Mission, Not Race

Notes from the Wikimania 2026 panel on the state of Wikimedia and AI

These are my remarks from “State of Wikimedia & AI 2026: Freedom, Equity, Reliability at a Crossroad,” a panel at Wikimania 2026 in Paris, on Friday 24 July 2026, 11:00–11:55 CEST, in the Est Room. I took part remotely. Questions are paraphrased in italics, and my answers are lightly edited for readability.

We are not in a race

What are we misjudging about this moment, and what does the world actually need from this movement now?

I see us misjudging by thinking in terms of a race. From the inside, this year can feel like a scoreboard — like losing something: readers arriving somewhere else, answers delivered by systems that never show your name.

But as I advise leaders in various jurisdictions — and spiritual leadership too — I think an AI race is the wrong frame. And not just for us. I say this to everyone. Because a race implies a finishing line, and the AI race has a finishing line called the singularity, which is defined precisely as something literally beyond human comprehension. Nobody can win a race whose end we cannot even reason about. The labs sprinting hardest cannot win it either.

We like to talk about a mission, because we are a movement. And a mission is different, because a mission has milestones. Missions are also what people align with.

Wikipedia never tried to win anyway. We tried to make sure that there would be many winners at any time. Anyone could read, anyone could correct, anyone could build on it. That refusal to win a race is precisely why we became the knowledge layer that everything else, including language models, now stands on.

So we didn’t lose a race of viewership. The Wikipedia movement is the reason there is a mission in the first place.

What does the world need from this movement right now? For us to think about the next milestones we are setting — and to set them publicly, with the people, not just for the people.

For example: what must the agentic systems that answer from our knowledge do? What is the code of conduct? Showing their sources. Provenance. Taking corrections. Returning something to the commons. All of these are potential milestones. If we frame them as milestones, others will align with us. If we frame it as a race, they only fear us or ignore us.

The movement has led the mission for twenty-five years. Just don’t let this be the year we are mistaken for a race with the machines.

Keep a Kami

Tell us about the small spirit you keep on your laptop.

I call it a Kami — knowledge artefact management intelligence. It’s a new word, a kind of retronym, from knowledge management, KM, and artificial intelligence, AI. Put the two together and you get Kami.

It is a small, bounded steward. It runs on an ordinary laptop. Its memory, its identity, its ledger of sources all stay local — with the people it serves. None of the language-model providers can hold us hostage, and all the inference engines are replaceable. Most of them run locally, but even people who cannot afford a laptop with 16 gigabytes of memory can do this through zero-data-retention API endpoints. And it is retirable by contract: when the need it was built for goes away, it goes away too. There is no lock-in.

Why did I build this? Jimmy described the reader whose machine will visit fifty pages while the human sees zero. That is arriving on every laptop now. So the only question is: who does it answer to? I want a Kami to answer to the relationship it actually serves — not to whoever is renting it out.

Honestly, I don’t trust my own epistemic discipline. I also hallucinate. And I don’t trust anyone to steer something unbounded. So the failure mode I fear is not the machine that refuses to care for us. It is the one that cares for us so well that we forget to care for each other. It is like sending our robots to the gym to lift weights for us: we lose muscle, and then we lose friends as well. A Kami is my hedge against that. It is small enough to inspect, local enough to own, and mortal enough to let go of. It is just scaffolding. That is the property I trust most: bounded, relational care.

There are house rules. Every claim must be traced to a ledger, or it is rejected — and that answers Netha’s fear of citations vanishing. Every reading session ends with a write-back, so the Kami returns something to the commons it read from. So the fifty invisible visits are not fifty extractions. It is not data oil. It is data soil.

What we really need to work on is the talk-page protocol: when two sources disagree, how to surface that conflict to a human instead of silently resolving it — and how to make talk-page protocols work inter-Kami as well. The text is the last step. The real work is judgment, discussion, consensus, and how to make that work. We have some ideas. We experiment with Habermolt, which is basically asking a Kami to enter into a Habermasian discussion. That is just one of the very early pilots.

So I also invite you to keep a Kami, and to engage in talk-page protocols.

House rules matter

Of the house rules, which one should our community hold onto?

Treat every surfaced conflict as an invitation — an edit button reappearing, exactly where Jimmy wants new editors to enter. There should be a super edit button, instead of just me and my friends, or people on Habermolt, looping back to resolve the conflicts. That edit button for the entire web should reopen through agentic surfaces, to everyone.

The Plurality is here — just unevenly distributed

AI development is concentrating in a handful of foundation models, while many languages and knowledge communities carry no commercial weight. Is a future of many locally governed AI systems, each accountable to its own community, actually possible?

Yes. Definitely. And that future is already here, and it is being unevenly distributed as we speak.

This year, many people prefer local inference — edge AI — because it is now, for the first time, faster and also much less hallucinatory than cloud models. The data soil outperformed the data oil.

And hundreds of language communities, each governing its own civic context, its own wiki and norms and talk pages, and interoperating — that is exactly how the movement has worked for twenty-five years. The concentration you describe — the mainframes of our age — is real. But frontier models cost billions to train, while smaller models, commercially speaking, are much easier to train together. We are already seeing Thinking Machines Lab with the Tinker tool, Nemotron, and many others enabling local communities to decentralize post-training, fine-tuning, steerability and so on. They don’t cost a lot of electricity. They don’t cost a lot of water for cooling. That is a far more preferable way for communities to develop.

The cost of “good enough” has collapsed, as Mozilla’s report shows. Capable open-weight models are now within about 3% of frontier inference performance — and that is without local context. So bounded, inspectable and retirable Kamis are now, for the first time, the better thing.

And a point of clarification, since Chris made a correction: of course the Commons is a gift, and a gift enforced is no longer a gift. I am not saying we become the toll booth of the knowledge on the internet. I am saying we need to require the agents to surface the write-back to the human. That is the ability to be a good Wikipedian. But I don’t think we need to require mandatory write-back such that people get punished for it.

To gain trust, first give trust

You wrote the preface to the Taiwan edition of Jimmy’s book on trust, and compressed it into a single line. What does that line demand of the people building AI?

That line is: to gain trust, first give trust.

Wikipedia has spent twenty-five years proving that goodwill can compound. Every open edit button, every published draft — everything is a deposit into the trust reservoir. And the interest is the knowledge layer of the whole world.

What it demands of the people building AI is a Taoist maxim: to give no trust is to get no trust. So if the frontier labs are going to redeem themselves in terms of the trust they have lost, they need to unilaterally trust the people: publish the sources before they are forced to; take corrections before they are demanded; give the commons back its due before any regulator asks. That is a very simple thing. But it is not a feature you ship at the end of the lawsuit. It should be the first move in the game.

Start small and concrete

From the room: what experimentation could design such a talk-page protocol, and what governance space would it need?

I would like to start always small and concrete. Just pick any mid-sized wiki that wants to experiment with this protocol, and choose one very bounded task. For example: agents that read articles and file structured talk-page reports. I mentioned Habermolt because it is written explicitly as a skill file — how to write back to the Habermolt talk page — so look at that. And see where sources are disagreeing, where citations are not routing, where ledgers cannot verify the claims, and more. Then ask what I call system-weeder agents — maintenance Kamis — to act on those.

There is much less at stake than it sounds. We are finally seeing X.com using Community Notes and bridging-based feed re-ranking, which may lead to more pro-social, bridging timelines — we’ll see. And we’ll also see if Elon really does open-source everything. But it has been experimented with for multiple years on Community Notes, in lower-stakes situations, not the main timeline. The measurements they did back when Twitter still called it Birdwatch — and the work that continues on Meta and Bluesky and so on — these are early experiments we can incorporate into Wikipedia space. They are early, but they are also very promising.

For a 19-year-old starting to edit in 2026

Optimize for fun.

My hopes for 2030

After an hour of listening: What did you hear, and what do you hope for our community?

I hear a movement running its own talk-page protocol live on stage. We surface disagreements. We show provenance. We refuse false certainty. We turn conflict into energy.

I also hear the question still hanging in the air: AI is a political project. Is the only work worth funding making AI more capable? Or is it to make it obey? Some call that alignment. But I don’t think so. I think there is another side of alignment. Some of us call it reverse alignment — you can find it at reversealignment.ai. It is about redesigning our institutions and norms and skills so that society can absorb AI safely, and also share its gains.

But today, for every dollar making AI more capable, only a fraction goes to the institutions and commons that decide whether the benefits are shared or seized. That gap is not a law of nature. It is a choice. And that is why authoritarianism and tech-broligarchy seem to grow: because surveillance gets cheaper faster than the right of appeal and write-back.

So we can resist by doing what the movement does best: by demonstrating the alternative norm. In Taiwan, we call it forking the government. We build a free version. We prove it works. We force the government to merge it back. History shows that institutions, when forked this way — in a way that can absorb a technology — can be rebuilt in the image of something that holds that technology best.

And the Wikipedia movement is already that, in the image of the commons. If we can extend that into the age of AI, so that extracted versions are outcompeted by the demo of provenance and verification and contestability — that is my hope for 2030: that every answer shows its sources.

That is the mission, not the race. The singularity — the race — can only ask who wins. But plurality — the movement — can ask who is missing. And the movement has only ever asked that question: who is still missing in the room?

Don’t ever stop.