Copenhagen and the Infrastructure of Knowledge. The AI Act, high-risk systems, and cognitive autonomy
Europe’s extension on high-risk AI systems makes validation no less urgent. The Fable 5 case showed, in forty-eight hours, what happens when an organization doesn’t know
I. A Transition Between Two Worlds (and Then Some)
In September 1941, Niels Bohr and Werner Heisenberg meet in Copenhagen. Germany has occupied Denmark for seventeen months. They are friends, master and pupil, two of the century’s greatest scientists — and the international system around them is undergoing, to put it mildly, a total and brutal reconfiguration. Heisenberg works on the Reich’s nuclear program, and the two can no longer speak to each other as scientists. The historian and playwright Michael Frayn reconstructed that meeting in Copenhagen without ever establishing what was actually said: not because the documents are missing, but because two people operating inside conflicting systems of power no longer share a common language. Physics, until a few years earlier an open international conversation, had become a strategic resource. Knowledge had stopped being neutral.
Eighty years later the same mechanism repeats itself — faster, at greater scale, through a far more pervasive cognitive infrastructure, and apparently with less trauma, at least if we compare our time to the slaughterhouse of the Second World War (a matter of perspective, admittedly: ask the citizens of the former Yugoslavia, or the Iraqis, or the Cambodians). The context is familiar. In 1989 the Cold War system officially collapsed — excellent news, by any measure. But the unipolar system that replaced it tried to harness the world and its many impulses under a single principle: the free flow of goods, services, and people (with enormous exceptions, in truth). The attempt was to replace the paradigm of power with the paradigm of wealth — ideally reserving power itself, in the form of soft power, as the exclusive resource of the planet’s only hegemon at that given moment. Which was, naturally, as always, the United States.
Anyone who loved Aaron Sorkin’s West Wing knows this story by heart. The series closes by prefiguring Obama’s election, in the hope of converting China to democracy through money, and in the explicit conviction that politics was over: the age of digital platforms had arrived. So-called globalization was the cultural software; the material condition was the telecommunications system woven from submarine cables and satellites orbiting the planet at various altitudes. The idea was naive and seductive at once. And, without trivializing much, the ultimate goal of domesticating the Chinese giant by ushering it into the WTO was precisely its greatest success and the beginning of its end. The system began to creak, for reasons there is no room to explore here.
Jump forward to today, and what we have is that same system — we are not, at this precise instant, anywhere else — crossed by political fractures running within and between the players. Which partly confirms that the hegemon is working hard to take itself out of the game.
II. What the AI Act Says (and What It Doesn’t)
One implicit consequence of this scenario: technology, never neutral to begin with, drops the mask and becomes instrumentum regni — an instrument of rule and a device for producing meaning. If in the Middle Ages the central idea conferring sense on the whole organized system of thought was God, today that cohesive role belongs to economic rationality: the concept of value, the ratio of costs to benefits, the bureaucratic mindset. Within this most modern of trinities, whoever isn’t winning the game must at least try to contain the distorting power of technological innovation — especially when that innovation sits in the hands of a hegemon that is not necessarily friendly.
That is a precise photograph of the Atlanticist middle powers, at a moment when the American administration does not seem to find them particularly congenial. With one fundamental caveat: administrations and their antipathies pass, but the center of American concerns runs less and less through their neighborhood. To survive, middle powers increasingly reach for their regulatory power, trying to set standards that limit the discretion of larger actors. It may be the thing the European Union currently does best. And it is what lets us travel from Bohr and Heisenberg to the AI Act.
Regulation EU 2024/1689 — the AI Act — is the world’s first comprehensive legal framework for artificial intelligence: adopted in May 2024, in force since August 1, 2024, applied in phases. Its purpose is to protect citizens, and their governments, from the destabilizing power of digital technologies, with everything that entails. The idea — much professed in the United States — that all this is merely the fruit of a bureaucratic mentality is myopic. Myopic, or instrumental.
Since February 2025, practices deemed to pose unacceptable risk have been banned: social scoring, subliminal manipulation, real-time biometric recognition in public spaces. Protection of the citizen against the overreach of the technological State — a moloch halfway between public and private — is the evident aim. Evident, too, is the regulator’s effort to preempt scenarios already prefigured in recent social fiction, and already operational in countries of less democratic spirit (to be kind) with a strong appetite for centralized power. Black Mirror is gripping to watch; nobody wants to live there. Not in Europe, given the choice.
Since August 2, 2025, the governance institutions and the obligations for providers of general-purpose AI models (GPAI) have been operational. The most demanding phase — the one covering high-risk systems — was expected on August 2, 2026. On June 29, 2026, the EU Council approved the simplification package known as the AI Omnibus, after the European Parliament’s vote on June 16. The deadline for high-risk systems moves to December 2, 2027. Sixteen more months.
What exactly are high-risk systems? Annex III of the regulation lists eight categories: critical infrastructure, education and vocational training, employment and human-resource management, access to essential services (credit, insurance, social benefits), law enforcement, migration and border management, administration of justice, democratic processes. These are not edge cases. They are systems already running in thousands of European organizations.
For these systems, Articles 9–17 impose four substantive obligations. A risk-management system that runs continuously, not a one-off audit. Data governance demonstrating the absence of systematic bias in the datasets used for training and validation. Technical documentation that makes the system inspectable from the outside. And an architecture of effective human oversight — not nominal oversight, not a string of code appended to the end of the process.
Fines for non-compliance reach €15 million or 3% of global turnover per infringement — hardly symbolic figures, and reminiscent of the extreme severity of Europe’s privacy regime. The extension to 2027 is an administrative fact. The problem the AI Act tries to regulate is an epistemic fact, and it does not move with parliamentary votes.
III. The Epistemic Condition of Organizations That Use AI
Most organizations deploying high-risk AI systems have no explicit model of how the system produces its outputs. This is not an accusation; it is a structural description of how systems get adopted. You buy a tool because it works, not because you understand its mechanism. It has always been so, from the introduction of automatic computation onward.
AI, however, introduces a specific discontinuity: epistemic delegation. We are not delegating an operation — we are delegating a judgment. We do not delegate the calculation of ballistic trajectories: we delegate candidate selection, creditworthiness assessment, radiological diagnosis, the ranking of available information. We delegate, that is, the production of the beliefs we then act on. Or at least of the raw material from which our beliefs take shape.
When epistemic delegation happens without explicit validation mechanisms, the organization lands in a peculiar position: it uses outputs as if they were knowledge while unable to answer the questions knowledge requires. It doesn’t know what the system knows. It doesn’t know what the system doesn’t know. It doesn’t know how the system behaves at the edges of its domain, or under what conditions it fails. This is the epistemic condition organizations must confront — not as a regulatory requirement but as a precondition for acting rationally in complex environments. The AI Act made it visible. It did not create it.
Meeting the obligations of Articles 9–17 means, first of all, answering a much older question: how do we justify the beliefs we use to make decisions that affect others? An organization that cannot answer this question about its own AI systems does not have a regulatory problem. It has a serious problem of method. And problems of method are not solved on the eve of a deadline.
IV. From Supercomputers to Fable
In the 1980s the Soviet Union needed supercomputers for physical, meteorological, and nuclear simulations. The Crays — the most powerful machines available — were American. The United States classified them as dual-use technology subject to export control through COCOM, the Coordinating Committee for Multilateral Export Controls. A European researcher working on a collaboration with the USSR could suddenly find access to cognitive infrastructure conditioned by the U.S. Department of Commerce, not by the scientific merit of the project. Science did not stop. It forked. Those with access to the Crays and those without produced incomparable results, developed divergent methodologies, built separate epistemic communities. The fracture did not run between those who knew and those who didn’t. It ran between those with access to the infrastructure and those who depended on someone else’s discretion.
The mechanism is identical to the one that produced the crisis of June 12, 2026 — perhaps still too early to call paradigmatic. On June 9, Anthropic launches Claude Fable 5, the first publicly available model of the Mythos class: the most capable ever released, with landmark performance in software engineering, scientific research, and visual analysis. On June 12, forty-eight hours after launch, the U.S. government issues an export-control executive order. Anthropic switches off both of its new models for every user worldwide, because it has no system for verifying user nationality in real time.
The declared trigger is a report by Amazon researchers flagging the model’s capability in generating exploits for software vulnerabilities. Anthropic contests the characterization: the same capability was present, to a comparable degree, in every competing model — including its own Haiku 4.5, the cheapest in the family. Independent experts agree the government overreacted. On July 1 the model comes back online.
The Crays and Fable 5. The technology changes: from supercomputers to frontier language models. The scale changes: from a handful of scientific institutions to millions of users. The logical structure stays put. Whoever controls the cognitive infrastructure controls the conditions under which knowledge is produced. And the structure that produced the crisis remains intact. Among the affected users were scientific researchers — chemists, biologists, physicists — who had received early access to Fable 5 and integrated it into their workflows, thereby manufacturing a vulnerability inside their own systems: without any awareness of it, by the effect of an American executive order.
One detail makes the picture worse: in the days before the suspension, the model was silently refusing or degrading legitimate scientific requests, without notifying the user of the downgrade. A researcher asking the model to analyze chemical structures received weakened answers without knowing they were weakened. Under these conditions, in what possible sense can the technology Anthropic provides be called a neutral instrument?
Transparency is the precondition of validation. A system that degrades its own outputs without declaring it cannot be validated, because the validator does not know what it is validating. To borrow the cooperative model once more: the researcher who actively integrates Fable 5 into a working routine is, in effect, cooperating with a cognitive infrastructure orchestrated by the U.S. government. In that context — even for perfectly defensible national-security reasons — the executive order is a defection, pure and simple.
This is why Article 13 of the AI Act, which imposes transparency obligations on high-risk systems, is no bureaucratic whim. It is the minimum condition under which knowledge produced by an AI system can count as knowledge at all. Because the understanding of the game here is asymmetric: researchers may not know that the system orbiting Fable has defected, and may therefore pay the cost of their cooperation without knowing it — and so without learning anything useful for the next move.
Simplifications and models aside, the structural knot holds. The chain of control over a high-risk AI system does not end with the organization that uses it. It includes the provider, the provider’s government, intelligence agencies, competitors who flag vulnerabilities to regulators. An organization without documented validation processes does not even know how many actors can interrupt its operations, or on what criteria.
Game theory has a precise concept for this: the shadow of the future. Cooperation becomes possible when actors know they will interact again and therefore have a stake in preserving the relationship. But the shadow of the future presupposes that you know who the other players are and what moves are available. An organization that uses opaque AI in environments of strategic interdependence blunts every other actor’s capacity for judgment — perhaps its own understanding of the system too — and this can tip the game negative-sum. Nobody wins. In the long run, perhaps not even the technology’s owner.
The validation framework behind this argument is documented — paper, code, and blind-test packet — in the Episteme Advisory Lab:
V. Science Stops Being Neutral
The suspension of Fable 5 is not an isolated episode. It is the surfacing of a structural tendency that the Copenhagen meeting had already anticipated, in the most tragic form possible: when a geopolitical system comes under tension, cognitive infrastructure becomes — at minimum — an object of control.
Anthropic now supplies platforms for the sciences: productive matrices in the full sense. This holds generally — one of AI’s most obvious effects is a rise in productivity. In the case of science, though, AI provides a cognitive infrastructure that is both process and product. And that infrastructure, as it happens, had to obey an American executive order within forty-eight hours. Whoever uses that platform to produce scientific knowledge is using a tool that can be switched off, modified, or degraded at any moment — without notice, without appeal, on the basis of political judgments that have nothing to do with the quality of the research.
Open science remains a fundamental goal, but science was never really neutral. It looked neutral inside a system that presumed itself neutral: the post-1989 order, with its multilateral institutions, its open markets, its rhetoric of knowledge as a common good. When that system comes under tension, science risks taking the shape of the power structure rather than the shape of truth. Not because individual researchers abandon their professional ethics — because the cognitive infrastructure they work on is subject to the same logics of control as any other critical infrastructure.
With the system in crisis, the question has left the seminar room and entered praxis. Dependence on foreign cognitive technologies is not only an operational risk. It is an epistemic risk in the most literal sense: whoever controls the infrastructure controls the conditions under which knowledge is produced. Building documented validation processes is, in this context, an act of sovereign cognitive hygiene: knowing what your system knows, how it knows it, and under what conditions it might stop knowing it.
On the other side — neutrality or not — we cannot abandon the regulative idea of truth altogether. It has to remain the common ground from which to judge the moves of other actors and design one’s own. With one great difficulty: documented validation processes cannot be built retroactively. This is their most important and perhaps least understood property. A risk-management system requires historical data: incidents, near-misses, behaviors at the edges of the application domain. Data governance requires traceability of choices made upstream — during training, during dataset selection, during evaluation. Human oversight requires an architecture designed to make it possible, not bolted on afterward as a wrapper.
None of this is produced in six months. There was no time for it by August 2026; it will not produce itself by December 2027. Whoever reaches 2027 with today’s empty structures will reach 2027 with the same empty structures — and twelve fewer months to fill them. The competitive advantage is not in compliance. It is in the epistemic capacity that compliance, properly understood, forces you to build. An organization that knows how its AI system works, has documented its limits, and has tested its behavior at the margins holds a more accurate model of the reality it operates in. That accuracy is not a normative value. It is a strategic one.
The extension, then, is an opportunity. A small one. Not an invitation to wait: a chance to do well what needed doing fast. To build not the minimum compliance required to pass inspection, but the processes that make compliance an epiphenomenon of a capacity already acquired. Proceeding is not necessity in the sense of the Greek ananke. It is a decision about the quality of the knowledge available to those who decide. The what precedes the how.
Sources
ANTHROPIC, “Statement on the US Government Directive to Suspend Access to Fable 5 and Mythos 5”, June 12, 2026.
AXELROD, Robert, The Evolution of Cooperation, New York, Basic Books, 1984.
COUNCIL OF THE EUROPEAN UNION, “Artificial Intelligence: Council Gives Final Green Light to Simplify and Streamline Rules”, press release, June 29, 2026.
EUROPEAN PARLIAMENT AND COUNCIL OF THE EUROPEAN UNION, Regulation (EU) 2024/1689, June 13, 2024, Official Journal of the European Union, L 2024/1689.
FRAYN, Michael, Copenhagen, London, Methuen Drama, 1998.
MASTANDUNO, Michael, Economic Containment: CoCom and the Politics of East-West Trade, Ithaca, Cornell University Press, 1992.
NOVET, Jordan, “Anthropic Says Trump Admin Has Lifted Export Controls on Claude Fable 5 and Mythos 5”, CNBC, June 30, 2026.
SIRCAR, Anisha, “Anthropic Disabled Fable 5 And Mythos 5 After A U.S. Export-Control Order. Here’s What Happened”, Forbes, June 16, 2026.