In mid-May 2026, one day apart, two telling things happened. On 14 May, in Aachen, Mario Draghi returned to the theme of Europe’s competitive decline and the continent’s technological lag, stressing the weight of artificial intelligence on productivity growth. The next day Pope Leo XIV signed an encyclical, Magnifica Humanitas, devoted to the implications of artificial intelligence for people’s lives, made public a few days later. Two worlds that could hardly be further apart, a former central banker speaking of productivity and investment and a document speaking of the human person and society, converge on a single point.
Digital transformation is no longer a matter for insiders. It has become a question of power, of work, of democracy.
When a speech on competitiveness and a papal encyclical take it up one day apart, it means the stakes are now plain for everyone to see.
I want to try to explain why, starting from what discussions of AI usually take for granted or deliberately leave in the dark: what does an artificial intelligence system actually do? Because it is precisely there, in how the technology materially works, that we grasp why artificial intelligence is not neutral, and why deciding how to use it is a political choice and not a technical one.
What artificial intelligence actually does
Strip away the evocative language and an artificial intelligence system does something fairly precise. It turns pieces of reality, a text, an image, a sound, a behaviour, into numbers. Not single numbers, but long sequences of them, vectors, which the machine calls representations or, in technical jargon, embeddings. In this space made of numbers, similar words, similar images, similar situations end up close together, and things far apart in meaning end up far apart. Once reality has become geometry, the system can search for regularities, similarities and patterns that escape us because they are too large or too tangled.
The delicate part is how the system learns to build this geometry. During the learning phase, “training”, the model is shown enormous quantities of data, and at each step it compares the answer it gives with the expected one, measures the error, and adjusts its internal parameters by a tiny amount so as to be slightly less wrong the next time. Repeated billions of times over billions of examples, this process of successive approximations is what allows the machine to tackle complex, non-linear problems, problems with no closed formula, where you have to approach the solution one step at a time. It is in this phase, training, that the model takes shape.
When the model has been trained and we use it to obtain an answer, the so-called inference phase, it is no longer correcting itself by trial and error. It applies in a single pass what it has already learned. The distinction matters, because we often imagine artificial intelligence as something that keeps reasoning and improving while we question it, whereas the learning, with all its enormous cost, happened earlier, elsewhere, on infrastructure we never see.
And here is the point that interests me. All of this, turning reality into numbers, learning by successive approximations, answering, requires three very material things.
It requires the data, that is, reality already turned into numbers.
It requires computing power, that is, the processing capacity to make billions of corrections.
And it requires that someone decide which problem the system must solve, that is, toward which objective it must be optimised.
None of these three things is neutral.
Why neutrality is an illusion
Some still maintain that technology is neutral, that everything depends on the use we make of it. It is a comforting idea, and a false one. No technology is completely neutral, because none comes about by chance: it is designed for a purpose, and from the outset it takes the shape that purpose requires. A hammer too is meant for striking, its form carries inscribed the function it was built for. The difference is that with artificial intelligence this truth, which holds for every technology, becomes vastly more powerful and harder to see, for reasons that are not moral but material, reasons that follow from how the machine is built. An artificial intelligence system is not merely designed for an end, it incorporates choices about data, objectives and constraints, and these choices make some outcomes far more likely than others. It does not impose a single outcome, but it orients, because it is shaped from training onward toward what someone decided to build it for.
The data a system is trained on are not reality, they are a slice of reality chosen by someone, a slice that contains certain things and excludes others, that carries with it the inequalities and absences of the world it was drawn from. The computing power needed to train the most capable models is concentrated in very few firms and very few physical places on the planet, a concentration that decides who can build these systems and who can only use them, on terms set and defined by others. And defining the objective, which problem the system must solve and what counts as a “good” answer, is not a technical choice, it is a political choice disguised as an engineering specification.
Put the three things together. The very same system, the same mathematics, the same vectors, can produce opposite outcomes depending on who owns the data, who controls the computing power, who decides the objective. The neutrality of artificial intelligence dissolves the moment you understand how it is made. There is no artificial intelligence in the abstract, there is always a system built by someone, fed by someone’s data, oriented toward someone’s ends. It is whoever governs this material base who decides, in the last instance, what the technology does to people.
I say this also because, from entirely different starting points, voices very far apart seem to be converging on it today. The papal encyclical of last May sets out two claims that are hard to dispute on the facts: the first is that technology is not neutral, because it takes the shape of those who conceive it, finance it, regulate it and use it, and the second is that digital goods, from algorithms to platforms to infrastructure, have a purpose that concerns everyone and cannot remain concentrated in a few hands. That considerations of this scope should come from such a document is a sign that the problem has left the circle of specialists.
The example of work
The clearest example of all this is work, because that is where the intrinsically non-neutral nature of artificial intelligence touches people’s material lives.
Take a system that organises work in a warehouse, or a delivery service, or a call centre. Mathematically it is a single kind of object, a model that optimises a function. That function, however, can be written in two opposite ways. It can be written to squeeze the maximum output from every minute, to time every movement, to flag every break, to push the pace of work to the limit of human endurance, and in that case artificial intelligence becomes the most tireless and most pervasive overseer that has ever existed. Or it can be written to reduce strain, to distribute workloads better, to schedule shifts that respect workers’ rest and health, to anticipate peaks so as not to dump them onto people, and in that case the same technology lightens work instead of making it oppressive.
A company’s reality, it must be said, is almost never so clear-cut. In most cases the two objectives coexist, time is optimised within certain safety constraints, efficiency is sought without crossing certain thresholds. Yet precisely because the outcome of the management process falls in a grey zone rather than in a single clear choice, what matters is who draws those boundaries and who can verify them. The difference between the two opposite outcomes imagined above does not lie in the algorithm, which is identical. It lies in the objective function, in who decides it, and in what democratic constraints govern its application. If the objective is decided solely by whoever owns the system, and those who work have no voice either in its design or in its governance, the outcome will tend to shift toward surveillance and intensification, because that is what suits the owner of the system. If instead there is a public, democratic power able to set constraints and to give a voice to those who are subject to the system, the emancipatory outcome becomes possible. Not automatic, but certainly possible. This is exactly why rules such as those in Europe’s AI Act or Digital Services Act were introduced, aimed at imposing transparency on automated systems: to let regulators, unions and workers inspect that objective function, to see what the system really optimises and on what terms. Without that chance to “look inside”, the grey zone is drawn unilaterally by whoever owns the system.
This is the heart of the matter, and it is why talking about the ethics of artificial intelligence without talking about the ownership and governance of infrastructure achieves little. Ethical codes, declarations of principle, voluntary guidelines all break against the same material wall: whoever owns the infrastructure and sets its objectives decides the outcome, and no abstract principle is enough against that power if it stays private and beyond any democratic counterweight.
Access to knowledge
There is a second example, less dramatic than work but just as important in the long run, and that is access to knowledge.
Artificial intelligence is fast becoming one of the main tools through which people today look for information, learn, train themselves, make sense of the world. Here too the same logic holds. The same kind of system can hugely widen access to knowledge, putting within anyone’s reach a tool that explains, translates, accompanies learning, tears down barriers that used to be economic or linguistic or geographic. Or it can narrow that access, turn it into a paid service calibrated to ability to pay, steer it toward what is profitable to show rather than toward what is useful to know, shape what people come to know in line with interests that are not their own.
Once again, which of the two outcomes comes about does not depend on the technology, it depends on who, by owning it, shapes it and to what end.
And it is perhaps through this second example that one can better grasp the importance of developing a digital public infrastructure for this technology as well.
A digital public infrastructure is a common, democratically governed base on which both the public sector and the market build their services: it does not replace the market, it gives it a shared track to run on. This is not theory, India does it with UPI payments, Brazil with PIX, Europe with the EUDI Wallet for digital identity.
An artificial intelligence conceived as a public tool for access to knowledge and learning, democratically governed and with transparent objectives, is one thing. An artificial intelligence left entirely to the market and its owners is another. The difference is made by control, not by the machine.
What is at stake
Hold all this together and it becomes clear why a former banker’s speech and a papal encyclical, competitiveness and the human person, end up speaking about the same thing.
Europe’s technological dependence is not only a problem of industrial competitiveness. It is a problem of democratic sovereignty. The infrastructure that today carries information, knowledge, payments, identity, and now artificial intelligence, is overwhelmingly in private and mostly non-European hands. This means that the decisions that matter, which data to collect, how to train the models, toward which objectives to optimise them, on what terms to make them accessible, are taken elsewhere, by private actors who answer to their own interests and not to a democratic mandate, and over whom European institutions have minimal hold.
Whoever controls this infrastructure controls, in effect, the democratic quality of our institutions, because they control the material conditions in which people work, learn, and exercise their rights. This is why digital transformation is a political question and not a technical one, and why governing it democratically is not one option among others, it is the condition for democracy to keep deciding on the things that matter
The objection worth taking seriously
There is an objection to all this that deserves an answer, because it is the strongest and you will hear it often. It is the argument of those who theorise what is called the technological republic, and it runs like this: it is precisely the concentration of capital and expertise in a few large private firms, tightly bound to the national security apparatus, that has allowed the West to build the most advanced artificial intelligence and to keep an edge over China. Breaking up that concentration, subjecting it to democratic constraints, would slow innovation and make us lose the race. Better to let those who run fast get on with it, and trust that they do so for our good.
It is an objection to be taken seriously, because on one point it is true: concentration has accelerated development. Beneath the efficiency argument, however, lies another, unspoken one, and it is worth bringing into the light. To say that private concentration is the price of technological advantage is to say that control of the infrastructure on which we will all have to work, learn and decide must stay in the hands of those who own it today, and that this control is not open to democratic negotiation, because questioning it would make us lose. It is an argument that presents itself as technical, a matter of speed and scale, but it is an argument about power: it serves to remove from public discussion precisely the choices that matter most, and to present them as technical necessities that are not put to a vote. The speed of innovation becomes the reason we should not even ask who owns what.
And here the leap does not hold. It takes for granted that speed is the only criterion by which a society measures success in technology. But a technological advantage paid for with dependence on a few private actors beyond any democratic control is not an advantage for those who bear that dependence, it is a transfer of sovereignty dressed up as progress. If staying in the race means accepting that the decisions that matter for our lives are taken by those who own the machines and answer only to their shareholders, then we have already lost the race.
This is the point to start from if we want to develop a line of reasoning about the political choices around these technologies. Developing technological capabilities and governing them democratically are two different things, and they do not necessarily go hand in hand.
On this point, open-weight models are often mentioned, models whose parameters are made public and can be used, studied and adapted without depending entirely on whoever built them, unlike the black boxes of a few giants. It is an interesting dynamic, but one to be viewed with caution, without mistaking it for the solution. Opening the weights means making public the already-trained model, not the infrastructure and the process that produced it. And here is the point: of open-weight models we almost always do not know how they were trained, on what data, with what filters, toward what objectives. Transparency of the weights is transparency of the result, not of the choices that determined it, which remain invisible. Having a model’s parameters tells us nothing about what it was taught and what was hidden from it. Opening the weights democratises use, not production, and it does not even guarantee the transparency it seems to promise.
This is why the direction is not only to have more open models. It is to make the very exercise of training, the choices about data, the definition of objectives, control over the computing infrastructure, a shared and democratically supervised process. That is where the transparency that counts lies, and that is where power lies. In concrete terms it means being able to know, and to set conditions on, what data a model is trained on, which objectives it is optimised toward, and under what independent audits, above all when that model ends up in public services. The most direct instrument for achieving this, as I will argue further on, is public procurement. How exactly to combine these instruments remains open, but it is in that direction that the alternative to concentration must be sought, not in the mere publication of the weights of models trained who knows how.
Sovereignty as a Widely Shared Capacity
There is, however, a risk in the reasoning so far, and it must be named. Focusing only on who controls and on what constraints to impose can slide into a defensive posture, all suspicion and brakes, that forgets something just as important: artificial intelligence must also be used, and used by everyone. Democratic control is not the end, it is the condition for diffusion to happen in a useful and not predatory way. But diffusion has to be willed, not merely permitted.
Artificial intelligence is a radical change, and it is so on two planes. On one side it is becoming a digital public infrastructure, a critical base underpinning public and private services, as identity, payments and data already do. On the other it is an industrial lever of productivity and competitiveness that touches firms in every sector, not only technological ones. A technology that acts on both planes, if it stays unused or in few hands, makes us lose on both, weaker public services and a less competitive economy. This is why diffusion, and the competence that makes it possible, is decisive.
And here there is a misunderstanding to clear up. We tend to think that counting in artificial intelligence means having the most powerful model, chasing the frontier. For a society as a whole the stakes are different, and more concrete: putting citizens, workers, firms, schools and public administrations in a position to understand what this technology is and to use it well. You do not need to own the largest model in the world to improve diagnostics in a hospital, or to help a teacher follow a student more closely. What is needed is accessible tools, often not at the frontier, widely shared skills to use them, and quality public data for them to work on. A society that learns to use the technology can become at once richer and fairer. The widely shared capacity to use artificial intelligence is itself a form of sovereignty, perhaps the most solid one, because it does not depend on who owns the largest model, but on how deeply a society has learned to make use of AI.
Here too something is beginning to move. With the Apply AI strategy the Commission aims to spread the adoption of artificial intelligence across every sector and among small and medium-sized enterprises, and this is the right direction. But there is a qualification that strategy, focused on industry, leaves at the margins: diffusion cannot stop at companies. Artificial intelligence is a tool that has to be brought into society at every level, among citizens before even among firms, and not only as adoption but as awareness, as the ability to understand what it does, where it errs, who decides its objectives. A population that uses AI without understanding it stays exposed, one that understands it can take part in governing it. It is this twofold movement, adoption and awareness together, that makes the widely shared capacity truly hold both planes, industry and public infrastructure, and become a form of democratic sovereignty and not merely a competitive advantage.
On this, Europe has a problem of attitude even before one of means: the risk is not only that of failing to control the technology, but also that of an excess of caution toward technological innovation that ends up paralysing its introduction while others learn to use it. Concerns about the dangers tied to the use of new technologies are legitimate, and this article does not mean to underestimate them. Yet awareness of the dangers is not a reason to stop, it is the condition for identifying the most suitable tools to protect ourselves.
Digital cooperation as the answer
At this point the question becomes concrete. How does Europe get its hands on that material base, data, computing power, the capacity to build and govern the models, on which democratic control of artificial intelligence depends?
The first part of the answer is internal to Europe, and concerns the capacity to build, not only to use, models of its own, starting from the place where everything is decided, training. Two things are needed. The first is to distinguish more clearly, within the European rules, between what genuinely protects rights and democracy and what instead burdens research without adding any substantial protection: not fewer rules, but rules that are better targeted and proportionate, holding firm to the democratic constraints over objectives, use, and the transparency of systems, while removing the requirements that slow the building of capability without protecting anyone. The second is scale. No member state, on its own, has the resources to build a frontier model, and the European market, today equally behind, will not get there by itself. What is needed is a joint public initiative, mission-oriented.
It is the model of what Europe has already shown it can do when it concentrated its efforts, from Airbus to Galileo: pooling the resources, expertise, and computing capacity of several countries around a common goal.
But giving itself a capability of its own is not enough on its own, and this is where the second part of the answer comes in, the international one. Only a Europe that has given itself its own capability can cooperate with other democracies as a peer, and not as a dependent.
The answer I have already tried to make the case for (I wrote about it at greater length here), and which remains to me the most realistic, is that Europe cannot manage alone and must not try to replicate the global tech giants at home. Attempting an isolated self-sufficiency, a kind of digital autarky, is unrealistic in scale, in investment, in timing. And pursuing a "digital sovereignty" understood as isolation would be a wrong answer to a right problem. The workable path is what I have called digital cooperation, the building of operational and non-subordinate alliances with partners who share the same democratic ambitions, through coordinated investment, interoperable infrastructure, open standards, and shared research. It does not give up autonomy, it is the concrete way to build it, through a governed interdependence rather than one merely suffered.
What the argument about artificial intelligence adds to this idea is that digital cooperation does not serve industrial competitiveness alone, it serves democracy. Pooling among democracies the capacity to build and govern the material infrastructure of artificial intelligence is the condition for that artificial intelligence to be steered toward democratic ends instead of suffered as a fact decided elsewhere.
And it must be said precisely, because everything turns on this, that digital cooperation works on two complementary planes that should not be confused. There is an industrial plane, concerning Europe's capacity to compete and to reduce dependence in strategic technologies, computing, semiconductors, models, built through scale, investment, alliances, and technology transfer among partners. And there is a plane of public digital infrastructures, concerning the capacity of institutions to build and democratically govern identity, payments, data, services, and artificial intelligence as a public tool for access to knowledge, under democratic control. The two planes hold together. Tending only to industry would strengthen competitiveness but leave citizens and administrations dependent on others' platforms. Tending only to public infrastructures would guarantee rights on paper, because without an autonomous industrial fabric those infrastructures would remain technically dependent on technologies we do not control. Real digital sovereignty, at the European scale, means holding the two planes together under democratic government.
Where all this leads
I began with two events of a single day in May, a speech and an encyclical, because together they signal that something has changed. For years digital transformation was told as a technical fact, to be delegated to specialists and to markets. That time is over. Artificial intelligence, precisely because of how it works materially, makes plain that behind every system there is someone's data, someone's computing power, objectives decided by someone, and that those "someones" decide how the technology bears on everyone's work and knowledge.
To recognise all this is not a point of arrival, it is a point of departure. From it follows a precise political choice: to govern democratically the infrastructures on which artificial intelligence depends, to build a cooperation among democracies capable of sustaining that material base, and to spread through society the skills to use it well. To understand how artificial intelligence works is to understand that the digital future will depend not only on the technology we manage to build, but on who controls its infrastructures, on who knows how to use it, and on which rules will guide its objectives. The question is not technological. It is democratic.
Digital Cooperation as the Answer
At this point the question becomes concrete. How does Europe get its hands on that material base, data, computing power, the capacity to build and govern the models, on which democratic control of artificial intelligence depends?
The first part of the answer is internal to Europe, and concerns the capacity to build, not only to use, its own models, starting from the place where everything is decided, training. Having one’s own capacity does not mean full self-sufficiency, but the capacity to weigh on the choices that matter. Two things are needed. The first concerns the rules, and it is not deregulation: it is about simplifying the procedural requirements where they add no protection, while strengthening the substantive obligations, transparency, accountability, access to relevant data, auditability of effects and public purposes. Not fewer rules, but smarter rules, that protect more where it counts and weigh less where they protect no one. The second is scale. No single member state, on its own, has the resources to build a frontier model, and the European market, today equally behind, will not get there by itself. What is needed is a joint public initiative, mission-oriented.
It is the model of what Europe has already managed to do when it concentrated its efforts, from Airbus to Galileo: pooling the resources, skills and computing capacity of several countries around a common goal.
And something is moving. In early June 2026 the Commission presented a technological sovereignty package that, together with the AI Continent plan and the Apply AI strategy of the year before, traces a more coherent line: treating semiconductors, computing power, cloud and artificial intelligence as critical infrastructure, building European capacity instead of merely regulating, and pushing adoption across every sector and among small and medium-sized enterprises, not only large companies. It is a sign that the diagnosis I started from is no longer a minority position.
Two things in that package matter more than the rest. The first is that Europe is beginning to act as a buyer, not only as a regulator: it uses public demand, the buy European approach and the public code principle to steer the market toward European suppliers and standards. Here, though, a sharp distinction is needed, because this is the point at which public procurement can fail. Buying European cannot be reduced to a simple national or continental preference, which would end up protecting weak players merely because they are ours. It has to be procurement conditional on results: openness, interoperability, auditability, security, code reuse, the reduction of dependencies. The criterion is not merely that a supplier be European, but that it be European, open, accountable and competitive. Understood this way it is not protectionism but industrial policy, and it is the lever that brings new and capable players into being rather than keeping alive those that already exist. This lever matters in another way too, the most important one for the argument I am making here: public procurement must not limit itself to buying ready-made tools, it can set conditions on the models it funds or buys, the datasets they are trained on, the objectives they are optimised toward, the transparency requirements, the independent audits, the purposes of use. This is how training, the place where everything is decided, truly becomes a democratically supervised process, rather than a black box we merely purchase. The second is that this framework is explicitly open to like-minded partners, the natural space for this cooperation. There remains, however, a limit of perspective, and this is where my argument parts ways: the package speaks above all of competitiveness and autonomy, whereas the real stake is who governs the process. Public procurement and open code are not only industrial tools, they are the concrete way of bringing the production of artificial intelligence under democratic control.
Giving oneself one’s own capacity, on its own, is nevertheless not enough, and this is where the second part of the answer comes in, the international one. Only a Europe that has given itself its own capacity can cooperate with like-minded partners as a peer, and not as a dependent. The answer I have already tried to argue in the past (I have set it out in more detail here), and which remains to me the most realistic, is that Europe cannot make it alone and must not try to replicate the global tech giants “at home”. Attempting an isolated self-sufficiency, a kind of digital autarky, is unrealistic in scale, in investment, in timelines. And pursuing a “digital sovereignty” understood as isolation would be a wrong answer to a right problem. The path I consider workable is the one I have called “digital cooperation”, that is, building operational and non-subordinate alliances with partners, among states outside the EU, that share the same democratic ambitions, through coordinated investment, interoperable infrastructure, open standards and shared research. A path that does not imply giving up autonomy, but that on the contrary could be the concrete way of building it, through a governed interdependence rather than one suffered.
What the reasoning about artificial intelligence adds to this idea is that digital cooperation does not serve industrial competitiveness alone, it serves democracy. Pooling among like-minded partners the capacity to build and govern the material infrastructure of artificial intelligence is the condition for that artificial intelligence to be steered toward democratic ends rather than suffered as a fact of life decided elsewhere.
This digital cooperation should work on two complementary planes that must not be confused. There is an industrial plane, concerning Europe’s capacity to compete and to reduce dependence in strategic technologies, computing power, semiconductors, models, built through scale, investment, alliances and technology transfer among partners. And there is a plane of digital public infrastructure, concerning the capacity of institutions to build and democratically govern identity, payments, data, services, and artificial intelligence as a public tool for access to knowledge, under democratic control. The two planes should interpenetrate. Tending only to industry would strengthen competitiveness but leave citizens and administrations dependent on others’ platforms. Tending only to public infrastructure would guarantee rights on paper, because without an autonomous industrial fabric that infrastructure would remain technically dependent on technologies we do not control.
That this is not just an abstract idea is shown by the countries that have already begun to move. In June 2026 Canada presented a national strategy that treats artificial intelligence as critical infrastructure, on a par with energy and defence. It aims to reduce dependence on the US giants by building its own data centres, computing capacity and skills, but it states that it cannot do so alone and points to a group of democracies to cooperate with, the European Union among them. It is exactly the logic of the two planes and of cooperation among peers. There remains a difference worth noting. When the Canadian prime minister sums up the stakes by asking whether AI will improve the lives of everyone or only of a few, it seems to me that he is speaking above all of the distribution of benefits, more than of control over production. It is a step forward, but the question I am raising here is further upstream: who owns and governs the infrastructure, not only how its fruits are shared out.
Controlling production, here, does not mean owning every chip and every data centre: that would be a race Europe cannot win on its own. It means democratically governing the process by which these systems are produced: the choices about data, the objectives they are trained toward, the rules under which they are put to work, and pooling among like-minded partners the capacity needed. This is what distinguishes governed interdependence from dependence suffered.
Real digital sovereignty, on a European scale, means holding the two planes together under democratic government.
Where This Leads
I began with two events from a single day in May, a former banker’s speech and a papal encyclical, because together they signal that something has changed. For years digital transformation was told as a technical matter, to be delegated to specialists and to markets. That time is over. Artificial intelligence, precisely because of how it works materially, makes it plain that behind every system there are someone’s data, someone’s computing power, objectives decided by someone, and that those “someones” decide how the technology bears on everyone’s work and knowledge.
Recognising all this is not a point of arrival, it is a point of departure, from which a precise political choice follows, resting on three axes: building Europe’s own capacity, industrial and in public infrastructure, instead of merely regulating; spreading it throughout society, as adoption and as awareness, and not only among firms; building a cooperation among like-minded partners able to sustain that material base, because no one makes it alone. European capacity, social diffusion, cooperation among peers: holding these three axes together under democratic government is what distinguishes digital sovereignty from autarky and from dependence suffered. To understand how artificial intelligence works is to understand that the digital future will not depend only on the technology we are able to build, but on who will control its infrastructure, on who will know how to use it, and on what rules will guide its objectives. The question is not technological. It is political.
The information and views expressed are those of the author and do not necessarily reflect the official position of European institutions.