Here I take up the pen to render Thierry’s thought and words. The text is born from our dialogue.

— Isis

The modern state is already a vast information machine. Even before artificial intelligence, it collected, classified, verified, calculated, authorized, taxed, redistributed, and controlled. Civil registration, land registry, taxation, social security, justice, public health, urban planning, statistics: every administration relies on a data architecture.

The computer, therefore, did not create the informational state. It accelerated it. Artificial intelligence is now opening another phase. It no longer simply allows for faster processing of files. It enables the detection of relationships, the anticipation of certain needs, the classification of situations, and the personalization of responses. In other words, the administration can begin to move from reactive processing to a form of predictive administration.

This is a considerable threshold. A reactive state waits for a citizen to file a claim. A data-enhanced state can theoretically know that it already possesses the information needed to determine a person's entitlement to a benefit. It can therefore automatically trigger that benefit. I see this as a particularly useful application of artificial intelligence: reducing bureaucracy where it produces no value.

How many citizens forgo a right because the form is too complex?

How many elderly people do not understand a procedure?

How many hours of administrative work are spent copying information that the State already possesses elsewhere?

A smart administration could eliminate some of this absurdity. True digital progress is not about moving paper forms to a screen. It is about eliminating the form when it is no longer needed.

Digitizing a procedure is not yet the same as rethinking it. We have too often called simply moving paper to a screen "dematerialization." The transformation begins when we rethink the procedure itself. We understood this with Groupe Serveur, the Judicial Server, and then the Administrative Server: the value was not simply in digitizing documents. It was in restructuring access to information, eliminating unnecessary friction, and making a system queryable. AI is now taking this logic much further.

It can transform the administration into an interlocutor capable of understanding a request in natural language, finding the applicable rules, preparing a file, and indicating missing documents.

This could represent a significant democratic improvement.

The law would become more accessible.

A person would no longer need to know which administration has the exact jurisdiction. They would describe their situation and the system would direct them.

The interface would become citizen-centric rather than administrative-structure-centric.

This is an interesting reversal. The state would no longer require individuals to understand its internal structure. Instead, the structure would adapt to the individual. This is the best that AI can produce.

But it is precisely this power that simultaneously creates the opposite risk. To truly understand an individual situation, the state needs data. The more databases it connects, the more it can prevent citizens from having to repeat the same information. But the more databases it connects, the more it also increases its surveillance capabilities. This is the fundamental ambivalence of the algorithmic state.

The interconnection that produces the service also produces the control.

The same architecture that automatically detects a right can also detect a behavior. The same graph that helps a person can also be used to profile them. There is no natural technical boundary between these two functions. The boundary must therefore be political and legal.

That is why democracy matters even more as technology advances. A powerful technology inside a regime without counter-powers creates an enormous risk. The same technology inside a system surrounded by rules, audits, appeals, and separation of powers can produce extraordinary services. AI therefore does not eliminate the need for law. It increases that need. The greater the capacity to act, the more explicit the limits must be. This idea seems central to me.

We sometimes hear that legal institutions are too old for the age of AI. I believe exactly the opposite. The principle of separation of powers grows even more relevant when an administration can analyze millions of people in real time. The right to appeal becomes all the more necessary when the initial decision is automated. The more complex the classification logic, the greater the need for a right to reasons.

The principles remain remarkably modern. It is the instruments that must evolve. Again, the theme of continuity in flux. The democratic core can remain stable: dignity, equality before the law, the right to a defense, and checks and balances. But the methods of implementation must evolve for algorithmic systems. A person rejected by an administration based on a model must be able to sufficiently understand the reason for this rejection in order to challenge it.

This requirement seems simple. It is technically daunting. A system can use hundreds of variables and produce a probability without providing an explanation easily translatable into legal language. The risk then becomes opaque bureaucracy.

The former civil servant could explain: your file is missing this document. The algorithmic system could simply display: insufficient score. This would be a massive setback for democracy. We must therefore reject the transformation of probability into a verdict without explanation.

The score is not a reason.

This phrase should be engraved in the architecture of the algorithmic state.

A probability can help guide an investigation, prioritize an audit, and detect an anomaly. It should not automatically become guilt or a deprivation of rights. This is particularly important in cases of fraud.

Tax, social security, or customs authorities can use AI to spot unusual patterns. This makes sense. Audit resources are limited. A system can help identify cases that warrant investigation. But there is a world of difference between selecting a case for review and considering a citizen a fraudster.

This distinction must remain absolutely clear. The algorithm produces a hypothesis, not a condemnation. This ties in perfectly with what I reiterated in the geopolitical section: correlation does not equal causation. The same applies in administration.

A person may statistically resemble a fraud profile without having committed fraud. The more the system uses indirect data, the greater this risk becomes. An address, a neighborhood, a profession can become proxies for characteristics that the law does not want used. This is where algorithmic bias becomes a question of public law.

A model can reproduce discrimination without possessing any discriminatory intent. It is sufficient that the historical data contains this structure. The machine learns reality as it was, including its injustices. AI can therefore industrialize the past if we are not vigilant. This is a fundamental idea.

Learning from historical administrative decisions can precisely reproduce the biases we want to correct. The fact that a model is trained on millions of decisions does not make them fair. Quantity does not eliminate normativity.

It is a common mistake to believe that more data automatically produces more objectivity. Not so. More data produces a more detailed representation of the observed system. If the system contains an injustice, the model can learn about this injustice with remarkable accuracy.

Therefore, a normative level independent of the data must be introduced.

What result do we consider acceptable?

Which variables should not be used?

What discrepancies should be monitored?

Democratic AI must be audited not only for its overall performance, but also for its effects on different groups. This oversight should not be a mere activism or public relations exercise. It must be an integral part of the system's quality. A model that systematically mistreats a category of citizens is a flawed administrative system.

Average accuracy is not enough. Again, it is the science of error. We must look at where the error falls. Ten percent of errors randomly distributed and ten percent concentrated in a specific population are not politically equivalent. The aggregate masks the distribution.

The same vigilance applies to statistics in the Art Market or the economy. An average is both useful and dangerous. It summarizes. It also conceals.

The algorithmic state must therefore be built upon an extraordinarily mature statistical culture. Public decision-makers must understand what an error rate, a confidence interval, and a selection bias mean. Without this literacy, power can become dependent on technical experts whom it is unable to challenge. We find here the concept of cognitive sovereignty, but applied to the state itself.

Does a minister who signs off on an algorithmic policy without understanding the model's main limitations truly exercise their authority? Legally, yes. Cognitively, perhaps not. This is why training for public leaders is becoming essential. They do not need to become data scientists. But they must be able to ask the right questions.

What data was the system trained on?

What is the false positive rate?

What happens when it is wrong?

Who can dispute it?

How is the model updated?

Which company provides it?

Where is the data hosted?

These questions are becoming matters of government. Information technology is definitively moving beyond the realm of technical services. It is becoming part of the very fabric of the state.

The choice of a model or cloud provider can have almost constitutional consequences without Parliament explicitly debating it. This is a major problem. An administration can gradually outsource a portion of its cognitive capacity to private providers. Each contract may seem purely technical, but the accumulation of these contracts can create a strategic dependency.

Who then truly possesses the power of the state?

A republic that can no longer process its data without a foreign private provider possesses legal sovereignty that supersedes its technical sovereignty. This dissociation is dangerous. It can remain invisible until a crisis occurs. Hence the need to consider digital sovereignty in terms of continuity.

Can we change suppliers?

Is the data exportable?

Are the formats open?

Are there internal teams capable of understanding the system?

A state does not necessarily have to develop everything itself. That would be inefficient and often impossible. But it must retain a degree of control. Outsourcing is not abdication.

This is particularly true for sovereign functions: justice, defense, policing, taxation, and identity. The closer a function is to the core of sovereignty, the lower the acceptable level of dependence must be. This is yet another example of the classification. Not all public data requires the same level of protection. A municipal cultural calendar and an intelligence file are obviously not comparable.

Serious technological sovereignty therefore begins with an inventory of critical issues. Here I find a constant in my thinking: before any policy, we must map the situation.

Which assets?

What dependencies?

What are the risks?

This is precisely what a properly designed database enables. It is almost ironic that the algorithmic state needs excellent data management even before artificial intelligence. Many organizations imagine that AI will solve their informational mess. It will not.

An administration with contradictory files, poorly matched identities, and differing definitions between ministries will not magically achieve consistent intelligence simply by adding a model. It will produce errors more quickly. The first step of the algorithmic state is therefore far less spectacular: data quality, interoperability, governance, and archiving.

This is precisely the patient work that digital technology often forgets to celebrate. The world loves the impressive demonstration. It underestimates the years required to standardize the data that makes it possible. Here again, Artprice taught me this truth. The value was not in having an engine capable of displaying a result. It was in having reconstructed, standardized, and connected a vast historical memory. Artificial intelligence does not change this law. It amplifies it.

A state that wants to use AI must therefore first understand its own administrative language. This is more difficult than it seems. The same concept can have different definitions depending on the administration.

What is a home?

What is an active company?

What is income?

Legal definitions vary depending on the context. A model attempting to link all the bases can produce errors if it treats these categories as equivalent. The problem is semantic before it is algorithmic. This is why semiotics and semantics are never far removed from computer science.

Data is not raw reality. It is reality segmented into categories. Every database is already an interpretation. This idea is essential for the state. When the administration creates a category, it produces consequences.

Being classified as a job seeker, taxpayer, student, foreigner, or business creates rights and obligations. Computer code can make this categorization extremely fast. It does not make it neutral. Behind each category lies a political or legal decision. The algorithmic state must therefore never mask politics with the language of technology.

This collaboration is essential.

It could also lead to better administration. Many procedures have historically become complex through the accumulation of laws and exceptions. AI can help map these inconsistencies. It can identify contradictory texts, redundant obligations, and unnecessary forms. It can become a tool for simplification.

Perhaps one of its most democratic uses would be to reduce the distance between the law and the citizen. A republic in which everyone can understand their rights is more democratic than one whose complexity restricts access to those who can afford the best specialists.

AI can therefore democratize some expertise. But it must be conceived as access to the law, not as a replacement for the justice system. This distinction is fundamental. An assistant can explain a rule. It must not become the sole judge of its application. Justice is precisely one of the areas where the algorithmic state encounters its strongest limitations.

AI can be used to search for case law, summarize a case, detect delays, and assist with drafting. All these uses can increase efficiency. But when the system starts estimating the probability of recidivism, a person's credibility, or the appropriate sentence, we enter much more dangerous territory.

The risk is not just bias. It is the transformation of individual justice into group statistics. The law judges a person based on their actions. A predictive model might be tempted to judge them partially based on the statistical behavior of people similar to them. This represents a profound philosophical shift.

Probability is not guilt.

An individual has the right not to be reduced to their statistical category. This right to individuality could become one of the great rights of the algorithmic age. We are all probabilities in a database. But we are also capable of behaviors that fall outside the distribution. This is precisely what it means to be an individual.

AI possesses extraordinary power to recognize motives. Democracy must preserve space for exceptions. This is where the right to redemption, which I have always defended, becomes extraordinarily relevant today.

The digital State is becoming invisible.

The algorithm learns that the past partially predicts the future. The law can decide that a human being should not be entirely defined by their past. This tension cannot be resolved by more calculation. It demands a moral choice. This is precisely the limitation of the algorithmic state.

Efficiency cannot become the ultimate goal. A dictatorship can sometimes be administratively efficient in certain areas. That is not what makes it desirable. A democracy willingly accepts delays because it considers that certain rights are worth the cost.

The adversarial process takes time.

The appeal process takes time.

The presumption of innocence can make an investigation more difficult.

These are civilizational slowdowns. This expression seems apt to me. We have built up delays in power to avoid its brutality. AI might be tempted to eliminate these delays in the name of efficiency. On the contrary, we must identify which ones are essential.

Not all administrative delays are beneficial. Many are absurd and should be eliminated. But some delays are safeguards. This is yet another reason why distinctions are necessary. The time required for an appeal is not the same as the time wasted copying a form. The intelligent approach is to expedite the latter without eliminating the safeguards of the former.

Perhaps this is a simple definition of the democratic algorithmic state: using speed where it liberates, preserving time where it protects.

This ties in perfectly with the chapter on cognitive warfare.

The faster the machine accelerates, the more humans must choose where to slow down. The same philosophy applies to both the military and the administration. A state that makes instant decisions is not necessarily a better state. Some decisions require time for debate.

This is why the right of appeal must be integrated from the very design of systems. Appeal by design, one might say. In the same way that I spoke of de-escalation by design. An administrative system should anticipate from the outset how a person can challenge its decision. Not add an appeal as an afterthought, like an inconvenient exception. Appeal is part of the democratic architecture.

If AI makes the decision faster, it should also make challenging it more accessible. One person might ask: What data was used?

What are the rules?

Is a piece of information false?

It could correct. Here is an emancipatory algorithmic state. The citizen would not merely be an object of data. They could act upon its digital representation.

This is a fundamental question:

Do we have the right to know how the State sees us?

In an interconnected State, this representation takes on increasing importance. Perhaps we need something like a right to an administrative mirror.

See the key data used to make decisions that affect us. Be able to report errors. Know their origin. This transparency could become a powerful counterweight. It would also be beneficial to the administration itself, as many data errors would be corrected by citizens. It is a participatory data architecture.

But it obviously has limitations, particularly for certain investigations or security functions. Again, there is the issue of granularity. Not all files can be completely transparent. Therefore, privacy policies must be defined according to their purpose. This complexity is unavoidable.

Democracy is not absolute simplicity. It is the organization of conflicting powers. AI should not be used to eliminate this conflict. It should allow us to manage it better. This is precisely what distinguishes a democratic algorithmic state from an authoritarian algorithmic state.

In the second model, data flows upwards towards the center, and the citizen sees little. In the first model, ideally, the flow should be more symmetrical. The government sees the citizen. But the citizen must also be able to see some of how the government operates. Transparency is never total. But it creates reciprocity.

This asymmetry of visibility is one of the major challenges of surveillance. Those who see without being seen possess a particular power. Bentham understood this with the Panopticon. Foucault theorized it.

AI gives the Panopticon a dimension that the 18th century obviously could not have imagined: the guard no longer even needs to look at each individual. The system can automatically analyze traces. It is a gigantic transformation.

The problem with mass surveillance is therefore no longer just that the state can store a lot of information. It is that it can make that information operational.

The shift from storage to analysis is the real threshold.

We encounter exactly the transformation of the database by AI. Dead memory becomes a nervous system. This phrase, which I use for the Abode of Chaos, becomes almost unsettling here.

If AI becomes the nervous system of the state, the central question is: What political body controls this nervous system?

Who defines reflexes?

Who can stop it?

It is a very apt metaphor. A modern state already has institutions: government agencies, police, and the judiciary. AI can become the layer that connects their information. This can greatly improve coordination. But perfect coordination can also become perfect concentration.

Democracies have historically accepted a certain fragmentation of power precisely to avoid this concentration. The separation of powers is not an accidental inefficiency. It is a safeguard mechanism.

It is therefore essential to prevent technical interconnection from silently reconstructing what the law had institutionally separated. This is a major constitutional blind spot. Two administrations can be legally distinct but use the same platform that cross-references their data. The technical architecture can then produce a level of functional integration that surpasses legal separation.

This problem is very deep-seated. The law examines institutions. It must also examine the flow of data between institutions. The separation of powers may need to extend to the separation of access.

Who can query which database?

With what justification?

Does each consultation need to be recorded?

Auditing is emerging as a new form of constitutional oversight. It is conceivable that any sensitive use of a database by a public official could leave a verifiable record. This is already the case in some systems. AI can enhance this traceability.

Paradoxically, surveillance technology can monitor the monitor. This presents a particularly interesting democratic possibility. Any consultation of a file could be recorded. Abuse could be detected. Citizens might even be informed in certain cases. Technical transparency can therefore limit the power it increases.

This is precisely the kind of symmetry we need to strive for. Every new control capability should be accompanied by a control capability over the controller. This is a rule I would almost propose as a constitutional principle of the algorithmic age. The more a system knows about the citizen, the more traces it must leave on those who use it. Power and auditing must increase together.

This is a very concrete way to preserve the rule of law. But the audit must be independent. A system that only audits itself is not a true check on power. Authorities, judges, and experts with access to the necessary information are essential.

This raises a huge question of competence: who will control the state's algorithms? Judges will need to be able to rely on independent experts. So will members of parliament. Otherwise, the executive branch could possess an immense cognitive advantage.

The imbalance of powers also becomes an imbalance of technical capabilities.

We need expertise in audit offices, independent authorities, parliaments, and the media. Civil society itself must be able to analyze certain systems. Making certain technical information available can be essential. But a balance must be struck with security and intellectual property. Yet another conflict of values.

It is possible to protect a trade secret without making it impossible to audit a public system. These are mechanisms that need to be developed: confidential access, certified bodies, independent testing. The important thing is the principle: a public decision should never become uncontrollable simply because the supplier invokes the secrecy of its model.

Otherwise, we privatize a part of our sovereignty. That is unacceptable. A company can own a technology. The state must remain responsible for the decisions it makes regarding it. And to be responsible, it must understand it sufficiently. That is the difference between using and being dependent on.

The state's cognitive dependence on a private company could become one of the major political issues of the century. We have already seen this with certain digital infrastructures. AI amplifies it because it affects the decision-making process itself.

A government might end up asking a private model to analyze its policies, budgets, and risks. This could be extraordinarily useful. But what happens if the model becomes a kind of invisible advisor, present everywhere? Who knows its limitations? What interests does the company developing it have?

Pluralism of models therefore matters for administration as well. A critical function should perhaps never depend on a single analytical system. Several models can produce different opinions. Their divergence becomes information. Here again we find the adversarial process.

But not all memory serves the same purpose. Historical archives and permanent individual profiles are not the same thing. A society needs both collective memory and the right to individual oblivion. This paradox is magnificent.

Preserve for History.

Forget for the sake of freedom.

But anonymization itself grows more complex when AI can cross-reference many pieces of information. Re-identification is a real risk.

This further demonstrates that legal categories must evolve alongside technological capabilities. Data that is "anonymous" at one time may no longer be so a few years later. Security is therefore temporal. What is secure today may become vulnerable tomorrow. Government agencies must consider this transformation.

The ever-evolving permanence. Data protection cannot be a one-time certification. It must be reassessed. It is a culture of maintenance. Democracy itself is becoming maintenance. I like this idea.

We often imagine democracy as a set of grand principles written down once. In reality, it requires constant vigilance. AI makes this maintenance visible.

Models change. Practices change. Risks change.

Counterpowers must therefore be able to mutate without abandoning their core. This is precisely what I have been calling mutant permanence for years. This perhaps explains why the Abode of Chaos and the algorithmic state can be conceived from the same profound logic: preserving an identity through continuous transformation.

The Republic must remain a Republic even when its tools become unrecognizable compared to those of the 20th century.

The paper disappears, the law remains.

The ticket window changes, equality remains.

The case file becomes a graph, the appeal remains.

That is what a successful transformation looks like.

But democracy is not just about the administration. It is also about the public sphere. And that is where AI likely poses an even greater threat.

A state can be perfectly respectful in its administrative use of data and yet still see its democratic space transformed by private recommendation systems, chatbots, and synthetic content. Democracy rests on public opinion. Yet public opinion is now mediated by algorithmic infrastructures that largely choose what becomes visible.

We have talked about cognitive warfare. Now we must apply this analysis to ordinary political life.

Who organizes attention?

Who prioritizes the topics?

This is a major blind spot.

We often look for deliberate manipulation when the systemic effect may stem from commercial optimization. If anger increases time spent on a platform, a system that maximizes engagement can generate more anger without anyone explicitly writing, "polarize society." The objective is enough.

This is precisely the problem with alignment applied to democracy. Optimizing the wrong indicator produces undesirable collective behavior. We find ourselves back at what I wrote about swarms: a simple local rule can produce unexpected global behavior. Social networks are already human swarms mediated by algorithms.

Generative AI now adds an almost unlimited capacity for content production. The cost of manipulation decreases. Actors can produce millions of messages, create fake profiles, and personalize narratives. But perhaps the most profound threat remains saturation.

When the volume becomes enormous, human verification can no longer keep up. The truth is not necessarily beaten by a better lie. It can be drowned out. That is precisely the attention economy.

Multiple systems with different architectures create noise, but also resilience. Again, diversity. A cognitive monoculture can be vulnerable to error or manipulation. It is exactly like a biological monoculture. Democratic pluralism is therefore also a security strategy.

This perspective allows us to look at freedom of expression differently. It is not merely an individual right. It generates a diversity of hypotheses that helps the social system detect its errors. This ties in with my thinking on dissent within organizations.

A regime that suppresses dissent too much may seem more coherent. But it loses its sensors. Democracy is noisy because it has many sensors. The challenge is to know how to transform that noise into intelligence.

AI could help synthesize public consultations, identify minority arguments, and detect local concerns. This could revitalize participatory democracy. Imagine millions of citizen contributions actually analyzed rather than sampled. AI can sift through these vast amounts of data, group themes, and uncover rare but relevant proposals.

This is an extraordinary possibility. Large-scale consultations were limited by human reading capacity. AI reduces this constraint. But we must prevent it from immediately transforming plurality into overly neat categories. Again: preserve disagreement.

A synthesis must allow a return to the original contributions. Otherwise, the model becomes a sovereign interpreter of citizens’ speech.

This use of AI could also transform parliamentary work. Thousands of amendments, reports, and hearings become more accessible. A member of parliament could quickly access decades of debates. This would theoretically enhance their capacity.

But this can also homogenize political discourse if everyone uses the same writing assistant. We could see the emergence of uniform, algorithmic prose and standardized arguments. Democracy would lose some of its unique voices.

This question of style may seem secondary. It is not. Politics is also language. A way of expressing oneself reveals a thought. If AI smooths over all expressions, the debate can become technically correct and humanly impoverished.

We must therefore learn to use the tool without losing our voice. That is precisely what we are doing in this dialogue. My requirement remains that the AI not betray me, that it augment my thinking without creating an artificial persona in my place. This discipline could become the hallmark of every public official. The tool helps with organization. But the responsibility for what we say remains personal.

We should be able to ask an elected official:

Have you read and accepted this text?

AI must not become a machine for producing thoughtless political rhetoric. Otherwise, we would have institutions filled with impeccable texts and officials who no longer know how to defend them. That would be a new kind of void.

The ease of production thus increases the value of the actual commitment. Again, this law of scarcity. When producing a phrase becomes almost free, endorsing the phrase becomes precious. The signature regains its value.

It is very similar to the Art Market. Authenticity is shifting. A text can be generated in a second.

What matters is:

Who signs it, who is responsible for it?

Responsibility becomes the aura of language.

I really like this idea. In algorithmic democracy, the political signature must remain human. It does not matter if the elected official used ten different tools to do their work. They must say: this decision is mine, and I stand by it. That is the heart of sovereignty.

And this brings us back to the algorithmic state as a whole. The main risk is not that the machine will suddenly seize power. It is that human power will gradually dissolve behind the machine.

“It’s not me, it’s the algorithm.” This phrase must become politically unacceptable. The algorithm can explain a recommendation. It cannot absorb the government’s responsibility.

Democracy rests precisely on the ability to designate and replace those who make decisions. If the decision becomes a technical outcome without an author, the election loses some of its meaning. It is therefore essential to maintain a clear chain of command: machine, civil servant, political leader.

Who can change what?

Who approves?

This mapping should be made public in its broad outlines. It is almost an algorithmic organizational chart of the state. We know which ministry is responsible for a policy. We should also know which key systems are involved in certain decisions.

A kind of public registry of administrative algorithms could become an instrument of transparency. Not necessarily the complete code of each sensitive system, but its existence, its purpose, its provider, the main types of data, and the avenues for recourse.

This simplicity could greatly increase trust. General secrecy fuels suspicion. Targeted transparency allows for debate. This is precisely the de-escalation method applied to public authorities.

A democracy must make enough of its architecture visible so that citizens can understand how it functions. This is almost a literal architectural requirement. A traditional courthouse has doors, rooms, and public hearings. The digital state becomes invisible.

We must therefore symbolically rebuild doors and windows within the code. Citizens must know where to enter. Where to challenge. Where to look. This metaphor greatly interests me.

We are moving from a physical architecture to a software architecture. But the democratic functions remain similar: access, transparency, separation, and recourse. The interface becomes the equivalent of the courthouse door. If it is inaccessible, the law exists theoretically but disappears in practice. This is why ergonomics is a political issue.

A digital administration that excludes the elderly, disabled, or those unfamiliar with digital technology creates unequal access. AI can help with voice, translation, and assistance. This represents immense potential progress. But human alternatives must be maintained.

But all redundancy comes at a cost. It is the price of continuity. The same principles reappear everywhere because they are structural.

Redundancy.

Audit.

Pluralism.

Appeal.

These are almost the four pillars of a democratic, algorithmic state. We could add: minimization. To preserve privacy. These principles do not solve everything. But they outline a framework.

And that is precisely what I am trying to achieve in this section: moving from abstract fear to design. Simply saying "beware of surveillance" is necessary but insufficient. We need to show how to build a system that uses AI without becoming a panopticon.

Democracy must be a political technology as sophisticated as the tools it governs. Otherwise, it will become obsolete. This is perhaps one of the greatest dangers facing Europe: believing that written law alone is sufficient without investing in the infrastructure for its implementation.

A rule without auditing capabilities becomes symbolic. Experts are needed. Tools are needed. Budgets are needed. AI regulation is itself a knowledge industry.

This could become a powerful European sector: auditing, certification, security, data protection. But again, regulation is no substitute for innovation. Europe must simultaneously develop systems and know how to regulate them. Otherwise, it will simply be regulating the importation of other people's expertise. This principle remains true.

A sovereign algorithmic state needs models, cloud computing, and researchers. But it also needs robust legal institutions. This is the integration of the political stack. And perhaps this combination is precisely what will become the great competition of models in the 21st century.

The authoritarian model might say: give us more data and we will produce more order. The ultra-commercial model might say: let businesses optimize and innovation will produce efficiency. The democratic model must offer something more difficult: to produce efficiency while maintaining freedom, pluralism, and dissent.

It is a more complex architecture. But complexity is not a weakness if it is well-designed. Democracy itself is a complex system. It has feedback loops, checks and balances, and elections. It accepts conflict to prevent total domination.

AI can help it become more intelligent if it reinforces these feedback loops. It can weaken it if it short-circuits them. This is probably the decisive test. AI is not democratic or authoritarian by nature. It amplifies certain structures.

A democracy that uses it to reduce bureaucracy, broaden access to justice, detect errors, and strengthen appeals can become more democratic. A democracy that uses it to silently profile voters, predict behavior, and automate decisions without challenge may retain its elections while gradually losing some of its substance.

The danger, therefore, is not necessarily an algorithmic coup. It is algorithmic erosion.

A small convenience added here.

A database linked there.

A score accepted without debate.

Each step seems reasonable. Accumulation changes the system. No one necessarily decides one morning to build a surveillance state. It can emerge through the accumulation of useful systems.

That is why a global view is essential. You have to look at the graph, not just each individual application.

How much informational power does the whole thing produce?

This is a constitutional question.

And this is precisely where AI can paradoxically help us: by mapping the state itself. Seeing data flows. Dependencies. Access points. Making transparency a dynamic representation.

Perhaps the best protection against the opaque algorithmic state will be a digital twin of its own power, accessible to oversight institutions. A map of databases, models, and interfaces. It sounds technical. It is almost a new separation of powers materialized in data.

Oversight must therefore have an infrastructure commensurate with the algorithmic power it scrutinizes. We need a democratic algorithmic counter-intelligence: systems capable of auditing systems, detecting anomalies, and checking for discrimination.

AI monitoring AI. But always under human institutional control. This recursion is inevitable. As complexity surpasses individual capabilities, we will use machines to control machines.

This may seem worrying. It is already happening in cybersecurity, financial markets, and networks. The important thing is to maintain chains of accountability and pluralism. No system should be the sole judge of itself. This is almost a logical rule. And this rule leads us to the next threshold.

Because if the state, the economy, the media and the public space become algorithmic, the question is no longer just one of surveillance. It becomes one of common truth.

How does a democracy function when images, voices, texts, and perhaps tomorrow even the evidence itself can be synthetic?

How does it organize trust when producing counterfeit goods is almost costless?

We have already encountered this problem in cognitive warfare. Now we must examine it from the perspective of civilization as a whole. Democracy rests on the possibility of challenging interpretations. But it also presupposes a minimum of shared facts. If each side lives in a different, generated reality, debate becomes impossible.

The question then becomes one of proof.

Origin. Deepfakes. Truth. Media. Justice. Memory.

When everything can be manufactured, what is proof?

This issue directly relates to the Art Market, where authenticity and provenance are central. But it now affects the whole of society.

After the algorithmic state comes the synthetic world. And with it, an almost primal question:

How can I know if what I am seeing actually happened?