Sovereign AI: The Ability to Choose
There is a simple test for whether something is truly under your control:
Who gets to say no?
Imagine that tomorrow an artificial intelligence capability important to your government becomes unavailable. Not because it failed, but because someone outside your country decided you could no longer use it.
Perhaps access to a model was revoked. A cloud service was suspended. A software license changed. A critical processor could no longer be supplied. Or a geopolitical dispute suddenly became a technology problem.
That is not a hypothetical risk. In late August 2026, OpenAI said it would wind down the contract supplying its models to Cursor after Cursor was acquired by SpaceX, with a proposed November 12 cutoff. Cursor may continue as a product, but users who depend on OpenAI models through it must adapt.
That is the difference between owning an interface and controlling the intelligence on which it depends.
How much of your government stops with it?
That is a Sovereign AI question.
A nation does not need to build every component of artificial intelligence itself. Few countries can. AI depends upon a global chain of semiconductors, energy, data centers, networks, software, models, data, and specialized talent.
But every critical dependency creates risk.
The risk is not dependency. The risk is in the dependency you cannot manage.
A nation can rely on international technology and remain sovereign. But when losing one outside supplier can disable a critical national capability, that dependency has become a sovereign security issue.
Every critical dependency should therefore have an answer to a simple question:
What happens if we lose access to it tomorrow?
Your Data Goes Somewhere
There is another side of the equation that is easy to overlook.
Every time we use AI, we may also give it information. Data sovereignty begins with understanding that outward flow: what leaves citizens, businesses, and government systems, and who receives it.
A business might submit intellectual property or a confidential contract. A government employee might provide a sensitive document. A citizen might upload a passport or national identification card and ask an AI for help completing a government application.
Where did that information go?
Was it retained? Who processed it? Can it be used for training? Who can access it? Under whose laws? Can it be deleted?
These are different questions, and answering one does not answer the others.
Keeping data physically inside a country can reduce some risks, but geography alone does not establish control. Infrastructure located domestically may still be operated, influenced, or controlled by a foreign provider and subject to another country's laws.
Data residency is only part of sovereignty. It is not sovereignty by itself.
For a nation, protecting information therefore means understanding not only where data is stored, but where it is processed, who controls the infrastructure, what happens to the information after processing, and which jurisdictions can reach it.
That is one side of AI sovereignty: protecting the information that goes into the system and retaining meaningful control over how it is used.
Sometimes Sovereignty Requires a Smaller AI, Not a Bigger One
Much of today's AI race focuses on bigger models, more computing power, and increasingly capable frontier intelligence.
But a government does not need the world's largest AI to perform every job.
A tax assistant does not need to solve advanced physics. An agricultural AI helping identify a crop disease does not need to write a screenplay. A remote government kiosk may be more valuable if its AI understands local languages, operates without an Internet connection, and keeps sensitive information nearby.
Small Language Models, or SLMs, make this increasingly practical. Smaller models can require less computing power, specialize around particular knowledge and tasks, and increasingly operate locally in computers, vehicles, kiosks, phones, and other devices.
Larger frontier models remain enormously important. Their breadth and reasoning capabilities may help tackle difficult, global problems in medicine, science, engineering, and fundamental research.
A nation will likely need both.
Sometimes sovereignty requires a smaller AI, not a bigger one.
Keep routine sovereign intelligence close. Reach outward for exceptional intelligence when the problem requires it.
As useful intelligence becomes cheaper and more abundant, simply having access to AI becomes less of an advantage.
The scarce resource may increasingly be trusted context: your knowledge, your data, your identity, your rules, and the authority to act.
Intelligence Has an Inheritance
There is another side of AI sovereignty: the intelligence that comes back. Cognitive sovereignty concerns the context, assumptions, and patterns through which an AI interprets what it is asked.
Data sovereignty asks where information goes. Cognitive sovereignty asks what history and assumptions shape the intelligence interpreting it.
Control also begins before an AI answers its first question.
In 2026, researchers created Talkie, a modern 13-billion-parameter language model trained on approximately 260 billion tokens of material published before 1931.
Talkie makes that point concrete.
A model trained only on material available before 1931 would say that the United States has forty-eight states. It could also describe Czechoslovakia as a single country. Both answers would be correct about the world represented in its training.
Today, the United States has fifty states, and Czechoslovakia has become Czechia and Slovakia. The model would not be trying to mislead anyone. Its understanding would simply be the point of view before the world changed.
But the more important limitation is not always that the facts become old. Sometimes a model inherits the assumptions embedded in its training.
Imagine asking that same model about cigarettes. The world represented in its training included advertisements that invoked doctors, health, energy, and safety. That does not mean every doctor agreed, or that those claims were true. It means the model could absorb the confident language of a society that had not yet confronted the harm.
The model would not need to be deliberately biased. Its answers could simply reflect what its sources treated as credible.
That is the deeper issue. A model does not inherit only facts. It can also inherit the language, assumptions, commercial influence, and authority signals of the world from which its training material was drawn.
This does not mean that every older model is unreliable, or that a country must build every model from scratch. It means that no model is neutral outside the information and assumptions embedded within it.
A sovereign AI strategy begins by knowing what is already inside a system, what can be governed around it, and when a different model, training approach, or local alternative is necessary.
Open-weight models give nations an important ability to possess, operate, and adapt existing models without depending upon foreign systems for every interaction. But downloading the weights does not erase the history that created them.
A nation should therefore care not only about where its AI operates, but also about where that intelligence came from and how it has been evaluated in its own context.
A nation can also build controls around a model. Retrieval systems can ground answers in authoritative national information. Policies can restrict unsafe outputs. Permissions can determine what a model may access or do. Oversight can require accountability when the system affects citizens.
Those layers, or scaffolding, can shape how an AI behaves in practice. They can prevent a model from repeating a dangerous or false premise. But they do not mean the model has forgotten what it learned. They create a governing layer around the intelligence.
That layer can protect a nation's laws, culture, and democratic values. It can also become a mechanism for suppressing information or enforcing unaccountable power. Sovereignty gives a nation the ability to make these choices. Citizen sovereignty requires that consequential choices remain transparent, proportionate, and contestable.
Training is not the only way a digital system can shape what people see. In August 2026, after President Trump directed U.S. federal naming systems to use "Lake America" for Lake Ontario, Google Maps began showing "Lake America" to users in the United States, "Lake Ontario" in Canada, and both names elsewhere.
The lake did not change. Google's display reflected its own policy of using official geographic data relevant to a user's location. But the default digital context seen by millions of people did change, without each person choosing that label.
That is a smaller version of the cognitive-sovereignty question. It is not only who trains the model. It is also which data sources, policies, jurisdictions, and system defaults determine the frame through which people encounter information.
When AI Stops Talking and Starts Acting
Most of us first experienced generative AI as a conversation. We asked a question and AI gave us an answer.
That is changing.
Agentic AI can increasingly access databases, operate software, communicate with other systems, update records, initiate transactions, and carry out actions on our behalf.
For governments, this changes the question from what does the AI know? to what is the AI allowed to do?
There is a significant difference between an AI recommending that a citizen benefit be denied and an AI having the authority to deny it, change the citizen's record, notify another agency, or initiate a financial action.
The AI may help reason through the decision. But identity, credentials, permissions, financial limits, access to government systems, and authority to execute consequential actions can remain outside the AI itself.
That question becomes even more concrete when AI moves from software into the physical world. A chatbot can give bad advice. An AI inside a vehicle, robot, medical device, factory system, or critical-infrastructure platform can move, stop, unlock, inspect, deliver, or damage something in the real world.
A nation may choose to use intelligence developed elsewhere. But the authority to release physical action, define safety limits, identify the responsible operator, and stop the system when necessary should remain under control it can verify.
More broadly, the authority to decide what that intelligence may access, what it may do, and when it must stop should remain under sovereign control.
The 2026 dispute between Anthropic and the U.S. government shows the same issue from another direction. Anthropic continued to support national-security work, but maintained restrictions on mass domestic surveillance and fully autonomous weapons. Whether one agrees with those restrictions or not, when a government depends on a hosted model, the provider's policies can become part of the government's operational boundary.
That does not resolve what national policy should be. It shows why authority, safety rules, and a credible alternative path cannot rest solely on a supplier's terms.
Sovereignty Belongs to Citizens Too
There is one more party in this discussion: the citizen.
A government could control its AI infrastructure, data, identity systems, and agents while giving citizens little ability to know when AI is making consequential decisions about them or how their personal information is being used.
An AI system can be independent from foreign control while still exercising too much unaccountable control over the people it is meant to serve.
As AI becomes increasingly personal and begins acting on our behalf, another simple question becomes important:
Whose AI is working for whom?
National sovereignty should protect individual agency, not replace it.
The Ability to Choose
Sovereign AI is not a product a government can simply buy. Nor is sovereignty established because a vendor puts the word sovereign on an AI platform or cloud service.
Sovereignty is a condition a nation must be able to verify.
Sometimes that will require domestic capability. Sometimes it will require trusted international partners, multiple suppliers, open technologies, reserve capacity, or the ability to move to another provider. Sometimes a small local AI will be sufficient. Another problem may justify reaching for the most powerful intelligence available anywhere in the world.
The objective is not technological isolation.
It is strategic choice.
That balance is already appearing in national policy. In September, the United Kingdom addressed the risk of losing access to frontier AI models developed overseas by describing a three-part approach: build domestic capability where the country has strengths, work with trusted partners where cooperation improves access and influence, and secure access to frontier capabilities where those capabilities are essential.
That is not technological isolation. It is strategic choice.
Understand your dependencies. Protect the information that enters AI systems. Understand the knowledge and context that shape the answers that come back. Protect the identity, security, and authority that cannot safely be surrendered. Build the people and institutions capable of understanding these systems. And preserve another path when a dependency becomes a vulnerability.
Sovereign AI is the ability to understand your dependencies, control and protect what matters, and choose another path when necessary.
In the end, the test is remarkably simple:
Do you still have the ability to choose?
~ Mark Munger, CTO, Brightside
Comments