On 21 July, part of Wikimania in Paris, about sixty of us spent a full day talking about AI on the wikis. SJ Klein and I organised it to show two things: that large language models (LLMs) are already on the wikis, quietly enabling many processes; and that building with AI has become cheap and easy enough that people can solve their own problems rather than wait for someone else to do it.
What we already run
Before the lunch we asked everyone to “tell us about your AI/agentic projects” and got ten answers. The full list is in the Etherpad. Some of the highlights:
- Andrew Lih runs a self-hosted coding agent and three tools on Toolforge, the movement’s shared hosting for volunteer software.
- Asaf Bartov is doing most of 2026’s development on a 20+ years-old digital library this way. Bill Beutler tracks English Wikipedia’s conflict-of-interest edit requests.
- Sandra Fauconnier wrote three Wikidata user scripts (small pieces of JavaScript that add features to the editing interface).
- Hay Kranen ported Depictor from Vue 2 to Vue 3 in two days, a migration he had never found time to do by hand.
- Olaf Janssen bulk-publishes digitised images to Commons.
- Christophe Henner runs automated quality checks over content.
Andrew and Hay also ransessions; the others were simply in the room and all of them had something running. This was a pleasant surprise for us – we couldn’t have asked for better evidence of active use.
Andrew made the case that the tools worth building now are high-impact, low-burden. Tools like linters, quality checkers and metrics dashboards produce value while not increasing the burden on human contributors. Andrew’s Wikipedia AI Skills repo includes plenty of useful information on Wikimedia APIs, so the next person does not have to start from scratch. And his question stands: Quarry and PAWS brought SQL and data science to any Wikimedian – what is the AI-era equivalent?
Jimmy Wales dropped in during the afternoon and said he had been finding LLM coding tools fun as someone with rusty coding skills. Which is the argument in one sentence: the barrier that kept most Wikimedians from building their own tools has moved.
Mark Graham made the same point while opening the day. He started coding at 16 and is 66 now, and says he’s having more fun than ever vibe-coding internal tools for the Internet Archive such as a .gov Explorer for archiving and organising government-published infomation, a Leiden Dissertation Explorer for digitised papers that were headed for the bin.
Netha Hussain spoke about two distinct use cases: AI-assisted coding vs using LLMs in tools themselves. She used LLM coding tools to build fixarchive, which replaces deprecated archive.today links. The tool itself is deterministic and still got pushback from people who did not want to use something made with AI. She also chooses not to put LLMs inside her tools, because she feels accountable for what they output. Built with AI and running on AI are different things.
Galder Gonzalez‘s talk on data visualisation was followed by a lively discussion on the proper quality controls and the optimal way of creating AI-assisted graphics. A word of caution – Vera de Kok spoke up about the interface on screen: an AI-generated form did not label its fields making the app inaccessible to screen readers.
The scale of the problem, from three directions
Fermiboson presented AIlog, which flags edit summaries that look LLM-written on the theory that the edit probably is too. They quoted a number: English Wikipedia gets a new editor attempting LLM edits roughly every thirty minutes. About 55% of its flags are false alarms, which he argued is fine: it only has to be good enough to triage. The issue is whether enough editors have time and energy to act on the output.
Dev Jadiya gave the argument that should reframe everyone’s roadmap. The same sentence costs two to three times more to process in Hindi than in English, because models are billed by token (the chunks text is split into) and Indic scripts fragment into far more of them. So the work costs more, fits less in context and is less accurate and English-only benchmarks hide all three. Nor is a “small wiki” small in the ways you would assume: Gujarati Wikipedia has thirty thousand articles and fifty million speakers. What it is thin on is humans to catch a bad AI edit: three administrators, eighty-four active editors in a month. The wikis that most need AI’s help are the ones it serves the worst. What works, he showed, is MinT, the Foundation’s open-model translation service: it took Indic wikis from 11,000 to 59,000 articles in three years, and those articles are deleted less often than ones written from scratch, because the machine drafts and the human decides.
The third direction walked in off the metaphorical street. At 15:01 that day, a French investigative journalist writing as Manhack posted on the event’s talk page. He has spent two years identifying AI-generated news sites by hand – more than 15,000 in French, over 1,900 of them cited as sources on French Wikipedia – with volunteers already cleaning up. He has not found anyone in France to help build it into a framework. If you can, reply to him.
Productive coexistence
Hay Kranen ran a structured community conversation about what “productive coexistence” with AI would mean. Eleven themes came back and you can see them in the etherpad. The disagreements are the interesting part: should humans be in control or merely “in the loop”, a phrase several thought had stopped doing any work. And the eleventh: coexistence with AI makers, who embed their values in their tools and whose companies are straining our infrastructure. While we did not have anyone building LLMs in the room, the open-weight/source makers were not too far: Apertus presented its 1.5 release at the Fringe the next day and Anastasia Stasenko of Pleias was at Wikimania. Hay summarized his thoughts for the pre-conference and Wikimania in a blogpost.
Abzeronow, a Commons administrator who introduced himself as an “AI skeptic (willing to keep an open mind)” set out the issues Commons is facing. Commons has long held that it does not dictate what other projects use: under its in-use policy, a file that any project uses is in scope. Generative AI is now testing the settlement. AIIP (AI images of identifiable people) already overrides in-use, and an ongoing RfC proposes going further and requiring watermarks on all AI-modified files. The problem is real and the question is how to reconcile the in-use policy with the proliferation of GenAI images.
Eva Martin of Wikimedia Deutschland brought the Futures Lab‘s rules for having these conversations at all: de-emotionalise the topic, facilitate it properly, make room for perspectives from across the movement and from outside it, favour experiments over decisions, make it fun and avoid the generic term “AI”. Her message was to engage deliberately and to protect the movement’s credibility, which she argued is now the asset that matters as AI becomes the intermediary between people and knowledge.
Britta Gustafson – named a Wikimedia Laureate the following day – pointed out that stating risks as bare descriptions, such as “polarisation in the community”, is less helpful than stating them as if abc, then xyz. Identifying specific concerns and potential consequences helps the conversation move forward.
Sourcing and knowledge integrity
Verifiability is the policy Wikipedia’s credibility rests on, and there are plenty of uncited facts and far more citations than there are people to check them by hand. We realised that we have pretty much all the prerequisites to solve this problem:
- Sources at scale. Mark Graham mentioned Internet Archive’s handling of over 100 terabytes a day and 55,000 links to papers and books added to Wikipedia in a month. Daniel Erasmus, at Internet Archive Europe, has 250 million candidate citations.
- Research. Elena Simperl put it in one line: “Content generation is cheap. Trust is expensive.” Her King’s College London group has benchmarks for detecting AI-generated edits and ProVe for Wikidata, and asked 26 editors what makes verification hard: paywalls, dead links, policy confusion.
- Distribution, and the day’s news. Sucheta Salgaonkar showed the Foundation’s Edit Check and Suggested Edits which can be configured by each wiki’s editors. Open-weight models are now public on LiftWing, the Foundation’s own model hosting infrastructure. Connie Moon Sehat proposed surfacing Perennial Sources – English Wikipedia’s list of which publications are considered reliable – in that same interface.
- A working tool. I demoed Source Verifier, which checks whether a cited source supports the sentence it is attached to, on a panel moderated by Kevin Payravi – who wrote Cite Unseen – with Gerrit Quaremba from Elena Simperl’s group and Sucheta Salgaonkar of the Foundation. The panel spent its time on the step after that: once flagging is semi-automated, who fixes what it finds?
Come in
The best line of the week came from Wiki Fringe, SJ’s unconference alongside Wikimania: the room is the gift; a fringe is the structure that holds space open instead of filling it. If you are building a tool, have a problem that can be solved with the help of AI, a corpus or a benchmark you want to run or you maintain a smaller wiki, please tell us. WikiProject AI Tools on English Wikipedia is the most active space; Waikiki on Meta is the movement-wide one.
Thanks to everyone who presented, everyone who wrote in the pad, and the person who turned up, was helped to find some source code, and left the builder tables at seven in the evening building a user script of his own. That was the idea.
Session notes are in the Etherpad; the programme is on Meta.
Can you help us translate this article?
In order for this article to reach as many people as possible we would like your help. Can you translate this article to get the message out?
Start translation