Kami

This week, in Paris, we celebrate Wikimania and a quarter-century of Wikipedia. Earlier this month, I wrote the preface for the Taiwan edition of Jimmy Wales's book on trust, and it fits on one line: To gain trust, first give trust. Wikipedia has spent twenty-five years proving that goodwill can compound.

A quarter-century of compounding goodwill — that is what our room has built, edit by edit.

This is why one number deserves a moment of stillness. Alek and the Open Future team measured what many of you already felt: last year, human visits to Wikipedia fell eight percent, while bot traffic rose fifty.

For everyone who has dedicated decades of evenings to this project, that is not a statistic. It is grief. Value now leaves through the same door that knowledge used to enter by — and nothing is deposited on the way out. Your invitation calls this a dark commons. Exactly right: darkness is what happens when answers stop showing their sources, and sources stop seeing their readers.

So here is what I am working on. I keep — and keep is the correct verb — a Kami. In Japanese folk tradition, a Kami is a small spirit of one particular place: one shrine, one river, one household. Never the whole world. I borrow only the shape of that idea, with respect and no claim on anyone's religion. My Kami is an acronym:

Knowledge Artefact Management Intelligence

This talk was prepared with such an intelligence, so let me describe its working conditions — the exact inverse of extraction. Locality keeps secrets; ledgers keep truth.

If any fact in this talk is wrong, the error is mine, not the tool's. Judgment stayed with me.

Why do I trust this pattern? Because Taiwan has run it at state scale. In 2014, roughly nine percent of the people approved of the government's digital services; by 2020, over seventy.

Nothing mystical happened in between. We ran Wikipedia's protocol inside public administration: publish first, verify together, assume good faith. And good faith is not naivety when every version is logged and every error can be reverted in seconds — it becomes an executable protocol.

And because this room deserves honest numbers more than flattering ones, here is my honest one. In 2024 we convened an Alignment Assembly on deepfake scam advertising: two hundred thousand SMS messages, four hundred forty-seven people giving their time in forty-four small groups. After deliberating, support for one flagship measure — requiring platforms to reveal how their algorithms work — fell by twenty-seven and a half points, to fifty-five point six percent.

That is not the process failing. A process that only ever confirms what people already believed is not listening; it is performing. The assembly delivered a recommendation — never a vote on anyone else's behalf — while the government's bill was already moving through our parliament. Scam ads on the regulated channels did eventually decline, but that took the statute, the reporting infrastructure, the ad-scanning systems, and platform enforcement working together. What deliberation earned was subtler and more durable: a public that had genuinely reconsidered, and a state that had to respond.

MOSAIC's second question asks what a Public AI stack looks like, and how to resource it. People keep telling me the answer must be enormous. I always say: The singularity may be near — but the plurality is here.

The stack does not need one superintelligent centre of gravity. It needs millions of small, bounded, interoperating Kami stewards — one per archive, one per language, one per community — each accountable to the people nearest it; each inspectable, correctable, retirable. That is not a smaller ambition. It is a different topology, and it is the only topology that survives contact with a commons.

So when the resourcing conversation comes this afternoon, my suggestion is: fund the boundary conditions.

Wikipedia already proved these norms scale for humans — twenty-five years, at planetary scale. The unfinished work — the work this consortium could genuinely lead by building — is standardizing the same norms for agents. That is not a research agenda. It is a specification, and the people in this room hold most of the pieces.

You have three hours ahead of you, and two questions on the table: governance of the new knowledge loop — what rules should govern this ecosystem, and how the commons and mission-aligned AI developers can collaborate to keep it flourishing; and infrastructure for a Public AI stack — what we build together, and how we resource it.

I have tried to answer with one word and one demonstration. The word is Kami: small, bounded, accountable — kept. The demonstration is that reciprocity is not a sentiment we request from AI systems. It is a property we can engineer. I know, because I work inside it every day.