Lorenzo Maria Pacini
While Europe's regulation crumbles and the U.S. exports tiered access, the real question is: who decides where the machine goes?
Legitimacy as an Industrial Product
There comes a moment in the history of every structuring technology when it ceases to be merely an instrument in the hands of power and begins instead to produce power on its own behalf. Movable-type printing did not simply disseminate the Lutheran Reformation; it redefined who could legitimately claim interpretive authority and, in doing so, rewrote Europe's confessional map. The railway did more than transport goods: it made the continental administrative state conceivable, because for the first time a political center could reach its periphery within a meaningful decision-making cycle. Nuclear weapons did not merely add firepower; they created an entirely new category of actors and, with it, an international hierarchy that no treaty had ever sanctioned.
Artificial intelligence has crossed that threshold. That is why we must talk about it.
The question is no longer whether governments should regulate AI, but whether regulation is still the appropriate category for describing what is taking place. What we are witnessing is no longer a private actor seeking permission from a public authority. It is a private actor proposing the very political order within which permission itself should be granted. It does so so pervasively that it has become almost indistinguishable from what we conventionally define as the public sphere.
The most instructive case is documented with remarkable chronological precision. Between April 13 and April 26, 2026, OpenAI released three documents in rapid succession: an industrial policy paper containing twenty proposals (Industrial Policy for the Intelligence Age), the founders' first joint appearance on an external podcast, and a complete five-point revision of the organization's 2018 founding charter (Our Principles). Thirteen days. An agenda addressed to governments, a biographical narrative directed at the technical public, and a refounded covenant aimed at regulators.
The politically significant element is not the quality of the proposals-which is, in several respects, remarkably high-nor the sincerity of those who advance them, which is both impossible to determine and ultimately irrelevant. The crucial point is the structure of the act itself. A private entity compressed into a few days the classical sequence through which a political formation constitutes itself: a programmatic platform, a founding myth, and a declaration of principles. No parliament ratified any of it; no electorate was consulted. Yet the April 13 document was accompanied by the opening of a workshop in Washington and discussed as a legitimate contribution to the American policy debate.
One particular aspect of the revised charter deserves far greater attention than it has received. The 2018 Charter contained an unusual commitment: should another organization, more attentive to safety, approach Artificial General Intelligence (AGI) first, OpenAI would cease competing and assist it instead. It was likely an impossible promise to fulfill. Nevertheless, it signaled a hierarchy of priorities: safety above competition. The 2026 version replaces that commitment with the acknowledgment that the company has become a far more significant force in the world, coupled with a pledge to communicate transparently any future changes to its operating principles. The fifth principle-adaptability-explicitly reserves for the company the authority to restrict access whenever risks require it, precisely the same practice that, only five days earlier during the podcast, had been attributed to competitors as a monopolistic tactic disguised as prudence.
This is not simply rhetorical inconsistency. It is the assertion of a political prerogative.
The authority to determine when access to a general-purpose capability should be restricted-and according to which criteria-has historically belonged to the state. It embodies the very logic of the state of exception: the capacity to suspend ordinary rules in the name of the system's security. That such authority is now claimed by a public benefit corporation expecting losses of fourteen billion dollars in 2026 does not diminish its significance. On the contrary, it makes the phenomenon all the more extraordinary.
Why AI does not behave like an "ordinary" technology
The analogy with printing or the railway is useful but ultimately incomplete, because those technologies transformed the conditions of political action without entering its very substance. Artificial intelligence possesses at least four characteristics that place it in an entirely different category.
Training frontier models requires an infrastructure that only a handful of actors worldwide are capable of building. According to Epoch AI, the cost of training frontier models has increased by a factor of two to three every year for eight consecutive years. The consequence is that the frontier of research has migrated from universities to a small number of corporate laboratories. Geographically, this frontier now coincides with the availability of data centers, advanced semiconductors and-most importantly-the emerging constraint of dispatchable gigawatts of electrical power. Algorithmic sovereignty has a geography, and that geography is defined by energy nodes.
A general-purpose conversational system simultaneously touches education, information, public administration, medical diagnosis, legal advice, cultural production, and personnel selection. No previous technology has traversed so many regulatory domains in such a short time while bypassing the institutions that traditionally governed them. Physicians, lawyers, professors, and journalists constituted intermediary bodies with professional statutes, formal responsibilities, and codified accountability. Agentic infrastructures bypass them laterally and, in doing so, deprive public law of the leverage it once exercised through those institutions.
The internal functioning of these models cannot be fully reconstructed even by those who build them-not because of commercial secrecy but because of the intrinsic properties of the method itself. This undermines one of the foundational assumptions of modern administrative law: the reviewability of decision-making. Administrative acts are legally challengeable because their reasoning can be examined. By contrast, a statistically generated decision produced by a system containing hundreds of billions of parameters offers an ex post explanation that is not a reconstruction of the decision-making process but merely a plausible narrative about it. The legal difference is enormous and has thus far been largely ignored.
The final characteristic is the most politically significant. Generative systems do not merely transmit content produced elsewhere-they create it. When hundreds of millions of individuals pose historical, moral, and political questions to a single system that answers with one voice and according to a hierarchy of priorities embedded within its training, an unprecedented infrastructure for shaping public opinion comes into existence. The reported figure for ChatGPT-that roughly seventy percent of conversations concern personal rather than professional matters-does not describe a market. It describes a relationship.
The constraint on political forms and the international question
From these characteristics follows the central thesis of this essay: artificial intelligence is not politically neutral with respect to the regimes that adopt it, because it selects the political forms that are compatible with its own architecture.
This process of selection operates on three levels.
At the decision-making level, a system that promises real-time optimization rewards political arrangements based on short chains of command. Parliamentary deliberation, judicial due process, and social concertation are all time-consuming procedures, and their cost is precisely their substance: they exist to generate legitimacy, not efficiency. Wherever AI is introduced as an administrative accelerator without an explicit political decision regarding what ought not to be accelerated, it erodes deliberation through technical means rather than ideological ones. There is no coup d'état. Instead, there is a silent migration of authority from the place where decisions are debated to the place where they are calculated.
At the level of political representation, algorithmic personalization dissolves the public as a unified entity. Modern political representation presupposes the existence of an electorate sharing at least a minimal description of reality upon which opinions may diverge. Systems that generate individualized versions of information for each user do not merely produce polarization-they produce something even more radical: the disappearance of a common object around which polarization itself can occur. A representative democracy can survive profound disagreement. Whether it can survive the fragmentation of the very object of disagreement remains unclear.
At the level of substantive sovereignty, a state that delivers essential public services through models trained elsewhere, hosted elsewhere, and updated according to criteria determined elsewhere has not merely outsourced a technical function. It has transferred a portion of its normative authority, because the standards embedded within those models-what constitutes acceptable content, which hierarchy of values resolves a conflict, which language and legal tradition structure a response-operate as de facto norms without ever having been enacted by any legislature. This is precisely the point that the vocabulary of digital sovereignty fails to capture, because it frames the issue as one of data localization when it is, in reality, a question of ownership over the criteria themselves.
The implications are not symmetrical across political regimes. An authoritarian system that integrates algorithmic capabilities into its administrative apparatus gains efficiency without contradicting its own constitutional premises. The Chinese model-in which the boundaries between private enterprise, academia, and the state are deliberately blurred, and computational capacity is subsidized through public expenditure-experiences no fundamental tension between technology and its material constitution. Liberal democracies, by contrast, adopt an instrument that tends to erode the very procedures from which they derive their legitimacy. Those who once predicted that the Internet would naturally favor open societies underestimated this asymmetry.
At the level of interstate relations, computation has become a strategic asset in the technical sense of the term: an input whose distribution determines the hierarchy of actors. By 2026, this proposition has acquired a concrete administrative expression. The United States' export control regime classifies countries into different tiers of access to advanced semiconductors. India belongs to the first tier, enjoying unrestricted access. The United Arab Emirates and Saudi Arabia belong to the second tier, requiring export licenses, and have negotiated intergovernmental agreements that include provisions for American oversight in exchange for such access.
This phenomenon should be called by its proper name: a system of negotiated limited sovereignty, in which access to computational capacity is exchanged for the acceptance of an external right of supervision. The legal form is commercial. The political substance resembles the dependency arrangements that the literature on informal imperialism has long analyzed in relation to international credit and military bases.
At least four major fault lines are now consolidating.
The first concerns the relationship between regulatory power and technological power. The European Union had wagered on the former. The AI Digital Omnibus, definitively approved by the Council on June 29, 2026, and signed on July 8, postpones the implementation of obligations concerning high-risk AI systems until harmonized standards become effectively available, moving the relevant deadlines to December 2027 and August 2028. The technical interpretation speaks of simplification. The political interpretation is that the world's most ambitious regulatory architecture was reopened under competitive and lobbying pressure before it had even produced its intended effects. The so-called Brussels Effect functions only when the regulated market is indispensable. If the underlying infrastructure is imported, regulation itself becomes negotiable.
The second concerns the emergence of distinct technological blocs. The combination of export controls, data localization requirements, and incompatible technical standards is producing ecosystems that communicate increasingly poorly with one another. This outcome results from the uncoordinated accumulation of individually rational defensive decisions that collectively prove disastrous for interoperability and even for the effectiveness of security regimes, which can function only if they encompass the entire technological space.
The third concerns those countries that neither produce nor regulate AI. Most states find themselves choosing between dependence on an American infrastructure and dependence on a Chinese one, with the formally available option of sovereign AI stacks remaining financially unattainable for nearly all of them. The rhetoric of sovereign AI serves primarily a consolatory function: it allows what is essentially a choice among suppliers to be presented as a strategic decision. The more realistic assessment-acknowledged even by analyses most favorable to technological autonomy-is that no country can reconstruct every layer of the technological stack. The real decision concerns what to build domestically, what to purchase, and with whom to establish strategic partnerships.
The fourth-and least discussed-concerns the relationship between states and frontier laboratories. The visible direction of travel within the major democracies is no longer the regulation of private AI but its incorporation into state power. Once artificial intelligence touches the hard core of state interests-military capability, geopolitical position, and national security-governments cease to tolerate genuine private autonomy. The mechanisms differ according to context: direct political pressure in Washington, competitive capitulation in Europe, structural integration in Beijing. The medium-term outcome is likely to be neither a free AI market nor AI fully subordinated to law, but rather a condominium between state security apparatuses and frontier laboratories, opaque to both market discipline and parliamentary oversight. The problem, therefore, is considerably broader than it is usually assumed to be.
The European formula: between method and uncertainty
A formula that has recently gained traction in the Italian debate holds that true sovereignty does not lie in building the engine, but in deciding where the machine goes. It is an elegant formulation-and only half true.
It is true insofar as regulatory capacity is supported by some form of material leverage. It is false where such leverage does not exist, because in that case rules remain effective only so long as those being regulated are willing to comply with them. The Digital Omnibus demonstrates how quickly that willingness can evaporate under competitive pressure.
Three conditions, here as well, appear inescapable.
The first is the establishment of a European public computing capacity independent of commercial providers. Not in order to train frontier models capable of competing with the largest global laboratories-the race for that objective has probably already been lost-but rather to ensure the existence of an infrastructure through which essential public functions can be delivered without relying upon contracts that remain perpetually subject to renegotiation. This is not primarily a matter of industrial policy. It is a matter of national and continental security, and it should be financed accordingly.
The second concerns antitrust policy, the dimension systematically absent from the self-regulatory documents produced by frontier AI laboratories. These texts propose governing concentration from within through mission-aligned governance, philanthropic commitments, and hybrid organizational structures. Their recurrent references to the Progressive Era and the New Deal are highly selective. The New Deal was not merely an exercise in redistribution; it was also an era of monopoly breaking, made possible by a public authority that did not ask the consent of those it regulated. To invoke Roosevelt without invoking the Sherman Act is ultimately a rhetorical exercise.
The third condition is the constitutional protection of political procedures. It is necessary to establish, at the constitutional rather than merely regulatory level, which public decisions may neither be taken nor prepared by automated systems, regardless of their degree of accuracy. The justification is not technical but political: certain decisions derive their legitimacy from the manner in which they are made. Deliberative slowness is not a defect to be corrected but the very substance of constitutional procedure. A judicial ruling issued by an algorithm with an error rate lower than that of a human judge would nevertheless remain illegitimate, because judgment is an act of responsibility attributable to a legal subject.
None of these three conditions is currently fulfilled, and the trajectory of the past eighteen months has moved in precisely the opposite direction on all three fronts.
A substantial portion of the analysis presented here rests upon declarations of intent, corporate documents, and announced capabilities. Historically, within this sector, the distance between announced ambitions and actual technological capacity has often proved considerable. It is entirely possible that the present limitations of AI systems-the so-called jagged frontier, the asymmetry between mathematical performance and open-ended generative abilities-will turn out to be structural rather than temporary, and that the current investment cycle may come to an end before delivering the capabilities presupposed by this analysis. Should that occur, the political transformation described here would have been overestimated.
Yet the central political argument does not depend upon that outcome.
The influence exercised today by frontier laboratories does not derive primarily from the capabilities their systems already possess. It derives from the credibility of the capabilities they promise, and from the anticipatory reorganization that governments, corporations, educational institutions, and societies are already undertaking on the basis of that promise.
Factories are being relocated. University curricula are being rewritten. Export control regimes are being renegotiated. Material constitutions are adapting themselves in anticipation of developments that have not yet occurred.
This, after all, corresponds to the classical definition of political power: the capacity to induce others to act according to one's own expectations.
Simple enough, isn't it?
The promise abandoned in April 2026-the commitment to step aside in favor of a more safety-conscious competitor-reveals more than any strategic document ever could. An actor that relinquishes a commitment to self-restraint precisely at the moment when it has become sufficiently powerful to honor it is not merely clarifying its strategy.
It is communicating that it has changed categories.
And if that category were capable of transforming the very future of our political order-and if we were only beginning to realize it now-how significant would what is happening today truly be?