37 countries described the problem. Nobody named the solution.
Yesterday, the WHO held a global conference on AI in health in Lisbon. 37 countries showed up. Ministers, WHO leadership, the European Commission. For seven hours, they described the same problem from different continents, in different words, and with different examples. But it was always the same problem.
Dr. Hans Kluge, WHO Regional Director for Europe, called it data colonialism:
"We must also avoid what some have described as data-colonialism. Where data generated in one country create value in another country without bringing benefits to the people who created them."
Dr. Alain Labrique from WHO HQ defined sovereignty as knowing who controls your model:
"Sovereignty is not isolation. Sovereignty is responsibility, knowing where your data sits, who processes it, who laws govern that data, and on whose terms that model can be taken away from you."
He went further. A ministry that cannot inspect its own AI does not own it:
"A ministry that cannot inspect a model, cannot audit its performance in its own population, and cannot withdraw it from service, does not own that system. It rents it. It inherits it. And the owner of these systems may not be accountable to your government. And that is not sovereignty."
Guinea-Bissau's health minister said they need investment that builds national capacity "rather than creating long-term dependence on external vendors." Timor-Leste demanded open standard mandates and warned that vendors "must not treat the developing nation as a testbed for unverified algorithms." India called for "models that are sovereign and also rooted on the Indian data." Brazil framed AI as a public good. Dr. Kluge closed by insisting that AI must be shaped "as a shared public good, never as a source of greater inequality."
Then Kyrgyzstan told a story that stuck with me. They piloted an AI system for interpreting vascular images. The AI's conclusions didn't match what their doctors were seeing. They dug in and found the problem: the model had been trained on a population living near sea level. Kyrgyzstan's population lives between 1,000 and 4,000 metres. The vascular system is different at altitude. The model was confidently, fluently wrong. What saved patients was not the technology. It was a clinician who trusted their own judgment over the machine.
I've been writing about this for years. The difference now is that 37 countries are saying it too.
The assumption nobody questioned
Every statement from that conference was correct. Every concern was real. But they all shared one unspoken assumption: that AI has to be a central node.
A model that lives in a data centre. Data that travels to it with every interaction. Value that accumulates where the model sits, not where the patient sits. A relationship where each query, each diagnosis, each conversation deepens dependence on a system the user can never inspect, never audit, and never own.
In a centralized architecture, each interaction decreases sovereignty. Every time a clinician sends a patient's data to a remote model, that data leaves their control. It creates value for the model's owner. It deepens the dependency. It makes the model smarter about someone else's population, and sometimes, as Kyrgyzstan discovered, it makes the model dangerously wrong about yours.
This is not a policy failure. It is an architecture failure. No regulation, no procurement clause, no open standard mandate can fully restore sovereignty over a system whose fundamental design requires data to leave the user's control. You can write the best data protection law in the world. If the architecture sends the data to a server in another country before it can produce a single output, your law is catching up to a design decision that already made the choice for you.
The policy world is circling the concept. Sovereignty, inspection, auditability, public goods, open standards, avoiding vendor dependence. These are all real and important. But nobody at the conference made the connection explicit: the mechanism that delivers all of them is on-device, open-source AI. The language hasn't caught up to the solution.
Sovereignty is an architecture decision
I have spent years building toward a different architecture. At Isaree, the model runs on the clinician's device. The data never leaves the phone. Each interaction increases capability without decreasing sovereignty, because there is no central node extracting value, accumulating dependence, or quietly drifting away from the population it was meant to serve.
This isn't a feature. It's the foundation. And it addresses every concern raised in Lisbon:
Inspection and audit. You cannot inspect a black-box cloud model. Labrique said a ministry that can't inspect its model doesn't own it. He's right. On-device, the institution can inspect, audit, and withdraw the model at any time. Ownership is built into the architecture, not negotiated in a contract.
Data colonialism. Kluge warned about data generated in one country creating value in another. When data never leaves the device, there is no data flow to colonize. The problem doesn't need regulating. It needs designing out.
Performance drift. The Kyrgyzstan case is not a one-off. Models trained on one population fail on another. On-device models can be adapted and validated on the local population, by the clinicians who know that population. The altitude problem is an architecture problem, and on-device architecture solves it.
Vendor dependence. Guinea-Bissau and Timor-Leste called for local ownership and against external dependence. Open-source, on-device, connected through open protocols. No lock-in. No recurring cloud fees. No model that can be withdrawn by a vendor on the other side of the world. Local ownership by design.
Public good. Brazil and Dr. Kluge framed AI as a shared public good. An open ecosystem of clinician-built, open-source agents is the practical realization of that frame. Not a proprietary asset extracted from the people who generate the data.
The diagnosis is correct.
The prescription is wrong.
The WHO conference made the diagnosis. 37 countries described the symptoms with remarkable consistency. But the prescriptions being proposed, better regulation, stricter procurement, open standard mandates, are all attempts to regulate your way back to sovereignty over a system that was designed to take it from you.
You can't do it. Not fully. Not as long as the architecture requires data to leave the user's control before it can produce value.
Sovereignty is not something you negotiate with a vendor. It is something you build into the architecture, or it is something you will forever be trying to regulate back.
I've been saying this for years. The difference now is that 37 countries just spent seven hours proving the problem is real. What they haven't realized yet is that the cure isn't better terms of service.
It's a different architecture.
check our website and see the difference: www.isaree.ai
We're building it.
Your,
Bart