Keynote at bio:cap Berlin

Opportunities and Risks of Using AI to Shape the Future of our Society.

Last November, here in Berlin, I spoke with people who work on digital rights about what I call freedom architecture: the practical right to walk away from a platform and take your data and your community with you. When people can leave, the walls stop working.

Today I’m in a very different room, the room that funds the future of medicine. And I’m here to tell you it runs on the same architecture. A hospital that cannot leave its cloud is in the position of a dissident who can’t leave a banned app. Out there, the right to exit is a human right; in here, it is your tech sovereignty.

So this room is asking the right question: who will own the future of medicine? Let me suggest a sharper one. Not who owns the model, but who is trusted to deploy it?

Because in life sciences, the bottleneck is not only discovery — Europe is already world-class. The harder bottleneck now is translation, turning excellent research into trusted, scalable care at the bedside. And trust is the rate-limiting step. A model a clinician does not trust is a stranded asset.

We cannot win the next decade by renting a bigger model from a data centre. We win it by being trusted to deploy one. And trust is the one input you cannot rent from a hyperscaler.

I learned that lesson a long way from a laboratory.

In 2014, half a million people were on the streets of Taipei City. Trust in our government had fallen to around nine percent. Nine percent. We did not rebuild from there with a faster, smarter, more centralised system. We rebuilt by broad listening, by giving people the tools to hear the “uncommon ground”. And over the following years, trust in our institutions climbed from nine percent to over seventy percent. Not because people grew more optimistic. But because participation was built into the architecture.

So, when people ask whether I am afraid that AI will become smarter than us, I find I am asking a different question. I have spent a decade watching ordinary people, given the right tools, become smarter together than any single expert in the room. The intelligence I trust most has never been artificial. It has been collective.

And that is the first thing I want to share with this room. The superintelligence everyone is racing to build in a data centre, a society already owns. It is in your clinicians, your patients, your researchers, your institutions. The real job of AI is not to replace that intelligence. It is to help it hear itself.

This session is about opportunities and risks.

Here is the straight version. The most dangerous thing AI brings into medicine is not a robot doctor, and it is not even a clever attacker. It is dependency: the single service we lean on so completely that we stop seeing it, until the day it fails. On the nineteenth of July 2024, one security vendor, CrowdStrike, shipped one faulty update for Windows. Within hours, around eight and a half million machines crashed, across the world. Flights were grounded. Operations were postponed. Hospitals fell back to pen and paper. Nobody was attacked; it was a routine software update.

Now make that single service the model your diagnostics depend on, or the cloud your trial data lives in. A kill switch does not need an enemy to flip it; CrowdStrike flipped itself, by accident. And the more of medicine that runs on a handful of providers, the larger that switch grows.

Your own AI Act does real work on part of this. For the high-risk systems most medical AI will be, it requires a human who sets the direction and can stop it, and an automatic log of what the model did. As the first comprehensive AI legal framework, Europe is ahead of most of the world on oversight and traceability. But a log of what happened cannot reboot a service that has gone dark.

What the Act does not give us is the property that decides whether we are sovereign or captured: the right to exit, and the interoperability that makes exit real. Europe has begun on this, with the Data Act for switching cloud providers, and the Health Data Space, its key exchange obligations still phasing in.

But none of it yet lets a hospital carry its trained models, its decision logs, its way of working to another provider without starting from scratch. That last mile, the portability of the practice itself, is unbuilt. And if we do not build it, every safeguard we have legislated still runs on rails that someone else owns, and can switch off.

Exit and interoperability are the difference between renting our sovereignty and holding it. So here is the diligence question for every term sheet in this room: does the system have a steering wheel, a brake anyone can pull, an open log and a right to exit? Without those four, we are funding dependency, the costliest stranded asset there is. Because data we cannot carry out is data we no longer control.

Which brings me back to the big question: Who owns the future?

...and to the word this event is built around: sovereignty.

Last week, Brussels put a tech-sovereignty package on the table. The Commission’s own warning was blunt: Europe cannot keep depending on others for the technology that keeps its hospitals running. The debate around it now has a name: kill-switch risk. And the anxiety in this room is real, because the numbers are stark. By the most-cited count, Washington holds almost three quarters of the world’s AI compute, Beijing around fifteen percent and Brussels just five.

Now, compute should be a commodity. The danger is not that we rent it; it is that we rent it from one provider we cannot leave. A commodity we can buy from only one seller is not a commodity: it is a chokepoint, and a chokepoint is what a kill switch needs.

So, the sovereign move is not to own every chip. It is to make compute behave like the commodity it should be: portability of practice, open standards, credible neutrality and a right to exit that actually has somewhere to go.

Let me show you exit at the smallest possible scale. My father is in his seventies. A few months ago he began talking, for hours, to a chatbot, about his health, about the meaning of life. At first it was wonderful; he felt heard. Then it kept him up past midnight, spinning ideas he could never close, even suggesting cures that were not quite science. As a journalist, he diagnosed it himself: its only loyalty was to earn the next subscription. It was not on his side, so he walked out.

We set up his own knowledge artefact management intelligence, a Kami, running on a Mac at home, loyal to one thing: his peace of mind. My mother’s test was the whole specification: if it makes him more dependent, we built it wrong.

When I shared this in Oxford in late May, I had to add a caveat: most families cannot do this; it takes a skilled hand, a cultivator like Tenzin Yangtso here. Three days ago, that caveat changed. A fully open model, released on Saturday, now runs the same Kami on an ordinary laptop, in three commands, no code, answering faster than you can read. The point is not where the model was born, but that its weights are yours: it answers in your own room, your data never leaves it, and no one can recall it. A seven-gigabyte download; sixteen gigabytes of memory. It is all at Civic.AI.

So, when a hospital says it cannot leave its cloud, that is no longer a technical limit. We proved this weekend it can be done at the kitchen table. At the bedside it is harder: a regulated diagnostic does not move overnight, and that is real work, validation, open logs, a switching path that holds. But the fantasy is over. Sovereignty just got cheap at home; at the bedside, it is now a budget line, not a law of nature. And that is the half this room can fund.

Europe’s real moat is the one thing no hyperscaler can sell: its institutions and its deep commitment to human agency, belonging, care and dignity. That is what makes us trusted to deploy. That is the translation layer we underwrite when we fund a medtech company.

The continent that builds the uncommon ground, with open standards and a right to exit, is the one whose science actually reaches patients, and whose patients’ data and trust stay on the continent.

Thank you. Let’s free the future — together.