August 2026

Wikipedia turned 25 this year, and Wikimania came to Paris to mark it. From 21 to 25 July, more than 1,200 Wikimedians gathered in person with 2,000 online under the theme Liberté, Équité, Fiabilité (Freedom, Equity, Reliability) to ask how the free encyclopedia stays trustworthy in the age of AI. Internet Archive Europe (IAE) came with a working answer.

Archive Citation Sources: evidential support, at scale

On Wednesday morning, IAE Programme Manager Beatrice Murch and Head of AI Daniel Erasmus presented Archive Citation Sources & Claims, a beta tool built on a simple observation: of the roughly 100 million outbound links in Wikipedia, only a fraction point to academic literature. The Internet Archive’s Turn All References Blue project fixed millions of broken links. Archive Sources asks the next question: can we help editors plug the holes in their citations — making published research more reliable, one missing source at a time.

The numbers behind the tool are considerable. It indexes 43.9 million open-access academic articles from scholar.archive.org, broken into 237 million passages, from which it extracts around 1 billion candidate claims that editors could add to Wikipedia. Paste in any Wikipedia paragraph, and the system retrieves relevant academic passages in under two seconds, then spends about half a minute matching specific claims to specific evidence, with an explanation for each match.

The crucial design choice is what the system looks for. “We’re not really interested in topic overlap,” Erasmus told the room. “We’re interested in evidential support.” A standard search finds articles about the same subject. Archive Sources finds passages that confirm, contradict, or update what a Wikipedia paragraph actually says. Fine-tuning the retrieval pipeline for that task lifted accuracy from 0.68 to 0.92 on the team’s benchmarks. In one live example, the tool flagged a Wikipedia claim that two monoclonal antibodies had been approved for Alzheimer’s treatment, and pointed to a paper naming a third.

The workshop lived up to Erasmus’s promise to “put the work into workshop.” Attendees tested the tool live, and two of them pasted in French and Spanish Wikipedia articles unprompted. It handled both, a promising sign for a tool tested so far only on English. Just as important, the whole pipeline runs on infrastructure in Europe, operated by the team itself, with no reliance on external model providers.

The questions from the room went straight to what matters: how should candidate citations reach editors? Talk pages, a standalone tool, integration with existing workflows? Those decisions belong to the community, and IAE is listening. One attendee working on biodiversity heritage asked whether other collections could be added. The answer was an enthusiastic yes.

Mark Graham, the man of many sessions

Internet Archive founder Brewster Kahle could not make it to Paris, but the Archive was well represented by Mark Graham, Director of the Wayback Machine, who seemed to be everywhere at once and several members of the Wayback team. On Thursday, Mark joined the panel “Collateral Damage? Human Creativity and Interaction in the AI Crawling Era” alongside Creative Commons CEO Anna Tumadóttir, where he described how the Internet Archive is working with news organisations on alternatives to blocking the Wayback Machine, such as rate limiting and purpose-limited access. On Friday he returned for “Bridging Institutions, Wikidata, and Generative AI: From Data Silos to Shared Infrastructure,” on how memory institutions and AI systems can share infrastructure rather than compete for it.

Mark also brought something brand new to Paris at the Preconference: Wayback Machine Labs, an experimental playground he started building himself only weeks before the conference. The Labs already hold more than a dozen working prototypes built on the Wayback Machine, with contributions from the Wayback Machine team and the GDELT Project. Among them: Perspectives, a way to explore how media covers a searched topic; FiveThirtyEightIndex, an index of over 38,000 FiveThirtyEight articles, datasets, and podcasts preserved after Disney deleted the site, and What Changed, a tool that finds government webpages taken offline or quietly modified. His invitation to the community is an open one: come play!

Wiki Fringe: the open door

In-person tickets for Wikimania sold out well before July, and a new registration process made them harder to come by. The Wiki Fringe, an independent, curated side event running alongside the conference, gave everyone else a way in: a space for discussions, demos, and experiments where Wikimedians and the wider open knowledge community could exchange ideas without an official badge. Initiatives like this keep the movement porous, and that matters. Free knowledge has never been an invitation-only project.

One mission, many time zones

Beyond the sessions and the networking, Wikimania did something no video call can: it put Internet Archive people from across the globe, and friends from mission-aligned organisations across the free knowledge movement, in the same room at the same time. Colleagues who usually work an ocean apart compared notes over coffee, in corridors, and late into the Paris evenings. The group photo took nearly as much coordination as a conference session. Archivists, it turns out, are easier to convene around an idea than around a camera. Worth every minute, though, when looking at the beaming faces.