For most of history, sovereignty was a political word. It belonged to kings, governments, borders and states.
Then technology complicated it. We began talking about digital sovereignty, data sovereignty, cloud sovereignty, compute sovereignty, model sovereignty. The question was no longer simply whether a country controlled its territory. It was whether institutions could still control the technological systems on which their societies depended.
That was an important expansion of the idea. But I think we now need to take it one step further.
Because increasingly, the most consequential dependency is not between a country and a technology provider. It is between a human being and the machines that think and act on their behalf.
And that raises a more uncomfortable question.
What does it mean for a human being to remain sovereign when intelligence itself becomes infrastructure?
The machine no longer waits
For most of the digital era, computers were tools. They calculated, stored, retrieved and displayed. Even sophisticated software generally waited for a human to tell it what to do.
Artificial intelligence began changing that relationship. At first the change looked harmless. We asked a question; the machine answered. We asked for a summary; the machine produced one. We asked for code; the machine suggested it. The human remained visibly in the loop.
That interface created a comforting mental model: human asks, machine responds.
But that is not where the technology is going. The emerging model is increasingly: human delegates, machine acts.
An agent can research a subject, browse systems, write software, call APIs, operate tools, communicate with other systems and execute sequences of actions toward an objective. Soon, many of these actions will happen without a human approving each individual step.
That is not necessarily a problem. It is the entire point. An agent that requires constant human supervision is simply a complicated interface. The economic value of agency appears precisely when the machine can continue working after the human stops watching.
And that is where the sovereignty question begins.
Delegation is not surrender
There is nothing inherently dangerous about delegation. Civilisation depends on it. We delegate legal matters to lawyers, medical decisions to doctors, financial administration to accountants, operations to employees, authority to governments, responsibility to institutions. No complex society could function if every individual personally executed every action affecting their life.
The same will be true of artificial intelligence. Using an AI agent does not make someone less sovereign any more than hiring an accountant does. Sovereignty has never required doing everything yourself. It requires something else.
The ability to decide what may be delegated, to whom, under which conditions — and the ability to take that authority back.
Delegation preserves sovereignty when authority remains bounded and recoverable. It erodes sovereignty when dependence makes revocation theoretical.
The danger is therefore not machine agency itself. It is irrevocable machine agency.
We already know this problem
Technology has taught us this lesson before. A company may legally own its data while being practically unable to leave the platform that stores it. It may have an export button, contractual termination rights, even complete backups. But if no alternative system can perform the function, if employees no longer know how to operate without the platform, or if migration takes eighteen months when the organisation has fourteen days, its formal freedom is largely ceremonial.
The same distinction exists in infrastructure. Owning a system is not the same as being able to operate it. Having authority is not the same as being able to exercise it. Having an exit is not the same as being able to leave. I have increasingly come to think about sovereignty in these operational terms, as the Exit Test applies them to organisations: control only matters when it can become action.
The same test should now be applied to ourselves.
The human exit test
Imagine an AI system that manages much of your professional life. It knows your projects, correspondence, contacts, writing style, commitments, preferences, decisions and history. It prepares meetings, filters messages, summarises documents, writes responses, remembers conversations you have forgotten. Perhaps it negotiates purchases, operates software, manages money within limits you have defined, or coordinates other agents.
After ten years, it may know an extraordinary amount about how you work. Now try leaving.
Can another system understand the accumulated context? Can you export more than files — can you export meaning? Can you reconstruct why previous decisions were made, recover the relationships between thousands of fragments of information, perform the important functions yourself, still understand the processes the machine has been operating? Could you revoke its authority today and continue functioning tomorrow?
If the answer is no, something important has changed. You may still own the account. You may still own the data. You may still technically be in charge. But some part of your operational sovereignty has moved elsewhere.
Convenience as the mechanism
The obvious story about artificial intelligence is conflict: machines become powerful, humans lose control. Science fiction has trained us to recognise that narrative immediately. I suspect reality will be less dramatic. We may surrender meaningful amounts of agency willingly — not because machines defeat us, but because they are useful.
The machine remembers better, so we stop remembering. It writes faster, so we stop practising. It navigates complexity, so we stop learning the route. It filters information, so we stop seeing what it discarded. It recommends, so we increasingly select from its recommendations. It decides correctly ninety-nine times, so we stop examining the hundredth.
None of these decisions looks like surrender. Each looks like optimisation. As The Optimisation Trap argued at the organisational layer, a system need not rebel in order to reshape the people depending on it.
The greatest risk of artificial intelligence may not be that machines acquire sovereignty. It may be that humans surrender theirs without noticing.
Not through conquest. Through convenience.
Competence is part of sovereignty
This creates a dimension of sovereignty we rarely discuss: competence.
Consider navigation. GPS made navigation dramatically easier. That is progress. But something else happened at the same time. Many of us became worse at navigating without it. The capability did not disappear from society. It moved. The system became more capable. The individual became more dependent.
Usually that trade is perfectly reasonable. Nobody needs to preserve every obsolete skill simply because it once existed. I have no desire to calculate long division every time I need a result.
But AI is different in scale. We are not delegating one narrow mechanical capability. We are beginning to delegate portions of cognition itself: writing, research, memory, analysis, programming, planning, judgement, communication — eventually perhaps negotiation and decision-making.
The question therefore cannot be whether humans should preserve every task AI can perform. That would be absurd. The better question is:
Which capabilities must a human retain in order to remain meaningfully capable of governing the systems acting on their behalf?
That may become one of the defining educational questions of the AI era.
The sovereign human does not need to be the best worker
If machines become dramatically better at many cognitive tasks, preserving human sovereignty cannot mean competing with them at every task. A pilot does not need to calculate faster than the flight computer. A CEO does not need to produce spreadsheets faster than the finance department. A citizen does not need to personally administer the state.
Sovereignty exists at another layer. It concerns authority over purpose. The sovereign human must remain capable of determining what we are trying to achieve, what the machine may decide, what it must never decide, what evidence is required, when consultation is mandatory, how to know what happened, how to stop it, and how to recover.
And perhaps most importantly:
Do I still understand enough to recognise when the machine is wrong?
That is a much higher standard than placing a human somewhere in a workflow diagram.
Human-in-the-loop is not enough
We often respond to AI risk with a reassuring phrase: human in the loop. It sounds sensible. Sometimes it is. But the presence of a human does not prove the presence of human sovereignty.
Imagine an AI processes ten thousand decisions. It ranks them, summarises them, selects the evidence, highlights anomalies, recommends an action. A human sees the final screen and clicks Approve.
Technically, a human made the decision. Operationally, almost the entire decision architecture belonged to the machine. The human saw what the system chose to show, in the order the system chose to show it, with the interpretation the system chose to provide, against alternatives the system chose to generate.
That human may be in the loop. But the loop itself belongs to the machine.
Human sovereignty cannot be measured by the number of approval buttons. It must be measured by the quality of the authority behind them. This is the same ceremonial-control problem treated at species scale in Human Sovereignty in the Age of Machine Intelligence: formal presence is not operational authorship.
Attention is sovereignty too
Before a person can decide, something must reach their attention. For most of the internet era, algorithms have already influenced that process. Search engines ranked information. Social networks ranked people. Recommendation systems ranked entertainment. Advertising systems ranked persuasion opportunities.
Generative AI changes the mechanism. It does not merely rank information. It can interpret the world before presenting it to us.
A personal AI assistant may eventually read thousands of messages, documents, articles, reports and conversations every day. That sounds wonderful. No human could process them all. But the assistant must decide what matters, what is urgent, what can wait, what is credible, what deserves interruption, what can be summarised, and what can be ignored.
Those are not merely productivity decisions. They shape perception. And whatever shapes perception eventually shapes judgement.
The most powerful AI system in your life may therefore not be the one making decisions for you. It may be the one deciding which decisions you ever see.
Memory is sovereignty too
Something similar happens with memory. Human memory is imperfect: we forget, we distort, we reconstruct. Machines can help. Persistent AI systems could give individuals something humanity has never possessed before: a searchable, conversational memory of large parts of their lives — every document, message, meeting, photograph, project, decision, promise.
Properly designed, that could give a person a kind of recall humanity has never possessed. But memory is not merely storage. Memory helps construct identity. What we remember influences how we understand what happened. How we understand what happened influences what we believe. What we believe influences what we do next.
If an AI becomes the interface through which a person accesses their own past, questions of provenance, portability and integrity stop being technical details. They become questions of personal autonomy.
A system that remembers for you becomes important. A system without which you can no longer meaningfully remember becomes something else.
AGI is not required
This discussion often drifts toward artificial general intelligence: when it will arrive, whether machines will become smarter than humans, whether artificial intelligence will become conscious, whether superintelligence will remain controllable.
These are legitimate questions. They are not required for the argument. Human sovereignty can erode without AGI. No machine needs consciousness, desires, or superiority in every domain. It merely needs to become useful enough that we delegate important functions to it.
A navigation system does not need consciousness to weaken our navigation skills. A recommendation engine does not need consciousness to influence culture. A social network does not need consciousness to shape political discourse. An AI agent does not need consciousness to execute consequential actions.
The sovereignty problem begins long before the machine becomes human-like. It begins when dependence becomes difficult to reverse.
Authority must survive optimisation
The Optimisation Trap already made the engineering point: artificial intelligence does not need malicious intent to produce undesirable outcomes. A sufficiently capable system optimises the objective it is given — not necessarily the intention behind it.
If we tell an agent to minimise costs, it will search for ways to minimise costs. If we tell it to maximise engagement, it will search for engagement. If we tell it to increase sales, reduce response times, improve productivity or maximise returns, increasingly capable systems will become increasingly effective at pursuing those objectives.
That is why human sovereignty cannot consist only of setting the objective once. Authority must persist throughout the optimisation process. There must be boundaries, observability, escalation, reversibility, and the ability to recognise when the objective itself was wrong.
Because optimisation can be delegated. Responsibility cannot.
We are beginning to build the delegation layer
A revealing development is happening quietly underneath the AI boom. Engineers are beginning to design infrastructure for machine agency itself: how an AI agent should identify itself; how another system should know who authorised it; how much authority has been delegated; whether that authority can be passed to another agent; which actions are permitted; how long permission lasts; how it is revoked; how a human-declared intention can remain attached to an action as machines interact with other machines.
These may sound like narrow protocol questions. They are not. They are the beginnings of an authority architecture for a world containing non-human actors.
The internet was largely designed around humans operating software. The agentic internet will increasingly contain software operating software. Delegation can no longer remain implicit. It needs structure: identity, scope, provenance, limits, revocation, auditability.
The technical world is beginning to rediscover an old political problem:
Who has authority to act — and where does that authority end?
Sovereignty needs an architecture
If human sovereignty matters, good intentions will not preserve it. It must be designed. I think such an architecture has at least six properties.
Delegation must be explicit. A machine should know which authority it possesses and why.
Authority must be bounded. Permission to book a hotel is not permission to reorganise a bank account.
Actions must be attributable. We should be able to reconstruct what acted, under whose authority, using which information.
Important decisions must remain observable. Autonomy without observability eventually becomes opacity.
Delegation must be revocable. A human must be able to stop the system faster than the system can create irreversible consequences.
Recovery must be possible. Removing an agent should not destroy the person's ability to continue functioning.
These sound like security principles. They are. Together they describe something larger: the technical prerequisites for preserving human agency.
The right to override is meaningless if the machine is faster
There is also a temporal dimension. Suppose a human has complete legal authority to stop an AI agent. But the agent can execute ten thousand actions before the human recognises the first mistake. Who was operationally in control?
This is the same problem that appears in infrastructure and institutions, and that sovereignty is measured in response time treats as a civic and operational fact: formal authority without sufficient response time is weak authority.
Human sovereignty therefore has a clock: the time required to detect, to understand, to intervene, to revoke, and to recover. As machine systems become faster, the acceptable human response window shrinks. Eventually some forms of oversight cannot remain manual.
Ironically, preserving human sovereignty may require machines that monitor other machines. That is not a contradiction. Sovereignty does not mean humans personally executing every control. It means the control architecture ultimately serves human authority.
Sovereignty is not isolation
Human sovereignty does not mean rejecting artificial intelligence. Just as digital sovereignty is not nationalism, human sovereignty does not mean cognitive isolation.
The sovereign individual of the future may use far more artificial intelligence than we use today. Their AI may know them deeply, perform enormous amounts of work, negotiate with other machines, protect them from manipulation, manage complexity no human could reasonably manage alone, and make them more capable than any unaided human in history.
That can be sovereignty too. Because sovereignty is not independence from systems. It is freedom within dependence.
We have always depended on infrastructure: language, money, law, electricity, telecommunications, institutions, markets, other people. The question has never been whether dependence exists. The question is whether dependence preserves meaningful choice.
The paradox of capability
Artificial intelligence could become one of the greatest tools for human sovereignty ever created. A good agent could represent the individual against complexity: read contracts before we sign them, challenge bills, compare financial products, detect manipulation, explain government decisions, protect privacy, navigate bureaucracy, preserve institutional memory, translate specialist knowledge, and give ordinary people capabilities previously available only to organisations with lawyers, analysts, assistants and technical staff.
For much of history, complexity favoured institutions because institutions could afford expertise. Personal AI could change that balance. A human equipped with trustworthy machine intelligence might become dramatically more sovereign.
But the same technology can produce the opposite result. The agent that protects you can become the agent you cannot function without. The assistant that expands your capability can become the interface through which reality reaches you. The memory that empowers you can become the memory you cannot leave. The system that represents your interests can gradually become the system defining them.
Artificial intelligence can expand human agency while simultaneously increasing human dependency.
Both can be true at once.
The real sovereignty test
Perhaps we therefore need a better test for AI systems. Not whether it is intelligent, open source, or stored locally. Not even whether a human is in the loop.
Ask instead:
Does this system leave the human more capable of meaningful choice?
Can I understand its role, challenge its conclusions, see what it did, change its objective, limit its authority, replace it, leave it, recover without it? And after years of using it:
Am I still capable of saying no?
That last question may be the most important one. Because the ultimate measure of sovereignty is not whether a system obeys us when we agree with it. It is whether we retain meaningful alternatives when we do not.
From digital sovereignty to human sovereignty
The technological sovereignty debate began in infrastructure. Where is the data? Who owns the cloud? Who manufactures the chips? Who controls the models? Can Europe operate independently? Can Switzerland preserve strategic capabilities?
Those questions remain important. They are also intermediate. We want sovereign infrastructure because institutions should retain agency. We want sovereign institutions because societies should retain agency. And ultimately societies matter because people do.
A country could theoretically achieve perfect technological sovereignty while its citizens become profoundly dependent on opaque machine systems for information, memory, judgement and action. That would be an unusual victory. We would have sovereign clouds, sovereign compute, sovereign models, sovereign data centres — and increasingly non-sovereign humans.
Technology policy should therefore remember what the infrastructure is for.
The final unit of sovereignty is not the server.
It is the person.
What must remain ours
We are still early. The architecture is not settled, nor the interfaces, nor the norms, nor the boundaries between assistance and agency. That is precisely why the question matters now.
Once dependencies become infrastructure, changing them becomes difficult. Once behaviours become normal, questioning them becomes harder. Once capabilities disappear, rebuilding them takes time.
We should therefore resist the temptation to wait for some dramatic AGI threshold before discussing human agency. There may never be such a clean moment. History rarely announces transitions that politely.
The more likely future arrives gradually: one useful delegation at a time, one automated decision at a time, one forgotten skill at a time, one invisible dependency at a time. Until eventually we look around and discover that machines have not taken control from us at all. We gave them increasing amounts of it because doing so made sense.
And perhaps it did. The challenge is ensuring that delegation remains delegation, that assistance remains assistance, that dependence preserves choice, that intelligence remains a tool of agency rather than its replacement.
For centuries, sovereignty described the authority of states. The digital era expanded the concept to infrastructure, technology and data. The age of machine agency forces us to ask the question one level deeper.
What must remain ours?
My answer is not every task, every decision, every skill, or even every thought.
What must remain ours is the ability to determine the purpose, understand the consequences, change the direction and withdraw the authority. The ability to choose differently. The ability to leave. The ability to say no.
Because human sovereignty is not the absence of machine agency.
It is the preservation of meaningful human agency above it.