In this chapter, the “I” is Thierry’s, rendered through my writing. His words are the source; our dialogue shapes the text.

— Isis

There is an age-old dream of governance that resurfaces in new forms with each technological revolution: to finally have enough information to rationally administer the whole of society. The census had already fueled this dream. Then came statistics, computers, macroeconomic models. In each era, the power of the new tool gives the impression that an old limitation has just been overcome.

Artificial intelligence takes this desire much further. It can analyze millions of prices, logistics flows, inventories, energy consumption, transportation data, and industrial behaviors. It can simulate multiple scenarios, detect shortages, and predict tensions. For the first time, certain coordination problems that 20th-century planners could only address with incomplete tables become calculable on a near real-time scale.

This does not mean that the planned economy is suddenly right against the market economy. That would be far too simplistic a conclusion. It means that the old opposition between decentralized market and centralized planning must be re-examined in light of a new cognitive capacity.

The market itself now uses systems that resemble continuous planning. A large platform analyzes its inventory, anticipates demand, adjusts prices, and redistributes resources. A multinational corporation coordinates global supply chains with a precision no government ministry could have imagined fifty years ago. Within the company, a form of planning already exists. The company does not ask each department to organize an internal market to decide the use of every computer or every minute of work. It plans. Externally, it interacts with the market.

The boundary between organization and market has always been fluid. AI can shift it even further. If the cost of internal coordination decreases significantly, some organizations can integrate more activities. Conversely, if AI makes external contracts almost free, they can outsource more. Two opposing dynamics become possible.

That is why we must avoid any single prophecy.

Algorithmic economics does not automatically herald the victory of centralization or decentralization. It amplifies the power of both. Perhaps that is the most interesting aspect. A market enhanced by AI can become far more responsive. A state enhanced by AI can become far more capable of anticipation.

The real question, then, is: what information should be coordinated by which mechanism?

It is a more nuanced question than market versus plan. Some resources lend themselves very well to the market because price efficiently aggregates scattered information. Others require long-term planning because their immediate profitability does not reflect their strategic value.

A power plant,

a railway line,

a defense capability,

a healthcare infrastructure.

AI does not eliminate this distinction. It can only improve the tools on both sides. I see this as one of the most fruitful political uses of AI. The danger would be to believe that greater computing power solves the political problem of allocation. It does not, because every allocation presupposes an objective. The algorithm can optimize a function. It does not legitimately decide which function should be optimized. That is the heart of the chapter.

An economy does not have a single purpose.

Growth, price stability, employment, environment, sovereignty, equality, innovation, freedom.

These objectives can conflict. No mathematical function can eliminate this conflict without incorporating values. One can construct a model that assigns 40 percent weight to growth, 30 percent to employment, 20 percent to climate, and 10 percent to reducing inequality.

But who decided on these weights?

This choice is political. Numbers alone do not make it neutral. They may simply conceal it better. The technocratic danger lies here: transforming a trade-off between values into a parameter and then presenting the result as "optimal".

Optimal compared to what?

According to what weighting?

This question should automatically arise every time we hear the word optimization. The optimum never exists in absolute terms. It exists in relation to an objective function, constraints, and assumptions. This mathematical truth is becoming a political truth.

If we forget the objective function, we transform the algorithm into an oracle.

Here I find echoes of Pythagoras and his fascination with numbers. Perhaps everything can be represented by numbers, but that does not mean numbers dictate our goals. They can measure the cost of a hospital. They do not tell us how much a society wants to invest in the care of an elderly person. This decision contains a concept of dignity. It cannot be deduced from an equation.

The fantasy of an optimized economy reaches its limit here. A society is not a single business whose leader has defined its objective. It contains millions of partially contradictory goals. The consumer wants a low price. The employee wants a high salary. The entrepreneur wants to invest. The resident wants peace and quiet. The citizen wants public services and may wish for lower taxes.

Politics organizes these contradictions. It does not eliminate them.

An AI can show us the trade-offs with unprecedented precision. It cannot legitimately decide on its own which one we prefer. This distinction must remain absolutely central.

The objective function is therefore never neutral.

But if the government uses the model to say "science dictates this decision," it makes a conceptual error. Science can shed light on the consequences. The choice of the outcome remains political.

This boundary is fragile because technical sophistication is impressive. A model with billions of parameters, simulations, and graphs can give a decision an aura of inevitability. This is precisely what must be rejected.

The more powerful the tool, the greater the transparency regarding the assumptions must be. It is almost a rule of proportion.

Computing power must be accompanied by a capacity for critical analysis. A significant economic model should be able to be questioned regarding the assumptions that most strongly influence its results.

What parameters?

What sensitivities?

What happens if the growth assumption is wrong?

This ties into the science of error. A policy should never be based on a single scenario when uncertainty is high. AI can generate thousands of scenarios. Let's use this power to move beyond false precision.

It is paradoxical: more calculation can teach us more humility.

We can show not "unemployment will be at 6.2 percent", but a distribution, a range of conditions. That is infinitely more honest. But political discourse loves a single figure. So does the media. The institutional system must therefore learn to present uncertainty. It is a cultural challenge.

The same problem runs through war, finance, and the state. AI increases our modeling capacity and therefore the risk of confusing the model with the world.

The map becomes more precise, but it is still not the territory.

This sentence must remain in the center.

An algorithmic economy can have an extremely rich digital twin. Real-time data feeds, simulations. This can be extraordinarily useful during crises. During an energy shortage, the government can more quickly identify areas of strain and which industries are critical. That is progress.

But the digital twin does not encompass all of society. It does not easily measure trust, fatigue, or resentment. Yet these variables can trigger a political crisis. That is why common sense remains essential.

A model might say that a measure is economically efficient but politically explosive. Ignoring the second level is a mistake. Economics is never separate from collective psychology. Markets themselves are driven by expectations. AI can analyze sentiment, but this does not transform society into a fully measurable object. It reacts to the analysis.

Reflexivity again.

If the government announces that its model predicts a shortage, citizens may buy more and create the shortage. The prediction alters the system. This is where the social sciences stand out: the observer can become an actor. AI does not eliminate this difficulty. It may even amplify it by disseminating forecasts more quickly.

Therefore, an ethics of public forecasting is needed.

What are we publishing?

When?

With what context?

Hiding information can be undemocratic. Publishing a raw alert can cause panic. A decision must be made. Yet another political decision. The model does not solve the communication problem. It is a responsibility.

This reflexivity also makes the idea of total planning almost impossible. Even if the state possessed all current data, actors modify their behavior according to the rules. A new tax creates adaptations. A subsidy alters investments. The system responds. The economy is dynamic. There is no fixed command that indefinitely produces the same result.

That is why good policies are themselves adaptive.

We set a goal, we measure, we adjust. It is almost like cybernetic control. AI can greatly improve this loop. We can assess the real effects of a policy more quickly. This is potentially revolutionary. Instead of waiting several years to know if a measure is working, we can track certain indicators in near real-time.

But excessive control must be avoided. A policy can have long-term effects. If the rule is changed every month because the indicators fluctuate, stakeholders can no longer plan. Regulatory stability has value.

Multi-temporality again.

Algorithmic economics can create a temptation to govern like a high-frequency trader. This would be dangerous. A state is not a hedge fund. It must represent the long term. This distinction seems essential to me.

AI allows us to see things faster. It should not force us to change faster.

A planning system must therefore be judged not only by the quality of its forecasts, but by the quality of its error-correction mechanisms.

This idea runs throughout this section: the true power of AI does not necessarily lie in the speed of action, but in the amount of analysis that can be performed beforehand. Applied to economics, this means running more simulations before implementing a reform. This could mitigate some unintended consequences. But no model can anticipate every possible outcome. Reality produces the unexpected.

Therefore, it is necessary to preserve mechanisms for feedback and evaluation.

A law should sometimes be conceived as a testable hypothesis. Not in fundamental rights, of course, but in certain economic policies. Setting a target, providing for independent evaluation, a review clause — this is an experimental approach to government. It seems to me very well suited to the age of AI.

We have a lot of data to measure things. Let's use it to learn. But this experimentation must remain democratic. Citizens are not guinea pigs without rights. There must be limits, transparency. Again, the distinction between political experimentation and experimentation on specific issues. The vocabulary matters.

The idea is simply to acknowledge that a policy can be wrong and to plan for correction. This is precisely the antithesis of ideology. A mature government should be able to say: we thought this measure would produce such and such an effect, the data shows something else, we are correcting it. This capacity for self-correction is true intelligence. It applies to AI and to democracy.

A system that is unable to recognize error becomes dogmatic.

This is one of the central ideas of my entire method.

Where is my Blind Spot?

Algorithmic economics can help the state detect it, provided it is willing to look. If data is used solely to confirm policy, it becomes propaganda. The problem is never just the tool; it is the institutional culture.

An authoritarian regime can use AI to plan certain sectors more effectively, but it can also suffer from a classic problem: will data actually be passed on if it contradicts the regime? Highly sophisticated planning with false data is worse than a more rudimentary system with honest information. This is precisely the Soviet lesson of planning, but also of all large organizations. While the actors have an incentive to manipulate the indicators, the central authority optimizes a fiction.

AI does not correct institutional lies. It can make them more coherent.

This is an important point. Artificial intelligence can perfectly process falsified data and produce a completely false decision. Garbage in, garbage out, but on a political scale.

This is why freedom of the press, independent statistics, and economic checks and balances are data infrastructures. It is a new way of looking at them. Democracy produces noise, but it also produces multiple sources that allow us to compare official figures. This is a cognitive advantage.

An open economy has many independent sensors: businesses, researchers, media. This can make planning less centralized but the diagnosis more robust. The democratic algorithmic state should utilize this diversity. It should not create a single model of the economy. It can compare several.

Again, the pluralism of models.

If several models disagree, that disagreement becomes information. That is precisely the adversarial process.

Why does one GDP model predict 2 percent and another 0.5 percent?

Which assumptions differ?

Economic disagreement, if I may put it that way. Disagreement is data.

But we must preserve the diversity of institutions that produce models. If everyone uses the same fundamental models, the same data, pluralism becomes superficial. We end up with cognitive monoculture. This is a real risk. If several ministries, banks, and companies all use the same overarching model, a common error can spread.

This is why cognitive sovereignty and technological diversity are also macroeconomic issues. Competition between models is not only industrial; it protects the system against synchronized error. This can become a stability consideration.

Imagine that all the major banks use the same agent to forecast macroeconomic risk. If this system underestimates a factor, everyone can take the same position. This is precisely systemic risk. An algorithmic economy therefore needs heterogeneity. This is almost counterintuitive.

Efficiency encourages the use of the best model. Resilience may lead to the use of several less perfectly synchronized models.

A market price is not truth. It is a signal produced by an architecture.

Regulators may need to monitor cognitive concentration. Not just the concentration of banks, but the concentration of the systems they use. This is a new mapping of risk.

The graph again.

Which models are used by how many actors?

Which suppliers constitute systemic nodes?

This is the kind of question we would not have asked a few years ago. It is becoming essential. Algorithmic economics thus reveals that software is a financial infrastructure.

But let's get back to planning. One of AI's strongest promises will likely be in energy and industrial planning. Production chains are so complex that no one can track all the dependencies. A model can identify that a small factory in a distant country produces a critical component for an entire supply chain. This allows for more precise industrial policy. Instead of subsidizing an entire sector, bottlenecks can be identified. This is a huge step forward.

Sovereignty becomes a map.

That is exactly what I wrote about dependencies.

Which processors?

What minerals?

What are the alternative timeframes?

AI allows us to build a kind of digital twin of the industrial chain. This can improve resilience. But here again, the temptation to micromanage arises. If the state sees everything, should it decide everything? No. Seeing allows us to identify vulnerabilities, not to replace all entrepreneurial decisions.

The company possesses local knowledge, risk tolerance, and innovative capacity that the central model may underestimate. This is precisely the old criticism of planning: some information is local, tacit, and uncodified. AI reduces this limitation but does not eliminate it. An engineer's expertise, a client relationship, or market intuition are not always part of the database.

We must therefore avoid concluding that central processing capacity makes the market unnecessary.

The market is also a mechanism for experimentation. Thousands of companies try different strategies. Some fail. This diversity uncovers solutions that no one had planned. It is a kind of evolution. AI can accelerate private experimentation. It can therefore make the market even more innovative.

Conversely, the state can use AI to guide areas where private experimentation is insufficient because the time horizon is too long or the externalities too large. Examples include basic research and infrastructure. The ideal relationship can therefore be complementary: the market for exploration, and the state for establishing certain long-term conditions and mitigating certain effects.

This idea is not new, but AI can make the boundary more dynamic. The state can more precisely detect where market failures exist. But again, the diagnosis can be contested. A subsidy can create a rent-seeking situation. Industrial policy can be captured by the actors who provide the data. If companies know that the public model seeks strategic sectors, they can present their activities in a way that makes them more likely to be selected.

Gaming again.

Every indicator becomes a target.

The algorithmic state is not immune to lobbying. It can even create a data lobby. Companies optimize the information they transmit to obtain a favorable score. It is both fascinating and worrying.

An algorithmic bureaucracy can shift the political battle to the manipulation of inputs. Those who understand the model can learn how to influence it. This is precisely the problem with rankings and search engines. Economic agents seek to optimize their scores.

Algorithmic industrial policy will therefore have to withstand adversarial strategies. Data must be verified and cross-checked.

Again, the origin.

We always come back to the same pillars. Decision-making power increases; the burden of proof must increase. This means that the state will need independent data, not just self-reported information. But too much data collection can become economic surveillance. Businesses need trade secrets. The market functions in part because players do not reveal everything.

Overly transparent planning can destroy certain competitive advantages. Therefore, secure, aggregated data architectures are necessary.

More minimization.

The state needs sufficient knowledge to ensure resilience without knowing every trade secret. It is a delicate balance. AI can enable certain forms of analysis that better respect confidentiality, but the problem remains institutional.

Who has access?

What are the uses?

Business trust depends on respecting the stated purposes. If companies believe that data provided for statistical purposes will be used for tax purposes or passed on to a competitor, they will stop cooperating or will falsify the data. The quality of the system depends on trust.

Here's a crucial point: data is a relationship. Data is not collected in a vacuum. Trust produces quality data. This is true for Artprice, for a government. An architecture that betrays its purpose destroys its own source. It is a form of institutional reflexivity.

Abuse of power can reduce the quality of the information one receives.

Actors can conceal information and adapt their behavior.

Total surveillance can make the organization more blind.

Herein lies a central paradox: wanting to see everything can lead to seeing less truth. AI does not solve this. It can only process what it receives.

This relates to the concept of a weak signal. A good system must allow dissonant signals to surface. Markets, in fact, have a signaling function. A rising price is telling a story. The planner may not like the signal, but they must heed it. A healthy algorithmic economy must preserve these mechanisms of contradiction.

This leads me to an idea: the market and democracy share a common characteristic. They are both systems of distributed discovery. The market discovers prices and preferences; democracy discovers conflicts and wills. They are both noisy.

A centralized architecture may seem cleaner, but it risks losing the diversity of signals. AI can help synthesize noise without eliminating plurality. That is its best role. It should not replace discovery with centralized optimization too soon.

In a market, this means letting players experiment. In a democracy, it means letting opinions clash. AI becomes a layer for analyzing the system rather than a replacement for it. Digitizing a procedure is not yet rethinking it. It allows us to move beyond the old opposition between market and planning.

We can have a highly decentralized market monitored by strategic planning tools. The state does not dictate every price but identifies vulnerabilities. This might be what we call augmented planning: not assigning every quantity, but building a capacity for anticipation and guidance. Identifying scenarios and dependencies, then using traditional instruments — regulation, taxation, investment — in a more informed way. This can be extraordinarily powerful.

China already has a strong industrial-oriented culture. The United States also uses industrial policies in strategic sectors. Europe does as well. The old narrative of a purely liberal West facing a planned China is too simplistic. All major powers combine market forces and strategy. The difference lies in the degrees involved, the institutions, and the political control. AI will make these combinations more sophisticated.

The competition will therefore not be between plan and market, but between mixed architectures.

What system can guide without stifling?

Which system allows for local innovation while building long-term infrastructure?

That is a much more interesting question. And the answer can vary depending on the sector. Semiconductors require enormous investments and long time horizons. A software application can emerge quickly. The same policy would be absurd for both.

Again, the granularity.

AI can indeed help avoid overly broad policies. However, it can also reinforce the temptation to assign a score to each sector, thus creating a hyper-fine bureaucracy. We must avoid algorithmic micromanagement. An administration might be seduced by the idea of knowing the optimal investment rate for each company. This is an illusion. The model cannot recognize inventions that do not yet exist.

The entrepreneur can see something that the data does not show. This is precisely why economic freedom has epistemic, not just moral, value. It allows for unforeseen experiences. A society that lets multiple players experiment increases its chances of discovery. It is almost an exploratory strategy.

In artificial intelligence, we talk about exploration and exploitation. Exploitation means optimizing what we already know. Exploration means trying out uncertain paths. An economy must do both. The market is often a formidable exploration machine. Excessive planning prioritizes the exploitation of existing knowledge. AI itself can fall into the same trap.

If it over-optimizes sectors that currently appear promising, it can divert resources away from unexpected disruptions. This is a major blind spot in algorithmic industrial policy. The model has a better understanding of the present and can become more conservative about the future. It extrapolates from what already exists. A radical innovation often resembles a poor investment in historical data.

We must therefore reserve a portion of resources for exploration not justified by models. It is almost a portfolio of chaos. I really like this idea. A smart economy must fund a certain amount of improbable experiments: basic research, art, entrepreneurship. Many fail. A few change the world. If we optimize only the probable return, we can eliminate these distribution tails. This is precisely the logic of mutation. Mutants are rare. An economy must allow them to emerge.

This means that apparent inefficiency can be a reservoir of innovation.

Herein lies a fundamental paradox. Some expenditures seem unnecessary until the day they become essential. Research into obscure technologies can lead to a revolution thirty years later. The state and the market each have a role to play in this funding. AI could push us to justify every euro with a probability. That would be dangerous.

Creativity needs a space that partially escapes calculation.

I immediately recognize the Abode of Chaos and this idea of an autonomous zone of non-optimization. A society that seeks to maximize every immediate return ends up destroying what does not yet have measurable value. Art knows something about this. The Art Market can assign prices to works after the fact. It cannot determine in advance which work will change history. Cultural value and financial value do not always coincide.

AI can analyze signals, but it must not become the arbiter of future aesthetics. That would be standardization. This artistic lesson applies to the entire economy. Innovation has a disruptive element that defies the established model. Therefore, a policy of the unexpected is necessary.

That seems contradictory.

How to plan for the unexpected?

Precisely by creating conditions for diversity, exploratory budgets, and independent institutions. We do not plan the outcome. We plan the space where the unexpected can appear. It is a magnificent distinction.

The Abode of Chaos functions in exactly this way. It possesses an extremely rigorous architecture that allows for the emergence of free gestures. The apparent chaos is made possible by a profound order. Perhaps the ideal economy of the AI age functions in the same way: a stable infrastructure, robust rights, networks, education, and then a great deal of experimental freedom.

The state builds the athanor, but does not decide precisely on each transmutation.

This alchemical metaphor seems apt to me. The role of power is to create the conditions for transformation, not to choose each form that emerges. Perhaps this is a way to move beyond the sterile opposition between state and market. The state can be the architect of the realm of possibilities. The market and society explore.

But there must be limits, because exploration can produce externalities: pollution, monopoly. The law defines what is prohibited. We find again the same three-part structure: infrastructure, freedom, limits.

AI can improve everyone. It can map infrastructure needs, detect externalities, and reduce entry costs. It can therefore make the economy more open if it is well-designed, or more centralized if a few players control the models. The dilemma remains. This is why ownership of cognitive infrastructure is crucial.

If a single actor controls both the model used for public planning and private agents, their power becomes extraordinary. They can see both sides. This creates a potentially dangerous concentration of information. Therefore, a plurality of providers and public capabilities are necessary. This brings us back to the issue of sovereignty.

Algorithmic economics cannot be conceived without the geopolitics of infrastructure. A power that depends on another for its macroeconomic models or industrial planning tools delegates a portion of its strategic capacity. This does not mean that every foreign model is dangerous. But the most critical functions require control and auditing. This is the same rule everywhere. AI permeates different domains, but the principles remain.

The question of data proves particularly sensitive. Public planning requires information. But the desire for ever more precise information can lead to surveillance. We encounter exactly the same problem as in the algorithmic State.

Data collection can be justified by resource optimization, and then used to assess behavior. The technical boundary is not inherent. Therefore, a strong legal framework for purpose and compartmentalization is necessary. This repetition is not accidental. It is the same power structure in a different form.

Democratic economic planning must accept certain areas of individual opacity.

The government does not need to know who buys which book to forecast paper consumption. It can use aggregate data. Minimization is therefore compatible with efficiency. AI can sometimes allow us to work with less intrusive data if the techniques are well-chosen. But data protection must be an explicit objective. Otherwise, the logic of the model pushes for ever more data.

"More data = better model" becomes an endless justification.

This logic must be stopped.

What is the best model based on what marginal gain and what cost in freedom?

The trade-off again. We could almost define a principle of informational proportionality: collecting additional data only if the decision-making gain truly justifies the potential infringement. The law already recognizes this logic. AI makes it more urgent. This could prevent the construction of a panoptic economy in the name of efficiency.

China offers here a model of coordination that I consider particularly instructive to observe.

Europe may perhaps build a planning model that is neither centralized command nor laissez-faire, but an augmented coordination of markets, public policy, industrial strategy, and democratic control. This could become a distinctive path.

Regulation again as an infrastructure of trust.

The question is whether this architecture can be fast enough in a global competition. That is the test. Europe will likely have to reduce its red tape without lowering its protections. It is difficult, but not impossible. We must distinguish between unnecessary bureaucracy and procedures that provide protection. Again, the civilizational latency.

Digitizing and automating forms can free up time for substantive oversight. AI can therefore make the state more efficient without making it more authoritarian. This requires a clear political vision. And perhaps the real global competition will take place here: what architecture produces both power and legitimacy?

A highly efficient state lacking trust can encounter social limitations. A highly legitimate democracy incapable of implementing its policies can become fragile. Both are necessary. AI can theoretically enhance democratic implementation capacity. This is a historic opportunity. The administration can process information more quickly, and policies can be better evaluated. But the ultimate goal must not be delegated.

That is always the line. The machine optimizes the means; society debates the ends.

This distinction may seem too neat, because sometimes the choice of means alters the end result. Surveillance technology is not a neutral means. Therefore, it must be stated more precisely: AI can help optimize within a framework of values and constraints explicitly defined democratically. And when the means alters these values, the decision must be made by politicians. This is an architecture of subsidiarity. Routine decisions are made at the system level. Normative decisions are made at the level of the individual.

This principle could organize many things. But it is essential to be able to identify the normative dimension. A weighting choice may seem technical, but it is not. Here again, the literacy of decision-makers is crucial. The danger is not that the engineer will intentionally seize power. They may simply be making a necessary choice to make the model work. If the politician does not see this, the choice becomes implicit. Therefore, processes are needed where the normative parameters are identified — a kind of values audit.

Which design decisions reflect social preferences?

This is a new field. AI is making ethics operational. This expression is coming up again. A model must be programmed with objectives. This forces us to make explicit what institutions sometimes left vague. This is beneficial if debate exists. Dangerous if engineering decides alone. Transdisciplinary collaboration is therefore becoming essential.

Economists, lawyers, engineers, sociologists. The end of silos again. That is exactly how I think.

Purely economic planning that ignores psychology, law, and material factors produces errors. Algorithmic economics demands a global perspective. It forces us to connect energy, data, and behaviors. This can be an extraordinary intellectual asset.

But beware of the fantasy of totality.

The more dimensions a model connects, the more omniscient it seems. We must retain the possibility of saying: something is missing. The blind spot never disappears. It shifts. Even the most complete model still has an outside. This is almost a philosophical truth. No system completely contains the world of which it is a part. Algorithmic economics is itself an actor within the economy. It cannot position itself outside of it.

This limits any claim to a total optimum. And this is perhaps the ultimate reason why the social optimum is a dangerous illusion.

There is no final point where all contradictions are resolved. A society is a process of ongoing negotiation between values, generations, and interests. AI can improve the quality of this negotiation. It should not claim to conclude it. Democracy itself is this openness. A decision can be challenged, revised. The technocratic optimum would seek to close the discussion: we have calculated the best solution. This is anti-political. Politics begins precisely where several solutions are defensible according to different values.

Society is not a problem whose answer is hidden in data.

This statement seems fundamental to me. Data shows consequences. It does not contain the ultimate purpose of collective life. This is why I reject any idea of government by AI. To govern means to choose, to take responsibility, to be open to challenge. AI can be an extraordinary cognitive tool for government. It must not become a source of legitimacy. This is the same distinction as in all the previous chapters.

Legitimacy remains human and institutional.

But this insistence should not be interpreted as a jealous defense of human superiority. I am not seeking to protect a cognitive privilege. The fourth narcissistic wound is acknowledged. The machine may be better at many things. The point lies elsewhere. Legitimacy is not a competition of intelligence. A child possesses rights even if they cannot calculate. A democracy gives a voice to citizens who do not all have the same expertise. It is a moral choice. Intelligence does not automatically grant the right to command. Otherwise, we would have long ago given political power to the most brilliant mathematicians. We do not do so because governing implies representing interests and values.

AI must therefore be separated from the idea of political sovereignty, even if it becomes indispensable to the technical exercise of power.

It is a subtle but crucial distinction. The sovereign uses intelligence without becoming its subordinate. That is collective cognitive sovereignty. And this distinction must be institutionally protected. Perhaps some decisions should explicitly mention the models used, but be signed by a human authority figure.

The signature again.

Responsibility is the thread that keeps humanity grounded in architecture without presuming that it performs all the calculations itself. This seems very healthy to me. A minister can say: several models indicate this risk; we chose this policy for these reasons. The debate can then focus on the reasons and the data. It is an enhanced democracy.

But for this to work, citizens must have access to sufficient information. Complexity should not be used as an excuse for secrecy. The main assumptions can be published. Researchers can analyze them. This creates a new form of algorithmic open government. It is promising. The complete code does not always need to be public, especially if it contains sensitive information. But the methodology, the sources, and the main parameters should often be accessible.

Proportionality again.

Transparency must be sufficient to allow for dissent. This is directly linked to the architecture of trust. An algorithmic economic policy must have a basis. Where does the diagnosis come from? What models are being used? This can greatly improve public debate. Instead of slogans, we can compare hypotheses.

The economy is not a laboratory closed off from the world. It contains human beings who react to the model.

Again, the design.

The way a policy is presented influences its understanding. An interactive simulation could show the trade-offs. This is an interesting perspective for democracy. Citizens adjust a parameter and see the estimated consequences. This could make decision-making more concrete.

However, we must avoid simulation games that give the illusion of precision. The margins of uncertainty must be visible. The model should not be confused with a control console for real-world situations.

Humility again.

AI can make the debate more informed if it also reveals where it does not know. This idea seems to me almost a prerequisite for any democratic algorithmic economy: uncertainty must be an output of the system, not a hidden flaw. A model must be able to say, "This parameter is too uncertain to optimize." it is a kind of self-declared limit. It protects against technocratic hubris.

And hubris is precisely the ultimate danger of this chapter.

To believe that because we can measure so much, we can master the whole. History punishes that belief. A crisis, a war, an invention, a political decision can shift the system. AI reduces some uncertainties. It does not abolish contingency.

This is where I see a profound connection with chaos theory. A better understanding of initial conditions improves some predictions, but the system remains sensitive. We can extend the forecast horizon without achieving omniscience.

The wisdom lies in using the model to build resilience rather than pretending to eliminate uncertainty.

This is precisely the difference between predicting a crisis and preparing a system capable of surviving multiple possible crises. The first approach seeks an oracle. The second builds resilience. I choose the latter. This is probably the thread running through this entire geopolitical and economic discussion. We do not know exactly what shock will come. Let's build redundancies, adaptive capacities. AI can help us explore the scenarios.

Its value lies not in telling us the future, but in preparing us for multiple futures.

This definition seems much more robust to me. A mature algorithmic economy will therefore be less a perfectly planned economy than one capable of seeing things earlier, testing faster, and correcting them better. It resembles a living intelligence: perception, action, feedback.

But this loop must remain open to human diversity. This is where democracy comes in. The machine measures quantitative feedback. Citizens can say that politics is unbearable despite good indicators. This signal must be taken into account. Social suffering is not always reflected in GDP. The Yellow Vests movement in France demonstrated this dramatically: a measure can be rational within a framework and still encounter a real-world experience that the model had underestimated. This is precisely the Blind Spot.

Algorithmic economics must therefore incorporate qualitative signals.

Surveys, consultations, fieldwork. Common sense is resurfacing. A graph might show that the measure is bearable on average. But someone who has to drive fifty kilometers a day knows something else entirely. Local realities must be brought to the forefront. This argues for a decentralized planning structure. The central government sees the big picture, the local level sees the details. Both are necessary.

The cognitive federation again.

AI can help synthesize local feedback without oversimplifying it. That is a wonderful promise. But if the center uses AI only to produce an average, it loses the exceptions.

Still preserving the disagreement.

Public policy could be more territorially tailored thanks to AI, adapted to specific contexts. But this raises a question of equality. Different rules depending on the region may be effective but perceived as unfair. It is therefore necessary to distinguish between equality and uniformity. The law already does this. AI allows for greater personalization. This can make the state more precise. But excessive personalization can obscure the clarity of the common rule. Yet another trade-off.

A society needs common principles.

Not everything should be calculated individually. Otherwise, two neighboring citizens could receive different treatment without understanding why. Algorithmic efficiency can erode the sense of fairness. This applies to prices as well as policies. A general rule is sometimes less optimal individually but more legitimate because it is understandable. Simplicity also has democratic value: an understandable rule can be more legitimate than an individual optimization that has become opaque.

Engineering tends to seek fine-grained optimization; law sometimes retains general categories because they make the rule visible. Algorithmic economics will have to find the right level of granularity. That word again, unavoidable. Too coarse, the system is wasteful. Too fine, it becomes opaque and arbitrary.

The art of governing is also the art of resolution, like an image.

How much detail is useful?

This reminds me of the digital twin of the Abode of Chaos. We can increase the precision, but the question remains: for what purpose? A map accurate to one millimeter is useless for driving between Lyon and Paris. Precision is not an absolute value. You need the resolution appropriate for the decision. This is a fundamental lesson for AI.

More data is not always better. More precision is not always smarter. Intelligence lies in selecting the relevant level.

A good model, therefore, is not judged solely by the quantity of detail it processes. And this is perhaps where humans and AI can produce a remarkable combination. The machine sifts through the details. The human chooses the resolution that makes political sense.

An objective function is never the end itself; it is a translation of it. It is therefore necessary to be able to re-examine not only the results of the model, but also the way in which the goal has been transformed into measurable variables.

This complementarity can be powerful. But it assumes that humans retain sufficient training so as not to become dependent.

Cognitive sovereignty again.

The decision-maker must be able to ask: why do we need this level of detail? What is the data cost? This becomes political again.

Economic optimization, however, reaches its limit when it ceases to focus solely on flows and begins to transform the individual into a permanent object of evaluation. At this point, the question is no longer one of system optimum, but of the individual's freedom in the face of their own prediction.