriorating | 2.3 (28 months) | 1.8 (21 months) |
| P2 | Corrected total, deteriorating | 2.5 (30 months) | 1.8 (22 months) |
| P3 | Published total, deteriorating | 3.4 (41 months) | 2.6 (31 months) |
| P4 | Mobile web, frozen at -14.8% | 4.4 (53 months) | 2.8 (34 months) |
Reproducing this requires no modelling library: fetch the monthly series, take rolling twelve-month sums, divide by the same sum twelve months earlier, fit a straight line to the resulting rate series, and compound forward.
Each scenario is labelled P1 to P4 below and referred to by that code for the rest of this essay.
| Code | Scenario | Rate today | Half of today's readership | Half of the April 2024 peak |
|---|---|---|---|---|
| P1 | Mobile web, rate keeps deteriorating | -14.8% | 2.3 years (late 2028) | 1.8 years (spring 2028) |
| P2 | Corrected total, rate keeps deteriorating | -14.7% | 2.5 years (early 2029) | 1.8 years (spring 2028) |
| P3 | Published total, rate keeps deteriorating | -8.4% | 3.4 years (late 2029) | 2.6 years (early 2029) |
| P4 | Mobile web, rate freezes at today's level | -14.8% | 4.4 years (late 2030) | 2.8 years (mid 2029) |
P1 is the central case and the one I will refer back to.
P2 is the same computation on the total once desktop contamination is statistically removed.
P3 is what the published figures imply if you take them at face value, and it exists so that anyone who rejects the desktop correction still has a number.
P4 is the floor: the world in which the deterioration stops today and never resumes.
I added as a reference the grey line, which represents our mobile readership peak in April 2024 at 129.6 billion page views over twelve months. we're already 22% below that.
Half of peak or half of today, neither is a distant scenario: on the central case, P1, we cross half of peak in 1.8 years, in the spring of 2028, and even with the rate frozen today, P4, we cross it in under three.
The question is no longer whether we halve. Now it's which starting point we measure the halving from: if you like grey, we use peak; if you like colours, we use right now.
And in the rest of this long essay, keep P4 in mind, it's the optimistic case. Even if the deterioration stops dead today, if the rate simply holds at -14.8% and never worsens again, we lose half our mobile readership by late 2030. Nothing in the last thirty months of data suggests the rate stops accelerating, but even if it did, I consider P4 the floor of the discussion. And to be brutally honest, I would expect the acceleration to continue due to additional external factors such as a more aggressive AI overview rollout and our generational reach gap.
So... P1, the central case, and we lose half of our current traffic in late 2028. Two years and four months. Wikipedia's 28th birthday could be a very gloomy one.
In January I argued we had until Wikipedia's 27th birthday to decide and its 28th to execute. On these numbers, the 28th birthday is not the deadline for execution. it's the date we arrive at half.
Eight months of non-decision and inaction cost roughly six points of permanent annual decline rate.
Is this a forecast? Obviously not. For many reasons (data consistency and quality, external trends, and a crabload of other things) there is no way to build a forecast. This is nothing but a projection based on the last thirty months of data, with a few choices I made.
But are the odds that this is directionally right? I do believe so. First, because actual data did confirm the trend I shared in January. Second, because I did not paint a scenario where the decline accelerates (which is actually the scenario I believe will happen); all I did is focus on the data that is somewhat reliable to infer the global decline pace.
In January I argued for urgency on the basis of a gap: the internet grew and we did not, for years, and that was a thing I have shared across the last few years. But for the first time we weren't stagnating any more, we were slowly declining in nominal numbers. That was an essay saying we were leaving the plateau and entering decline (or, I hope, only the Trough of Disillusionment if we act).
So that concludes, to me, the question of whether we're going through an episodic decline or a systemic one. Everything points to systemic.
Now to some newer data points, not very encouraging, and some ideas for a path forward.
Let us work with the basis that traffic is currently falling at roughly 15% a year and that the rate has worsened in 28 of the last 30 months. The feedback I did get across projects and languages is "as long as we can create content it doesn't really matter".
I believe that readers are our number one contributor recruitment source (it absolutely doesn't mean all editors become editors by reading (I wouldn't be surprised other stimuli like teachers, meetup, news articles, Tik Tok, etc trigger the first edit but without massive readership all of those other stimuli loose a lot of power)). But as long as it's a belief, it's hard to use very strongly in discussions. So, a bit like in January when I went on the hunt for signals of what was to come, I have been doing the same for editorship.
My core belief is not only that community renewal depends on readership, but also that the impact of declining readership lags behind and that the acceleration would be much more violent than for readership. Because I believe the social fabric of our projects is key to what makes people stick with our movement for 20+ years. We could debate whether that same fabric may be pushing editors away, but today I focus on our communities' renewal.
So, to work on that, I started a new personal project, Wiki Economics, where I'm using existing indicators, adapting them to our world and analysing them, like the Gini indicator. The project is yet to be deployed on all wikis (I need to ask for much more storage and memory to do that once the ingestion process is solid enough) but it helped me validate and invalidate some ideas I had.
Long story short:
I will now share all the weak signals that help me paint that trend, but if I'm correct, in the next six months, each month, the editorship decline will continue and most probably accelerate.
The Foundation publishes contributor counts split by monthly activity level, which lets us separate people by how deeply they're involved.[6] I will use these four rings in the rest of the analysis, and refer to them by code:
| Code | Ring | Edits per month | Contributors, 2026 | Change from peak |
|---|---|---|---|---|
| C1 | Casual | 1 to 4 | 125,894 | -24.1% (peak 2020) |
| C2 | Occasional | 5 to 24 | 40,575 | -16.7% (peak 2021) |
| C3 | Regular | 25 to 99 | 14,279 | -9.1% (peak 2021) |
| C4 | Core | 100 or more | 9,470 | -9.3% (peak 2021) |
| All contributors | 190,218 | -20.5% (peak 2020) |
From that sheer data, there are three things we need to keep in mind, in my opinion:
So, in parallel with our readership decline, we do see an editorship decline too. But most importantly, at least to me, we can see contributors with small numbers of edits declining more. Meaning that our capacity to onboard and "train" the next generation of editors is already impaired.
About nine thousand people clear a hundred edits in a typical month. That is the C4 ring, across the twenty-two largest Wikipedias combined. Not nine thousand per language: nine thousand in total, and about fifteen thousand once Commons and Wikidata are counted alongside.
I checked that figure against the obvious objection, which is that summing per-project counts double-counts anyone active on several wikis. Working directly from the MediaWiki history dumps, I rebuilt the number the honest way: every main-namespace edit by a registered non-bot account across the twenty-two wikis, summed per person rather than per project, deduplicated by username.[7]
The answer is 9,107 people in July 2026, against 9,142 from summing per-project figures. The two effects cancel almost exactly: counting cross-wiki work adds people who clear the threshold only in aggregate, deduplication removes people who clear it on two wikis separately, and the net difference is under half a percent.
Because this way of capturing data is not sustainable, the goal is for Wiki Economics to take on the regular data computing.
Anyway, we have now assessed, in my opinion, that we do have an editorship decline.
To start trying to "correlate" (that is virtually impossible, we're looking for the breadcrumbs here), I plotted every ring of contributors against readership, using trailing twelve-month rates so that seasonality cancels and single months cannot distort the picture.
The ordering is visible without any statistics. Readership moves first, in red. C1 follows it closely. C2 and C3 move later and more gently. C4, the core, is the flattest line on the chart until very recently, and then it turns.
Cross-correlating each ring against readership in a proper table gives us this:
| Code | Ring | Best-fit lag | Correlation | Sensitivity |
|---|---|---|---|---|
| C1 | Casual | 1 month | 0.91 | 0.54 |
| C2 | Occasional | 7 months | 0.75 | 0.40 |
| C3 | Regular | 8 months | 0.70 | 0.40 |
| C4 | Core | 8 months | 0.68 | 0.44 |
| All contributors | 2 months | 0.88 | 0.48 |
This is a long data-heavy paper, it's a bit boring; well, the lag is actually very boring too. It rises monotonically with commitment, from one month to eight. And the most boring part to me (because it's part of my work) is that it's exactly how funnel conversion decay works in every other industry.
The sensitivity column tries to strengthen the correlation a bit and shows that for every ten points of readership decline, a ring loses between four and five points. All rings show that impact; the difference is only in when they feel it.
I want to be careful about causation here. Correlation at a lag is not proof that readers becoming editors is the mechanism, and there are other stories that fit: the same underlying shift in how people use the internet could hit reading and contributing independently, with contribution simply responding more slowly. That alternative would produce a very similar chart. What the data does establish is the ordering and the timing, and the ordering is stable across every ring.
The C4 ring is the one I want to look at a bit more, because it will be the last impacted AND will have the more lasting impact if it declines. From 2023 until the middle of 2024 it was flat or growing. It turned negative in August 2024, and for eighteen months it has been declining so very gently that it almost looked like a statistical blip. Nobody had any reason to worry: around one percent a year, which is well inside the range of normal churn in any industry. But the issue is not in the month-to-month or year-on-year data, it's in the cumulated impact across those eighteen months.
Then, in Q2 2026, it changed and it's decline started to accelerate:
| Month | C1 casual | C2 occasional | C3 regular | C4 core |
|---|---|---|---|---|
| 2024-01 | -0.52% | -1.86% | -0.96% | -0.18% |
| 2024-07 | -2.63% | -2.94% | +0.15% | +0.20% |
| 2025-01 | -4.05% | -3.54% | -1.45% | -0.86% |
| 2025-07 | -5.46% | -3.59% | -2.75% | -0.93% |
| 2026-01 | -9.41% | -4.34% | -2.59% | -1.26% |
| 2026-03 | -11.23% | -5.29% | -2.68% | -1.90% |
| 2026-05 | -11.86% | -5.71% | -2.65% | -2.82% |
| 2026-07 | -10.91% | -5.40% | -2.52% | -3.80% |
The C4 decline rate has tripled in just a few months and is not a rounding error any more. We never ran the analysis of how many new C1 editors we need to get a new C4 editor, but finding like-minded souls and getting them hooked (not on a feeling) on editing takes a lot. Everyone we lose takes a long time to replace.
Considering the lag is eight months, the next part is not a prediction in the usual sense. it's a bit more structured, as regressing the core's rate on readership eight months earlier, and then feeding in readership figures that have already been recorded, gives this:
| Month | Implied rate | Driven by readership in |
|---|---|---|
| Aug 2026 | -4.58% | Dec 2025 |
| Oct 2026 | -5.46% | Feb 2026 |
| Dec 2026 | -6.39% | Apr 2026 |
| Feb 2027 | -7.00% | Jun 2026 |
| Mar 2027 | -7.24% | Jul 2026 |
By March 2027 the core of the movement is on course to be shrinking at over seven percent a year, roughly double today's rate and six times the rate of eighteen months ago. Every input to that number is based on readership we have already lost. If the correlation exists, it also means that no action on readership will change that trend: that lever is in the past already.
I would rather this forecast were wrong, and it may be. Readership decline doesn't account for the whole decline, but it most probably accounts for a good chunk of it. And the direction has been consistent for years. The alternative is to believe the two are absolutely not related, which makes our problems worse to some extent.
In the discussions of the last few months, people also pointed out to me regularly that edit numbers are up and so overall things are OK. Which is absolutely true, edit numbers are up indeed.
Total content edits across all projects are running about 5% above last year. Anyone who wants to argue that the community is fine will reach for that figure, and on its own it does look pretty convincing... or does it?
| Year | Contributors | Edits per contributor | Share who are core (C4) |
|---|---|---|---|
| 2015 | 218,082 | 49.7 | 3.99% |
| 2019 | 228,397 | 95.0 | 4.00% |
| 2021 | 239,034 | 90.3 | 4.37% |
| 2023 | 224,731 | 94.3 | 4.45% |
| 2025 | 202,138 | 105.5 | 4.85% |
| 2026 | 190,218 | 111.7 | 4.98% |
Contributors are down 20.5% from the 2020 peak. Output per contributor is up 125% since 2015 and is still climbing. The work is getting done because a shrinking group is doing more of it, and the share of all contributors who are in the core ring has risen every year since 2019.
This is the pattern I described in January as concentration risk. Rising output per contributor is the reflection of better tools, better templates, better semi-automation. In a world where editorship grows, it's an amazing situation. But in a world where you have a very small decay, efficiency hides the decline even more and gives reasons not to look at it.
So if we have more edits but fewer core contributors, it does increase the concentration of work.
Using the same per-person dataset described above, here is how the work is actually distributed across everyone who made at least one main-namespace edit in a month.
| Month | Editors | Edits | Gini | Top 1% | Top 5% | Top 10% | Median | 90th pct | 99th pct |
|---|---|---|---|---|---|---|---|---|---|
| 2026-01 | 183,656 | 5,885,302 | 0.890 | 45.6% | 76.1% | 86.4% | 2 | 39 | 571 |
| 2026-02 | 171,365 | 5,488,660 | 0.889 | 46.1% | 75.7% | 86.2% | 2 | 40 | 551 |
| 2026-03 | 179,273 | 5,913,884 | 0.893 | 47.8% | 76.7% | 86.7% | 2 | 39 | 563 |
| 2026-04 | 173,521 | 5,370,344 | 0.887 | 45.5% | 75.6% | 85.9% | 2 | 38 | 543 |
| 2026-05 | 177,530 | 5,687,034 | 0.889 | 45.8% | 75.9% | 86.2% | 2 | 39 | 561 |
| 2026-06 | 171,087 | 5,412,265 | 0.890 | 46.6% | 76.1% | 86.3% | 2 | 39 | 544 |
| 2026-07 | 170,593 | 5,706,558 | 0.895 | 48.6% | 77.2% | 87.1% | 2 | 39 | 557 |
Half the encyclopedia is written by one percent of the people who write it.
I would like to take a moment on the Gini coefficient. Gini runs from 0, where everyone contributes identically, to 1, where a single person does everything. Income inequality in the most unequal national economy on earth sits around 0.63 as a reference. If we were considering edits as the high-value asset of our work (which is highly debatable, but it helps assess concentration statistically) it would mean that we're operating in a very, very highly concentrated model.
Extreme concentration is normal in volunteer systems (and, writing that, I'm thinking I should do the same calculation on open source repositories to compare when I can) and is not by itself an issue or odd: the people contributing the most are there because they choose to be, doing work the movement depends on and should be grateful for.
The Gini rose from 0.890 to 0.895 and the top-1% share from 45.6% to 48.6% across 2026, which points in one direction, but I would not build an argument on seven observations, and July is a seasonal low. Like the C4 decline, it doesn't set things in stone, it gives a direction, and gives another breadcrumb.
But it does give us a way to measure how much we rely on C4 editors, and clearly sets C3 and C4 rings as the most critical thing to take care of. More on that later.
Because if total edits remain the same, it does mean fewer people do more. Of course tool efficiency helps, but it may also mean that active C4 editors are actually compensating for the people who are missing.
It may also mean that the decay is not only hidden by efficiency improvements but also by burning a bit more volunteer energy to cope with it. it's not evidence that we have found a way to do more with less.
Now, in my opinion, we have the perfect recipe for a very, very, very slippery slope. The more we lose people, the more we will burn people out. And I would even say that the temptation to increase rules and regulations is going to be (already is?) very tempting as a way to mitigate workload, but research tends to prove that more often than not they don't mitigate workload, they mostly move it around.
A note on where these numbers come from. The distribution figures above are not published anywhere. I computed them from the raw history dumps for this essay, which is a slow and awkward way to answer a question we should be able to ask casually.
So I have started building a tool: wiki-economics, on Toolforge, which treats a wiki's contributor base the way an economist treats a population, reporting distribution and concentration measures per wiki over time rather than headline counts. it's early and incomplete: it doesn't yet cover all wikis, and the cross-wiki per-person view described here is not in it.
I mention it because the analysis in this section should be something any Wikimedian can reproduce in a browser in ten seconds, and because I would rather show unfinished work and be corrected than publish numbers nobody else can check.
This may not interest everyone, to make this a bit easier to read I collapsed it by default.
Two datasets. Ring counts, rates and the lag analysis come from the Analytics API. The per-person distribution figures (Gini, percentile shares, the deduplicated 9,107) come from the raw MediaWiki history dumps, because the API cannot answer questions about how work is distributed across individuals.
Distribution dataset. MediaWiki history dumps, 2026-07 snapshot, 2026 partitions for the twenty-two wikis listed below. Rows filtered to event_entity = revision, event_type = create, page_namespace_historical = 0, event_user_is_anonymous = false, both bot flags empty, and usernames not beginning with a tilde (temporary masked-IP accounts). Edits are then summed per username across all wikis before any bucketing, so a person working on three projects is one person. Validated against the Analytics API: English Wikipedia agrees within 0.2% on both editor counts and every activity band across all seven months; French agrees within 4%, the residual most likely being revisions later deleted. Gini is computed on the per-person monthly edit counts of everyone with at least one edit.
Sources. Wikimedia Analytics API, endpoints metrics/editors/aggregate/{project}/user/content/{activity-level}/monthly, metrics/edits/aggregate/all-projects/user/content/monthly and metrics/registered-users/new/all-projects/monthly, pulled 25 August 2026. No authentication required. The same series are browsable at stats.wikimedia.org.
Coverage. The editors endpoint doesn't accept the all-projects aggregate, so contributor figures are the sum of the twenty-two largest Wikipedia language editions: English, German, French, Spanish, Japanese, Russian, Italian, Chinese, Portuguese, Polish, Dutch, Arabic, Indonesian, Ukrainian, Persian, Turkish, Hebrew, Korean, Swedish, Vietnamese, Hindi and Bengali. Commons and Wikidata were fetched but are excluded from Wikipedia totals, since their contribution patterns are structurally different. Edits and new registrations are all-projects figures and are noted as such.
Definitions. Contributors are registered users editing content pages; anonymous and bot edits are excluded by the user editor type. Activity levels count edits to content pages in the calendar month.
Rates. All rates are trailing twelve-month: the sum of the last twelve months against the twelve preceding them. Seasonality cancels by construction. These are lagging measures, so a July 2026 value describes August 2025 through July 2026.
Lag estimation. Pearson correlation between each ring's trailing rate and the mobile web readership rate, evaluated at every monthly offset from -36 to +48, over the window from January 2018. The reported lag is the offset maximising correlation.
Forecast. Ordinary least squares of the C4 rate on the readership rate eight months earlier, fitted from January 2018 (R² 0.46), then evaluated on readership figures already recorded through July 2026. it's a projection of an observed statistical relationship, not a causal model, and no claim is made that readership causes core contributor decline.
What this doesn't show. Nothing here demonstrates causation, identifies which contributors left, or distinguishes departures from reduced activity. It establishes the ordering, the timing, and the direction.
So far, it was mostly measurement. Of course I had to make some assumptions and decisions, which I shared along with the explanations. But after this point you enter a full opinion section based on measurement. it's now what I think the measurements mean and where we seem to be going.
This is the core of what I have been trying to share for some time now. Some people may even remember me in 2018 in Berlin comparing our growth rate to the rest of the web. But now we have stronger data to start building more "solid" hypotheses.
Correlation is not causation, but it feels like it to me. If we look at the data, at expected behaviours and dynamics, we're looking at the ripple effect that started a few years back and, I believe, got hidden by the COVID surge. we're not unique; other industries are actually going through the exact same issue, one of the most visible being the video game industry.
Numbers were mostly stagnating, COVID hit, inflating all metrics and hiding the previous situation. When the bump stops delivering and reality hits back, it hurts. Top that, in our case, with systemic changes in how content is produced and consumed, and the slap becomes violent.
it's a ripple effect, and it takes time to be visible from the banks of the lake. But I believe we now see it.
| Measure | Now | Trend | |
|---|---|---|---|
| Readers | Mobile web, trailing 12 months | -14.8%/yr | Worsening ~9pp per year, 28 of last 30 months |
| Published figure | All access, trailing 12 months | -8.4%/yr | Overstates by ~6pp; desktop is two-thirds automation |
| Casual contributors | C1, 1-4 edits/month | -10.9%/yr | Follows readers at a 1-month lag |
| Occasional | C2, 5-24 edits/month | -5.4%/yr | 7-month lag |
| Regular | C3, 25-99 edits/month | -2.5%/yr | 8-month lag |
| Core | C4, 100+ edits/month | -3.8%/yr | 8-month lag; rate tripled in seven months |
| Concentration | Share of edits by top 1% | 48.6% | Gini 0.895 |
I never saw those numbers before in our movement (but I could have missed them), nor did I see any dual analysis. Which means that we may be actively making (or not making) decisions based not only on lacking metrics but also on wrong metrics. While things are accelerating, we're not operating with the proper context.
An organisation that cannot see its own decline cannot decide to act on it. Before any strategy, we need trust. Trust in what the reality is, what the situation is, in each other, and so on.
Before going to the next sections, a last opinion: given the level of concentration AND the decline acceleration of C3 and C4, we no longer have a margin of error or indecision.
The following is not a forecast of what will happen. it's what the last thirty months imply if nothing changes: readership continuing on its measured trend, and the core following it at the eight-month lag we can observe.
| Date | Reader rate | Readers (Jul 2026 = 100) | Core rate | Core contributors |
|---|---|---|---|---|
| Jul 2026 | -14.8% | 100 | -3.8% | 9,107 |
| Jan 2027 | -19.8% | 90.7 | -6.5% | 8,850 |
| Jul 2027 | -24.9% | 79.7 | -8.7% | 8,497 |
| Jul 2028 | -34.9% | 55.5 | -13.2% | 7,549 |
| Jul 2029 | -45.0% | 33.0 | -17.7% | 6,370 |
| Jul 2030 | -55.1% | 16.3 | -22.1% | 5,090 |
| Jul 2031 | capped | 6.7 | -26.6% | 3,840 |
The two dates that are currently driving me, like a final countdown (yes pun intended I'm old enough to make it now).
Late 2028: half our readers are gone. Two years and two months from now. That is Wikipedia's 28th birthday, which in my January essay was the deadline for having finished executing a strategy. it's now the date we arrive at half of our current traffic, dividing our new editor stream, our revenue stream, our weight in discussions with public institutions, museums, and so on.
Early 2031: half our core contributors are gone. Four and a half years. The contributors we lose in 2030 are not being renewed by readership that disappears in 2028 or 2029, and the readership that disappears in 2029 is on a curve we can already draw.
I don't believe these curves run to their ends. Nothing falls to zero, some irreducible core of people will keep coming directly, and I have deliberately not modelled a floor because I would have to invent its value. Treat 2027 and 2028 as meaningful and everything past 2030 as an illustration of a direction.
But note what the shape does to the usual reassurances. The contributor curve looks gentle for two more years. It will be used, in good faith, as evidence that the community is holding up while readership falls. The gentleness is the lag, not resilience. By the time the contributor curve looks alarming, the readership that determines the following year is already lost.
In January I said we had two years to decide and three to execute. Eight months later, the readership decline rate has worsened by roughly six percentage points, the core contributor decline has tripled, and the deadline I set for finishing has become the date we reach half.
Eight months of process cost more runway than eight months of calendar.
One: we need to stop considering readership growth / recovery altogether. This is a big change for a movement, and for an organisation, the Foundation, that was intrinsically built around reader numbers. But that ship has sailed. At almost -20% year on year, we have entered the realm of non-recovery. Imagine a ship with a hole. There is a moment where the hole is so big that you will consume a lot of energy to slow the inflow down temporarily, but the pressure is so great that the bandaid suddenly fails and the situation gets worse abruptly. You don't fight to save the ship; you let sink what is not salvageable.
Two: we need a short-term absolute focus on C3 and C4 on all projects. In that downward spiral, C3 and C4 editors are the people where we can have the most impact and where we can slow the decline down. And it's also the place where actions can actually make a dent. And this is about people, workload, energy: we need to make sure we reduce organic churn as fast as we can. Renewal is broken; while we fix it we need to protect our communities and make them last.
Three: we need to milk wikipedia.org readership right now. This is where I will get the tomatoes (though I love tomatoes), but we need to stop thinking of wikipedia.org readership as something to protect. Editors we need to protect. Open content we need to protect. But wikipedia.org readership is going to be gone, and while that happens we need to leverage it for two things we direly need: money and people. We don't want to fight trying to increase readers, we want to make sure every single value-aligned reader becomes an editor. Readership teams should be solely focused on editor recruitment.
If I sum it up: the long-term solutions will need a lot of work and discussion (though we need that fast), but for those to exist we will need money and people. So we focus, in the short term, on:
Since January, a lot of things have happened. What I published in January was meant to be a wake-up call. But by June little had changed and my frustration grew to the point where I realised that we were all, me included, responsible for the stagnation.
I did start, some before June even, different initiatives, discussions and work, but we may need a framework. A way to structure discussions, actions and learnings.
On those, there are a few things that are absolutes:
The other thing that changed is that I hoped the Foundation would step up and lead the change. But it's becoming abundantly clear they will not, or at least not with the intention of being ready to reconsider everything we're doing in order to salvage our mission. Part of me still hope to see that change, but I stopped expecting it.
So what follows is not a plan for the Foundation, but the way I'm structuring things in my own mind and how I try to allocate my time and work, and also how I think we could work on our topics as communities.
| Topic | Goal | Speed |
|---|---|---|
| Big Discussions | Document consent about what matters, and what is out of bounds | Months |
| Big Projects | Infrastructure and products the movement doesn't have yet | Years |
| A Thousand Experiments | Everything we need to try, most of which fails. From editor tools to 3D content, videos, citation checkers, and so on | Weeks |
Those are not sequential. Waiting for the discussions to conclude before experimenting would be another version of the mistake we have been making for a decade. And actually each should inform the others, and people should not feel compelled to contribute to all, or any. Because on top of that, the day-to-day needs to happen and, back to my earlier point, the priority is supporting C3 and C4.
The movement has not had a real strategic conversation since the Movement Charter process fractured. I don't think people stopped caring, but that attempt cost so much and returned so little that nobody wants to be the one to start another one. That is an understandable reaction but it's a dangerous one: the questions did not go away when the process collapsed, quite the opposite.
But the part that is true is that it cannot be a similar process with a closed group working in the dark. What I believe in is creating a space where consent can emerge and dissent can exist to build.
That choice of word, consent, is absolutely deliberate. Consent is not consensus.[8] Consensus asks whether everyone agrees, which on a question of any real weight is a machine for producing nothing. Consent asks whether the material objections have been heard and addressed well enough that a responsible next step becomes possible. Silence is not consent. An objection is not a veto. And the formal decisions stay exactly where they already sit: with the communities, the affiliates, the committees and the boards that hold the authority. The consent machine documents and grounds the lay of the land to make decisions easier and more connected.
The difference matters because it changes what a disagreement is for. In a consensus process, an objection is an obstacle to be worn down. In a consent process, an objection is information: it tells you where the risk is, who is exposed, and what a better version of the proposal would look like. Making disagreement not only heard but actually useful is the whole design.
Lodewijk published an article on the topic recently on why we need to innovate in our decision making[9] and he has been working (full disclaimer: I do give him a little assistance) on a tool that could help make our decision making richer: proto.wiki.
So I started Next 25, which I see as a campaign that needs to end as fast as we can. It began as a set of community members who assembled in the first half of 2026 to share concerns, and now I have made a proposal to turn it into a six-month public experiment to build a consent-creation engine, starting by using that process on itself to decide how it should work.[10]
The goal I see in Next 25 is an umbrella to start having those discussions and those decision-making innovations, but the campaign becomes irrelevant when consent has been adopted or decided to be dropped absolutely.
But this is one way. There are dozens of other ways, other places, other intents. All I know is that we need to start having those discussions again, and fast.
Earlier in this "essay" I shared how wikipedia.org, our golden goose, providing editors, money and public power, is coming to an end. But to be able to keep pushing our mission forward, we need those anyway. So we need to find ways to get them back, at the very least editors and money.
Did I not mention content creation and curation? Because on that ground we already have great tools. It doesn't mean we cannot improve them, but in our hordes of issues, getting thousands of people to collaborate across languages and cultures is not one of them.
So what comes after wikipedia.org as a stream?
Most probably not only one thing. We were very lucky in that sense, and probably many different things.
In those discussions, the bigger issue I saw, especially at Wikimania talking with people I have known for years and who are very smart, is that every time I mentioned those big projects, the push-back was "well sure, any ideas?". Since January I have been careful not to share my ideas too much, not because I don't think they're worth sharing, but because I was being mindful of giving space for other directions to emerge. But I also realise that sharing them gives room to think outside what we're doing today. So here are some ideas of mine and a couple that are not from me.
When I look at our assets and where the evolution of current trends leads us, the paths cross each other. One thing we're the only ones in the world to have is 25 years of building an engine to assess the authority of facts. That engine is human driven, but it's absolutely amazing.
On the other hand, by design, LLMs need to be creative, but inference costs need to drop for providers. The issue being that to drop costs they route requests to smaller models and we degrade final state quality.
A great way to improve that without increasing costs is to actually use a deterministic authoritative engine. A way for a small model to push "I will build my answer on facts 1, 2 and 3, grade them" and an engine to assess confidence in facts.
A verification layer that sits at inference time and tells you not what is true, but what is established, what is contested, and how contested.
Hallucination (I hate that word, they're not a bug, they're the core feature) is seen as a model problem, but it's an infrastructure problem. Nobody can build that layer alone credibly. They would need the knowledge, experience and data to create the engine, but also to build it as a commons to make it auditable, with mathematical confidence rather than editorial judgement. It becomes (will be) something states, publishers and AI companies all need and none of them can make.
The fun part is that it gives value to what we have been doing in the talk pages and history pages for 25 years.
Citations are the poor relation of our projects and simultaneously the thing that makes them trustworthy. Turning every citation into a structured, addressable object, across all languages, would make editors' lives easier immediately and make automated gap-finding possible at scale.
But doing that makes it something of a lot of worth for a lot of people on top of us, including researchers, GLAMs, and again frontier labs.
This is absolutely not a brand new idea. There are currently existing initiatives, intentions and attempts.[11] But they're on the fringe. This is exactly the kind of big project that is easy to start; it just needs a clearer direction, better communication and actual real resources.
A multimodal, multilingual, vectorised dataset produced by humans, with provenance attached to every claim.
Datasets are what AI needs to be trained on, and if LLMs train mostly on text, there are a lot of other AI types and use cases (including in LLMs and SLMs) where being able to provide an always up-to-date dataset is the best way to ensure knowledge, in a neutral form and not designed with a very intentional bias, finds its way through consumer-grade LLMs.
But it also becomes a golden dataset for much narrower models and use cases.
This is Wikimedia as a seed vault for human knowledge. It builds on what we already have, what people use from us, evolves it and makes it something that may be worth paying for.
All that being said, let us be honest, there are a lot of technical challenges, but wouldn't it be fun?
The most speculative, and the one I care most about.
I believe we're entering an AI age and the only ways back from it are Dune- or Foundation-level events. Whether we like it or not.
In that world, knowledge distribution is based on LLMs. When we entered the game of knowledge we were competing with peer-reviewed encyclopedias. We did not compete by playing the same game with an open licence. We changed the game.
We competed with institutional intelligence using collective intelligence.
We competed with editorial boards using decentralised open production.
I do see people advocating open weight models and open source LLMs, but they're trying to play the same game as the labs. LLMs are trained on massive infrastructures, in a highly centralised way, with absolutely no full auditability.
What if we went about LLMs the way we went about printed encyclopedias, by changing the game?
Theoretical research is fun, and lacks a product, but we know there are ways to train large models in a very decentralised way. In a much more granular way. There are a lot (A LOT) of technical challenges, but imagine if instead of relying on one huge LLM we were relying on thousands of very expert models. A decentralised mixture of experts. Each trained and supported by experts. Imagine if Women in Red were training and maintaining a model dedicated to that. And the video game project in jawiki one on video games in Japanese. But this would go beyond us: the Museum of Black Civilisations, the ALMA Observatory or the Louvre could operate their own experts. They could take ownership how their data can be used in a larger model.
And the way SETI@home or Folding@home did, instead of building gigantic data centres to train huge models, we distribute training in small chunks on people's computers.
What would be open is a model where the corpus, the recipe, the evaluation, the governance and the artefacts are all public and forkable. And the way to build it's not to out-spend laboratories that out-invest us by orders of magnitude. it's to change the shape of the problem: many small expert models, independently trained, composed at inference time, contributed by communities and institutions who keep their own data and retain the right to withdraw. Recent research has made that credible rather than merely hopeful.[12]
Everyone has seen their catalogues and databases pillaged. Maybe we can try to build one that is made respecting and supporting each contributor.
And that comes with two challenges (well more like a thousand but two overarching ones), and I'm not sure which would be the greater: a technical one and a governance one.
Is it possible exactly as I imagine it? Probably not, I honestly don't know right now. But neither was a free encyclopedia in three hundred languages.
The first two tiers are slow by nature. But time is our enemy; the last eight months have made that, I think, abundantly clear.
And the first two tiers don't exist in a vacuum. They exist in a reality where we need smaller wins, more incremental work, and we also need to manage today's issues. For that we need experimenting.
In my opinion what blocks experimenting is not money, tooling or permission. it's that we have made failure socially expensive (and as I proofread that, I realise that it also applies to organisations and explain why we see this leadership gap).
A tool that doesn't work, a workflow nobody adopts, a prototype abandoned after three weeks: these are treated as embarrassments to be quietly deleted rather than as the ordinary output of trying to do things. I have been prototyping a way to make authors more prominently visible (from an idea from Amir). I got support, but I also got quite strong push-back stating I was not allowed to do it prior to discussing with the community.
Experimenting, sticking one's neck out, is dangerous and complex socially. So people either don't try, or they try in private and are very, very cautious about not telling anyone.
A movement that cannot tolerate failure cannot experiment, and a movement that cannot experiment has only one strategy available to it, which is to keep doing what it's already doing, slightly harder. The data in the first two parts of this essay is what that looks like after two decades.
But innovation is bubbling. We never had as many new tools made as this year, thanks to increased capacity from LLMs and appetite to innovate. But it's done discreetly: one failure doesn't benefit the others. Infrastructural parts are done over and over and over again.
So the first thing is simple: builders and innovators, come out of the woods and share what you are building. Celebrate your errors, your failures and your successes. And let us work together to make it acceptable to fail in public.
The second thing is coordination. Plenty of people are already experimenting in isolation. we're unaware of each other, we're solving the same problem three times in three languages, and none of what we learn accumulates. We lack the place to even see that.
This is why I have been building Toolhub Evolved: an extension of Toolhub.[13] Toolhub stays the source of truth, but I have been giving it a boost. Well, many additional features, but the most meaningful one, you ask? it's now an automated directory of all running user scripts and gadgets on all projects. Next? Enriching the data layer to be able to identify neighbouring tools that are heavily used, and start connecting people, ideas and needs.
We have over 50,000 different tools. I'm quite sure we could find a couple of them that are functional duplicates and would benefit from each other.
The third thing we're lacking is consent to experiment, which ties back to my Big Discussions point. I know of some very, very, very interesting experiments going on right now, but people seldom share them because the first feedback is tough to read, as it will be about what is wrong. That is our culture. We need to change that.
That is where I say I hope it's my last very, very, very long essay. That the next one will not be by me, but by someone listing all the cool things we have done and changed.
And now, well, I will keep pushing. I will try to make Next 25 irrelevant by creating a consent engine that doesn't need a campaign to make it exist. I will try to get Toolhub Evolved to a state where we will be able to learn about tools better across languages and projects. I will try to get SP42 working as a composable interface for tool builders and editors. I will try to get hints of human authorship right below the title of an article to attract new editors on as many wikis as possible. I will try to find more crazy people to dream about some big projects.
I will try a lot of things, but what I will not do is accept the trend I described without doing the change I call for.
.png){kind=small}
pageviews/aggregate/all-projects/{access}/user/monthly. The user agent filter is the Foundation's own bot exclusion. The same series is browsable at stats.wikimedia.org. Data pulled 24 August 2026.
all-agents and user series on the same endpoint and the same months.
all-projects aggregate, so the figures below are the sum of the twenty-two largest Wikipedia language editions, covering the large majority of contributor activity. Commons and Wikidata are excluded from the Wikipedia totals and discussed separately.
Discuss this story
(This allows for greater visibility of discussions, makes archiving easier, and prevents discussions becoming disconnected from articles during the publication process)