There is a question we hear more and more often:

“Who controls artificial intelligence?”

But that is already the second question.

The deeper question comes first:

Who has the power to decide what artificial intelligence will do within society — and who has the power to hold those decision-makers accountable?

Because AI, by itself, is not political power. Power lies in the network of infrastructures, data, capital, computing resources, algorithms, corporations, public institutions, and political decisions that determine where, how, and for whose purposes AI will be used.

So the problem is not simply to put “rules on AI.”

The deeper challenge is to build democratic institutions capable of governing a new form of technological power.

And this changes the entire conversation.

1. The Wrong Picture: AI as a “Mind” That Someone Must Control

Public debate often personifies artificial intelligence.

We talk about what AI “wants,” what a model “thinks,” or whether AI might “escape.”

These may be interesting philosophical questions. But socially and politically, they can distract us from something more important.

A model does not decide by itself to use AI for hiring, credit scoring, policing, education, advertising, public services, or military applications.

Someone deploys it. Someone purchases it. Someone defines its purpose. Someone chooses the data. Someone sets the limits. Someone benefits. Someone bears the consequences.

The real political question, therefore, is not:

“How do we control an intelligent machine?”

It is:

“How do we govern the social relations of power within which the intelligent machine operates?”

That is the much larger question.

2. AI Creates a New Problem for Democracy

Classical democracy knows, at least in principle, how to control political power.

We have:

elections,
parliaments,
courts,
independent authorities,
a free press,
trade unions,
civil-society organizations,
rights of appeal,
public scrutiny.

But technological power does not coincide with state power.

A private company may possess computing resources, data, expertise, and infrastructures that no individual public institution possesses.

At the same time, governments may deploy systems that citizens cannot meaningfully understand, while public institutions themselves may become dependent on privately owned technological infrastructures.

A new political problem therefore emerges:

a society may possess democratic institutions without possessing an equivalent democratic capacity to govern the technologies on which it increasingly depends.

That is the real gap.

3. Technological Power Is Not Located Only in the Algorithm

Imagine an AI system that evaluates applications for a public benefit.

Who controls it?

The answer “the programmer” is inadequate.

We must ask:

Who decided that the process should be automated?

Who selected the data?

Who defined what counts as a “normal” applicant?

Who decided what level of error is acceptable?

Who audits the algorithm?

Who has access to the code and the data?

Who can challenge the decision?

Who compensates the person who was harmed?

And most importantly:

Who decides whether the system should be used in the first place?

This is where the discussion moves from the ethics of AI to the political economy of AI.

4. Transparency Is Not Enough

We often hear:

“The algorithm must be transparent.”

But transparency alone is not democracy.

An algorithm can be completely transparent and still serve a socially contested purpose.

At least four different levels are therefore necessary:

1. Transparency

We must know that a system is being used, by whom, and for what purpose.

2. Explainability

We should be able to understand why it produced a particular result, to the extent that this is technically and legally possible.

3. Contestability

A citizen must be able to say:

“The machine got it wrong.”

And there must be a genuine process through which that claim can be heard and reviewed.

4. Accountability

There must be a specific person, organization, or public authority that ultimately bears responsibility.

If the answer to an error is:

“That is what the model produced,”

then we do not have accountability.

We have transferred responsibility to the machine.

And that is democratically dangerous.

The OECD AI Principles explicitly connect transparency with the ability of people affected by AI systems to understand and challenge outcomes, while treating accountability as a distinct principle. (oecd.org)

5. From “Human-in-the-Loop” to “Citizen-in-the-Loop”

Here we need another important distinction.

We can have a human-in-the-loop and still have no democratic control.

An employee may formally have the final decision.

But if:

they do not understand the system,
they do not have enough time to review it,
they assume that “the algorithm knows best,”
they have no meaningful authority to override it,
and they themselves are evaluated according to how closely they follow the system,

then the human is technically in the loop but, in practice, the decision has already been transferred to the machine.

The goal should therefore become:

citizen-in-the-loop — the citizen inside the loop of technological power.

The citizen should not appear merely as a “user” of technology.

The citizen must appear as a rights-bearing participant and co-author of the rules governing its use.

UNESCO’s Recommendation on the Ethics of Artificial Intelligence explicitly connects human oversight with responsibility, participation, education, accountability, and the protection of rights. (unesco.org)

6. Democracy Must Govern Not Only Outcomes, but Infrastructure

Here we reach perhaps the most difficult point.

If a society allows critical AI infrastructures to become concentrated in a handful of private actors, it may regulate applications without actually controlling the technological foundations on which those applications depend.

Democratic debate therefore has to extend to:

computing power,
data,
cloud infrastructure,
energy consumption,
access to research infrastructure,
standards and technical protocols,
foundation models,
public research funding,
public digital infrastructure.

Because whoever controls the infrastructure often possesses greater power than whoever controls the application.

7. The Answer Is Not “Let the State Take Over AI”

That would be an oversimplification.

The state is essential for creating and enforcing rules, but governments themselves can deploy AI in ways that restrict rights.

So the choice is not simply:

private power or state power.

The democratic objective is:

private and public technological power subject to social, legal, and institutional accountability.

That requires multiple checks and balances.

8. We Need Institutions of Technological Accountability

Imagine a system similar to what democratic societies have created for other powerful social functions.

Not one “AI Authority” controlling everything.

Instead, an ecosystem of oversight:

Independent supervisory authorities

With technical expertise, genuine enforcement powers, and adequate resources.

Independent technical auditors

Not merely corporate compliance departments.

Judicial review

Citizens must have meaningful ways to challenge automated decisions.

Parliamentary oversight

Legislators cannot effectively oversee technologies they do not understand technically.

They therefore need independent scientific and technical institutions capable of translating technology into political knowledge.

Social oversight

Trade unions, civil-society organizations, universities, journalists, human-rights groups, and professional associations must be able to investigate and challenge systems.

This direction is no longer merely theoretical. The EU AI Act has established a multi-level governance architecture involving the AI Office, national competent authorities, the AI Board, a Scientific Panel, and an Advisory Forum. Its principal enforcement and governance provisions are now entering their operational phase in 2026. (digital-strategy.ec.europa.eu)

But there is an important point here:

the existence of institutions does not automatically equal democratic control.

The crucial questions are:

Who has access?

Who has resources?

Who can scrutinize whom?

And who ultimately has the power to say “no”?

9. Society Needs the Right to Say “No”

This may be one of the most underestimated democratic rights.

It is not enough to be able to ask:

“Explain why the AI did this.”

Sometimes we must also be able to say:

“We do not accept this technology being used for this purpose.”

Democratic governance therefore requires red lines.

Some applications may need to be prohibited.

Others may be allowed only under strict conditions.

Others may remain open to experimentation.

This logic is already reflected in European and international frameworks. The EU AI Act prohibits specific AI practices considered unacceptable because of their risks, while the Council of Europe Framework Convention allows for measures including bans or moratoria on certain AI applications where necessary. (digital-strategy.ec.europa.eu)

The underlying principle is fundamental:

technological capability does not automatically constitute social legitimacy.

The fact that we can do something does not mean that we should do it.

10. Democratic Governance Must Begin Before Deployment

One of the greatest mistakes would be to wait for harm to occur and only then attempt to repair the system.

Democratic governance must move further upstream:

before design, before development, before deployment.

For every significant AI system, we should ask:

What social problem is it supposed to solve?
Is AI actually necessary?
Who will benefit?
Who might be harmed?
What data are being used?
Who has access to them?
Who has a financial interest?
Who controls the provider?
How is the system audited?
How can a citizen challenge a decision?
Who can shut the system down?
Who pays when harm occurs?

This goes beyond a conventional algorithmic impact assessment.

It suggests something broader:

a democratic assessment of technological impact.

11. And Here We Return to the Ancient Idea of Politics

There is a paradox.

Artificial intelligence is extremely new.

But the fundamental problem it creates is ancient:

How can a society organize power so that no concentration of power becomes unaccountable?

That is the classical problem of politics.

Ancient democracies had no algorithms.

But they understood that power must be visible, contestable, limited, and accountable.

Our contemporary problem is that technological power can become invisible.

It does not always appear as an order.

It appears as:

a “recommendation,”
a “prediction,”
a “score,”
an “optimization,”
an “automation,”
a “neutral calculation.”

And precisely because of this, we must recover the political dimension of technology.

12. The Most Dangerous Point: When Politics Begins to Look Like an Engineering Problem

Imagine that an algorithm tells us:

“This is the most efficient allocation of resources.”

Immediately we should ask:

Efficient for what purpose?

Because efficiency is not politically neutral.

Do we want:

maximum economic output?
equality?
speed?
security?
freedom?
social cohesion?
environmental sustainability?

These values can conflict.

No algorithm can democratically decide for society which value should take precedence.

It can optimize a goal.

It cannot legitimize the goal itself.

Choosing the goal remains a political act.

13. Democracy Therefore Needs Technological Literacy

Not every citizen needs to become an engineer.

But citizens need enough knowledge to understand:

what a model is,
what a dataset is,
what probability means,
what error means,
what an automated decision is,
what profiling means,
what data ownership means,
who has access,
how an appeal mechanism works.

UNESCO explicitly places education, public awareness, digital literacy, AI literacy, and citizen participation within the democratic governance of AI. (unesco.org)

Because a citizen who cannot understand the technology depends upon those who can.

And when that happens, technical expertise becomes political power.

14. The Great Democratic Principle: No “Black Box” Should Acquire Political Sovereignty

This does not mean that every algorithm must be open source.

Nor does it mean that every citizen must have access to the code of every model.

It means something deeper:

No technology that exercises significant power over human beings should exist outside mechanisms of social accountability.

Even when the code is a trade secret, there must be:

independent audits,
technical access for regulators,
rights for affected individuals,
decision records,
appeal mechanisms,
sanctions,
the ability to suspend systems,
and ultimately the ability to prohibit them.

The Council of Europe explicitly places this question within the triangle of human rights, democracy, and the rule of law, and its Framework Convention addresses the entire lifecycle of AI systems. (coe.int)

15. Democracy Itself Must Change What It Means by “Participation”

In traditional democracy we ask:

“Who do I vote for?”

In digital democracy we must add:

“Which systems do I allow to make decisions that affect my life?”

“Which data may be collected and used?”

“Who has access to them?”

“Who can evaluate me algorithmically?”

“Who can exclude me?”

“Who controls those who evaluate me?”

This means that democracy in the twenty-first century cannot be only representative democracy.

It must also become democracy of infrastructure.

16. A Possible New Social Contract

If we wanted to condense this entire argument into a new social contract for AI, we could formulate seven principles:

1. No consequential decision without an accountable human institution.

The machine cannot become the scapegoat.

2. No consequential automation without a right to challenge.

Citizens must have meaningful avenues of appeal.

3. No critical technology without independent oversight.

Self-regulation alone is insufficient where social stakes are high.

4. No technological power without social participation.

Those who bear the consequences must have a voice in the rules.

5. No compulsory technological interaction without meaningful alternatives where fundamental rights are at stake.

A person should not be excluded because they cannot or do not wish to interact with an automated system.

6. No critical infrastructure without public capacity to understand and govern it.

The state and society must possess sufficient technical capacity not to become entirely dependent upon the providers themselves.

7. No technology above democratic decision-making.

Society must retain the right to say:

“We will not do this.”

17. So the Real Question Is Not “Who Controls AI?”

The original question appears to ask for a controller.

A minister.

A committee.

An independent authority.

A corporation.

A state.

But that is the trap.

We should not search for a sovereign actor who will control AI.

We should build a system in which no single actor can control technological power alone.

That is much closer to the logic of democracy.

The answer is not:

“Who will control AI?”

It is:

“How do we design institutions so that the power generated by AI remains distributed, visible, contestable, and accountable to the society over which it operates?”

And here perhaps lies the great political question of our time.

It is not whether machines will become more intelligent than us.

It is whether democratic institutions will become intelligent enough not to allow technological power to become more powerful than democracy itself.

Epilogue: From the “Ethics of AI” to the “Politics of AI”

For years, much of the discussion focused on AI ethics:

AI should be fair.

AI should not discriminate.

AI should protect privacy.

AI should be safe.

All of these are necessary.

But they are not enough.

Because behind every ethical rule lies an institutional question:

Who enforces it?

And behind every institutional question lies a political question:

Who has the power to decide?

And behind that lies the deepest democratic question:

Can a society democratically govern the technologies that are reshaping the very structure of social life?

This is the point at which the discussion of artificial intelligence ceases to be merely a discussion about computers.

It becomes a discussion about democracy.

And perhaps, ultimately, about the very meaning of politics:

not the elimination of power, but the organization of power so that it can always remain subject to those over whom it is exercised.