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
At a certain point in the evolution of every transformative technology, it stops being a mere tool wielded by those in power and starts generating power autonomously. Movable-type printing did more than spread the Lutheran Reformation; it changed who could rightfully claim interpretive authority, effectively redrawing Europe’s confessional landscape. Railways did more than move goods—they made the concept of a continental administrative state feasible, as political centers could for the first time reach their peripheries within workable decision cycles. Nuclear weapons went beyond providing firepower; they introduced a novel class of actors and established an international hierarchy without formal treaty approval.
Artificial intelligence has passed this pivotal point. That is why discussion about it is crucial.
The issue is no longer if governments should regulate AI, but whether regulation is the appropriate framework to capture what is occurring. The dynamic no longer features a private entity seeking approval from a public body. Instead, a private actor is proposing the political framework within which permissions themselves should be granted. Its influence is so pervasive that it nearly merges with what is traditionally considered the public sphere.
The most instructive example is documented with exceptional chronological detail. Between April 13 and April 26, 2026, OpenAI unveiled three major documents in rapid succession: a policy blueprint featuring twenty recommendations (Industrial Policy for the Intelligence Age), the founders’ first joint appearance on an external podcast, and a comprehensive five-point overhaul of the original 2018 charter (Our Principles). Thirteen days. This sequence delivered a government-focused agenda, a narrative targeted at technical audiences, and a renewed covenant aimed at regulators.
The political significance lies not so much in the proposals’ quality—which is quite high in several respects—nor in judging the sincerity of their authors, which is unknowable and irrelevant. What matters is the structure of the act itself. In just a few days, a private organization compressed the traditional phases of political formation: setting a platform, crafting a founding myth, and issuing a declaration of principles. None of this had parliamentary approval or electoral legitimacy. Yet the April 13 document was accompanied by a workshop in Washington, recognized as a valid input to the U.S. policy debate.
An element of the revised charter deserves much closer scrutiny. The 2018 Charter made a unique pledge: if another group more focused on safety arrived first at Artificial General Intelligence (AGI), OpenAI would halt competition and assist them instead. That promise was likely unachievable, but it established a hierarchy—safety trumped competition. The 2026 version replaces this with an acknowledgment of the company’s increased influence worldwide and commits to transparent communication about any future adjustments to its operating principles. The final principle—adaptability—explicitly grants the company the right to limit access when necessary due to risk, echoing a practice recently criticized as a monopolistic maneuver disguised as caution during a podcast discussion just days before.
This represents more than mere contradictions in rhetoric. It asserts a political prerogative.
The power to decide when and how to restrict access to a general-purpose capability traditionally belonged to the state, embodying the principle of a state of exception—the capacity to suspend ordinary rules to protect the system’s security. That this authority is now claimed by a public benefit corporation expecting $14 billion in losses in 2026 only highlights the exceptional nature of this development.
Why AI does not behave like an “ordinary” technology
While analogies with printing or railways provide some insight, they fall short because those technologies altered the context of political action without penetrating its core. AI stands apart due to at least four defining traits.
Training frontier models demands infrastructure that only a limited number of actors worldwide can create. Epoch AI reports that training costs have doubled or tripled annually for eight consecutive years. Consequently, the cutting edge of research has shifted from academia to a few corporate labs. Geographically, this frontier depends on data centers, cutting-edge semiconductors, and crucially, access to dispatchable gigawatts of power. Algorithmic sovereignty has a spatial dimension defined by energy hubs.
AI-based conversational systems intersect simultaneously with education, public information, government administration, healthcare diagnostics, legal advice, cultural output, and recruitment. No previous technology has so swiftly crossed numerous regulatory fields while circumventing traditional governing institutions. Professionals like doctors, lawyers, teachers, and journalists acted as intermediaries with defined roles and accountability. Agentic AI undermines these mediating bodies laterally, thereby weakening public law’s formerly effective leverage through them.
These systems’ internal workings cannot even be fully reconstructed by their creators—not due to trade secrets but because of the intrinsic nature of the technology. This challenges a basic tenet of modern administrative law: that decision-making can be reviewed and scrutinized. While administrative decisions are contestable because their rationale can be examined, statistically generated outputs from models with hundreds of billions of parameters only offer post hoc explanations that are plausible narratives, not true process reconstructions. This legal gap is substantial but largely unacknowledged.
The politically most consequential trait is that generative systems don’t just relay content made elsewhere—they create it. When millions ask a single AI platform historical, ethical, and political questions, receiving answers with a harmonized voice and embedded priority hierarchy, a novel infrastructure to shape public opinion emerges. The statistic suggesting that about seventy percent of ChatGPT interactions are personal rather than professional doesn’t indicate a market—it reflects a relationship.
The constraint on political forms and the international question
From these factors emerges the core thesis of this essay: artificial intelligence is not politically neutral regarding the regimes that implement it, as it favors political structures compatible with its architecture.
This selection happens on three levels.
On the decision-making level, a system promising real-time optimization favors political frameworks with short command chains. Parliamentary debate, judicial process, and public consultation are deliberately slow—a cost valued for producing legitimacy over efficiency. Where AI accelerates administration without political decision on what should not be expedited, deliberation erodes through technical means rather than ideology. This is no coup, but a quiet shift of authority from deliberative sites to computational centers.
Regarding political representation, algorithmic personalization fragments the public into individualized realities. Modern representation depends on an electorate sharing some basic consensus on reality, even if opinions vary. Systems that tailor information individually don’t just deepen polarization—they cause the disappearance of a shared object around which debate can occur. While representative democracies survive sharp differences, it’s uncertain if they can endure fragmentation of the very subject of disagreement.
At the substantive sovereignty level, states providing vital public services through externally trained, hosted, and updated models do more than outsource technology—they cede certain normative powers. The standards baked into these models—defining acceptable content, prioritizing values, and structuring responses through language and legal traditions—function as de facto norms without legislative enactment. The concept of digital sovereignty misses this point by focusing on data localization, rather than control over the embedded decision-making criteria.
The effects are asymmetrical across regimes. Authoritarian states integrating AI into their administration gain efficiency without conflicting with their constitutional basis. China’s approach—blurring lines between private firms, academia, and government and funding computational power publicly—faces no fundamental conflict between technology and state structure. Liberal democracies, however, adopt tools that undermine the very processes from which they draw legitimacy. Predictions that the Internet would inherently benefit open societies underestimated this imbalance.
In international relations, computing power now represents a strategic asset—its distribution shaping the hierarchy of players. By 2026, this reality had an administrative form. U.S. export controls categorize countries by tiers for semiconductor access: India enjoys unrestricted access, while the UAE and Saudi Arabia require licenses and accept U.S. oversight as part of government agreements.
This phenomenon demands an accurate label: a system of negotiated limited sovereignty, where access to computational resources is traded for external supervision rights. Legally commercial, politically it resembles dependency relationships described in informal imperialism literature regarding international credit and military bases.
Four significant fault lines are crystallizing.
First is the tension between regulatory authority and technological power. The European Union had bet on regulation. The AI Digital Omnibus, finalized by the Council on June 29, 2026, and signed July 8, delays the rollout of rules affecting high-risk AI systems until harmonized standards are available, pushing deadlines to late 2027 and mid-2028. Though technically framed as simplification, politically it shows the world’s most ambitious regulatory framework reopening under competition and lobbying pressures before it made an impact. The so-called Brussels Effect only works when the regulated market is indispensable. If the core infrastructure is imported, regulations become negotiable.
The second fault line involves emerging technological blocs. Export controls, data localization rules, and incompatible technical standards foster ecosystems increasingly isolated from each other. This fragmentation arises from individually rational defensive choices that collectively undermine interoperability and security regimes, which require universal coverage to function effectively.
The third deals with states that neither develop nor regulate AI. Most countries face a choice between dependence on American or Chinese infrastructure, with sovereign AI platforms financially out of reach for nearly all. The idea of sovereign AI mainly comforts by framing supplier selection as a strategic decision. Realistically—even acknowledged by autonomy advocates—no country can rebuild the entire technology stack. The core question is what to develop domestically, what to procure, and with whom to form alliances.
The fourth and least discussed concerns the dynamic between states and frontier AI labs. In major democracies, the trend is no longer regulating private AI but integrating it into state power. Once AI touches critical state interests—military, geopolitical, security—private autonomy is no longer tolerated. Mechanisms vary: direct political pressure in Washington, concession through competition in Europe, structural assimilation in Beijing. The likely medium-term result is neither an open AI market nor one fully subject to law, but a condominium between state security agencies and frontier labs outside market and parliamentary control. The scope of the issue exceeds common assumptions.
The European formula: between method and uncertainty
A notion gaining traction in Italy suggests that true sovereignty lies not in creating the engine but deciding where the machine is directed. This is a neat idea—yet only partially correct.
It holds true when regulatory capacity is backed by tangible leverage. It fails when leverage is absent, since rules only hold as long as those regulated comply willingly. The Digital Omnibus shows how quickly such willingness can erode under competitive pressure.
Three prerequisites seem unavoidable here as well.
The first is building a European public computing capacity independent from commercial providers. The goal isn’t to compete with the largest global labs by training frontier models—a race likely lost—but to maintain infrastructure that guarantees essential public services without reliance on perennially renegotiated contracts. This transcends industrial policy; it’s a matter of continental security and warrants adequate funding.
The second concerns a robust antitrust policy, notably absent from frontier AI labs’ self-regulation documents. These propose internal governance through aligned missions, philanthropy, and hybrid structures. Their frequent allusions to the Progressive Era and the New Deal are selective; the latter was not mere redistribution but a period of monopoly breaking enabled by public authority acting without regulated consent. Invoking Roosevelt but ignoring the Sherman Act remains rhetorical.
The third condition involves constitutional protection of political processes. It’s vital to constitutionally designate which public decisions cannot be made or prepared by automated systems, notwithstanding their accuracy. The reason is political, not technical: legitimacy often depends on how decisions are reached. Deliberative slowness is a feature, not a flaw, of constitutional procedure. An algorithmic judicial decree—even if statistically more accurate than a human judge’s—would lack legitimacy, since judgment entails responsibility by a legal subject.
Currently, none of these three conditions is met, and recent developments have moved contrary to all three.
This analysis relies extensively on corporate statements, intentions, and claimed capabilities. Historically, the sector has often shown a wide gap between promises and practical capacity. It’s plausible that AI’s current limitations—the so-called jagged frontier, the disconnect between mathematical accuracy and genuine generative creativity—are structural rather than fleeting. The current investment cycle may end before delivering the expected powers, potentially reducing the scale of political transformation anticipated here.
Nevertheless, the core political argument stands independently.
Frontier labs’ influence does not mainly derive from their existing capabilities. Instead, it comes from the credibility of the powers they promise and from the anticipatory adjustments by governments, corporations, educational systems, and societies based on those promises.
Factories shift location. Universities alter curricula. Export controls are renegotiated. Material infrastructures evolve preemptively, preparing for capabilities yet unrealized.
This aligns perfectly with the traditional definition of political power: the ability to make others act according to one’s expectations.
Simple, isn’t it?
The pledge abandoned in April 2026—the commitment to yield to any safer competitor—reveals more than any strategic outline ever could. An actor relinquishing self-restraint exactly at the moment when it has the strength to uphold it is not merely clarifying strategy.
It signals that it has entered a new category.
And if this category can reshape the future of political order itself—and if we are just beginning to grasp this—how profound is the significance of what is unfolding today?
About the Author
