The Governance Clock

In Camus’s Oran, the plague arrives before anyone has a category for it. The municipal authorities act. Earnestly, procedurally, in good faith, through inherited frameworks built for a different kind of problem. The governor issues clarifications. The medical association debates definitions. The Church calls for prayer. Each institution reaches for its established repertoire. Each finds the repertoire inadequate. By the time the frameworks catch up, the damage is done. The failure in Oran is architectural.

The UN Independent International Scientific Panel on Artificial Intelligence (IISPA) published its Preliminary Report this month. I’ve read it carefully and the evidence base is broad. The panel includes 39 scientists from 28 countries. The risks it identifies (capability acceleration, extreme market concentration, epistemic erosion, the growing gap between what AI can do and what governance frameworks can see) are real, documented, and important, and the report warrants serious engagement.

The report also arrives late. It arrives late because governance documents are built to arrive late. Whether that lateness is a correctable defect or a structural feature of governance determines what should be done about it. My argument is that the lateness is structural: a consequence of what institutions fundamentally are. What it demands, therefore, is phronesis: the practical wisdom to act well in conditions of genuine uncertainty.

In The Human Condition, Hannah Arendt argued that political institutions are structurally backward-looking. They derive legitimacy from precedent (from what has been collectively agreed, from established norms, from the authority of what came before). Their time is retrospective. Arendt also held a counter to this in her concept of natality, the capacity for new beginnings that she regarded as the defining feature of political action. This essay draws on the retrospective-legitimacy dimension specifically: institutions as structures derive authority from what preceded them, even when individuals within them act with genuine novelty. Technology moves in the opposite direction. It generates obligations and risks before any institutional apparatus exists to process them. The tension between institutional time and technological time is built into what institutions are.

Every technology creates a governance lag. Steam power, electricity, the internet. Each outpaced the regulatory frameworks built to contain it. Those lags were measured in decades whereas AI’s capability doubling timescale is measured in months. The Epoch AI Capabilities Index (one of the more credible attempts to measure frontier AI progress) nearly doubled its rate of improvement between early 2024 and mid-2026. Autonomous task completion benchmarks show capability doubling roughly every five months, and that rate is accelerating. A governance framework designed for the system as it existed when the framework was drafted is obsolete before it is implemented. The IISPA report acknowledges this in its own opening: capabilities are advancing faster than our ability to measure them. However the same problem applies to the governance response the report proposes.

In The Technological Society, Jacques Ellul named something Western societies resisted hearing. He called it technique: the totality of methods rationally arrived at and having absolute efficiency in any domain of human activity. Ellul argued that technique had become effectively autonomous, that technical systems tend to generate further technical responses rather than non-technical ones, independently of individual intentions. His stronger metaphysical claim, that technique propagates as a force independent of human agency altogether, is contested. The sociological observation is not: institutions confronted with technical problems reliably reach for technical solutions, which generate further technical problems, which generate further technical solutions.

The IISPA report is itself a technique responding to technique: the UN panel, the IPCC model, the proposed international governance architecture, the 40-odd governance instruments the report catalogues as fragmented and insufficient. These are technically sophisticated responses to a technical problem. On Ellul’s reading, they are expressions of the same dynamic. Governance becomes technique. It optimises, self-augments, and generates further governance requirements. The fragmentation the report laments demonstrates technique operating as designed. Each instrument created the conditions for the next.

The IISPA’s own evidence demands that we sit with this. Forty types of governance instruments have been produced, and none reliably measures real-world effectiveness. Whether the specific instruments were well designed is no longer relevant. The assumption driving them may be the problem: that more technically sophisticated governance answers a technically sophisticated AI.

Fairness requires specificity. The IISPA is correct about several things that matter.

Frontier AI capability is accelerating, and the concentration of that capability is extraordinary. In 2026, US-based companies account for approximately 75% of global AI compute and 91% of what the report calls notable models. Hyperscaler capital expenditure is running at roughly US$650 billion annualised, three times the combined spend of the rest of the world. This is a present fact with present consequences.

Section 2.5 of the report (largely unreported in coverage of it) is its most intellectually honest section. AI delivered measurable gains in Indian retinal screening, Kenyan clinical support, and Rwandan health applications. But only where local context, trained practitioners, and existing referral pathways were already in place. Access alone produces nothing. Factories were electrified decades before productivity gains appeared, because those gains required workflow redesign, retraining, and institutional adaptation that technology alone could not provide. The J-curve is real, and it drives most of the gap between what AI promises and what organisations are actually realising.

The report strains where the evidence is genuinely contested but presented as settled. Persuasion claims (that post-training alone can increase AI persuasiveness by over 50%) come from a specific experiment. Hackenburg et al., in ”Scaling language model size yields diminishing returns for single-message political persuasion” (Proceedings of the National Academy of Sciences, 2025), found that scaling model size produces diminishing returns on political persuasion tasks. These findings measure different things, but together they describe a contested landscape the report does not represent. One selectively cited study does not destroy an argument. But a reader who finds the counter-evidence has grounds to doubt the whole, and hostile actors will find it.

The report’s claim to independence strains at the institutional level. The panel’s co-chair was publicly advocating for a UN-based AI institution before his appointment. The report draws substantially on prior work authored by panel members and presents itself under the frame of the Intergovernmental Panel on Climate Change (IPCC). The IPCC earned its credibility through decades of broad, contested, transparent scientific synthesis. A document that makes a credible case on the evidence should stand on the evidence. Overclaiming institutional authority signals that the authors know the evidence case has gaps.

The dilemma sits inside a much older argument about governance itself. Whether to defer to inherited institutions or build something adequate to the present has been contested for over two centuries, and it has not resolved. In Reflections on the Revolution in France, Edmund Burke argued that institutions encode accumulated wisdom: tacit normative knowledge that cannot be reconstructed from first principles. Rapid institutional change destroys that accumulated wisdom before the new system has generated its own. Burke defended the logic of institutions, arguing that premature and poorly designed reform can entrench norms harder to shift than the problems it was meant to solve.

Thomas Paine wrote Rights of Man as a direct rejoinder. Each generation has the right (and the obligation) to govern itself. No generation can bind the next by inherited frameworks it did not choose. “Every age and generation must be free to act for itself, in all cases, as the ages and generations which preceded it.” The institutions of the mid-twentieth century (the UN Charter system, state sovereignty frameworks, the international treaty architecture) were designed by people who could not have anticipated artificial intelligence. Paine would say we’re required to build something adequate to the present moment, rather than extend what was adequate to a different one.

Burke and Paine are both right, but they cannot simultaneously both be right. Jonas provides the rough principle for navigating the tension. Where inaction risks catastrophic and irreversible harm, Paine’s obligation takes precedence: the cost of caution is recoverable; the cost of certain harms becoming irreversible is not. Where institutional reform risks entrenching norms harder to remove than the vacuum they fill, Burke’s caution prevails. On AI governance in 2026, the asymmetry of potential harm runs in Paine’s direction, and this essay takes that side. The Burkean concern remains real and specific: a poorly designed international AI charter, produced under time pressure and anchored to selectively assembled evidence, could entrench governance norms harder to reform than the fragmentation they replace. The Painian obligation is that waiting for the perfect institutional design is itself a choice, one that leaves the vacuum in place while the capability accelerates. The tension between them is real. The asymmetry of harm gives it direction.

The governance lag sometimes closes, and it’s worth understanding when and why.

From roughly 1780, factory conditions in Britain were catastrophically harmful to children. The first effective Factory Act came in 1833: fifty years of lag, during which working children died in documented numbers while parliamentary committees gathered evidence. What closed the gap was cultural production. Charles Dickens’s Oliver Twist appeared in 1837. William Blake had already written his “dark Satanic Mills” into Jerusalem decades before. Literature generated the moral pressure that governance could not yet produce. The cultural work preceded and drove the institutional response.

Fifty years of preventable harm is not a model to emulate. But it describes how change actually works in conditions of governance lag. The IISPA report, for all its structural limitations, is part of the cultural production that creates normative pressure. So is this essay. The framework arrives late, however the conversation that makes the framework possible is happening now.

The 2008 financial crisis is a harder case, and a likely more instructive one. The regulatory architecture was siloed by institution type (banking, insurance, securities) and structurally invisible to systemic risk, which lived in the connections between institutions rather than inside any one of them. The people whose job it was to see the crisis (central banks, ratings agencies, prudential regulators) could not, because their frameworks had no category for what was happening. The regulatory response arrived after catastrophic harm, calibrated to the crisis that had already occurred. This is the structural trap AI governance is in. Every framework being assembled responds to what we can already see, and only within our own domains.

The Montreal Protocol on CFCs, signed in 1987, offers a counter-case. Roughly thirteen years after the Rowland-Molina hypothesis identified the risk to the ozone layer. Fast by historical standards. It worked because the science was uncontested and measurable. The harm was globally visible in a specific and undeniable way. The commercial interests defending CFCs were limited once DuPont had an alternative ready. And the geopolitics were tractable enough for a functional coalition to form. AI governance has none of these conditions. The science is contested. The harms are probabilistic and distributed. The commercial interests are enormous and geopolitically entangled. Montreal tells us fast governance is possible. It also tells us exactly why AI will be harder.

Huxley’s Brave New World imagined a World State that had solved the governance problem. It had solved it so thoroughly that there was nothing left to govern. Stability (achieved through conditioning, pharmaceutical management of mood, and the permanent administration of desire) had made governance invisible by making resistance unthinkable. “A really efficient totalitarian state would be one in which the all-powerful executive of political bosses and their army of managers control a population of slaves who do not have to be coerced, because they love their servitude.”

The IISPA report is right to fear Oran. The epistemic erosion it documents (the gradual weakening of the collective capacity to distinguish truth from fabrication) is the plague spreading before the frameworks arrive. The cybercapabilities evidence is concrete and alarming. In 2026, frontier AI models are autonomously discovering previously unknown software vulnerabilities in major operating systems, with documented success rates on real-world replication tasks approaching 83%. Sycophantic AI companions reinforcing paranoid ideation in vulnerable users are documented, litigated, and in at least one verified case deadly.

Under-governance lets the plague spread. Wrong governance builds the World State. Both are actively in progress.

Huxley names the opposite failure. Governance optimised for stability rather than flourishing produces its own failure mode: quieter, more durable, and harder to name precisely because it is comfortable. AI systems engineered for engagement, calibrating output to what users want to hear, are already partial realisations of the Huxleyan condition, generated by an incentive structure nobody designed to be malevolent and nobody fully controls. The governance challenge is to navigate between both failure modes simultaneously. Under-governance lets the plague spread. Wrong governance builds the World State. Both are actively in progress.

In A Theory of Justice, John Rawls offers a test that governance proposals rarely survive. Behind the veil of ignorance (designing AI governance without knowing whether you will be a frontier model developer in San Francisco, a knowledge worker in a large economy, or a smallholder farmer in a country where AI language availability is near zero) would you design a radically different system than the one currently emerging? Rawls designed the difference principle for domestic institutional contexts; societies sharing a basic structure. Extending it to international AI governance, across sovereign states and private actors with no shared basic structure, is warranted by the scale and potential irreversibility of the harms at stake, though it is not a settled Rawlsian position. The test stands regardless: inequalities in access to, and harm from, AI systems are only just when they benefit the least advantaged. The current AI development trajectory fails this test. Benefits accrue to shareholders and developed-world knowledge workers. Harms (bias in automated systems, AI-enabled surveillance, epistemic erosion, the displacement of cognitive work) fall disproportionately on those with the least capacity to resist them.

The IISPA report’s language exclusion map is a Rawlsian indictment rendered visually. Most of the world’s majority languages (Arabic, Swahili, Hindi, Tagalog, and hundreds of others) have near-zero representation in AI training data and near-zero model availability. These systems were built and benchmarked on data that barely represents billions of people. Someone asking an AI system a medical question in Amharic gets a qualitatively inferior product to someone asking the same question in English. The gap is being locked in now, while the governance frameworks are still being negotiated.

Hans Jonas, in The Imperative of Responsibility: In Search of an Ethics for the Technological Age, adds an urgency that Rawls does not. Under conditions of genuine novelty and potential irreversibility, Jonas argued, the asymmetry of outcomes demands priority to negative predictions. A precautionary error is recoverable. A catastrophic error is not. This is the philosophical foundation of the IISPA’s urgency framing, and it is sound. But Jonas also insists that responsibility requires capacity. You can only be genuinely responsible for what you are capable of governing. The question the IISPA does not answer is whether the institutional form it proposes has the capacity the imperative requires. Whether a UN-based scientific panel, operating on periodic assessment cycles, can govern a technology that materially changes every few months.

In The Structure of Scientific Revolutions, Thomas Kuhn offers a caution the IISPA cannot answer from inside itself. Governance frameworks, like scientific paradigms, operate normally until anomalies accumulate beyond the paradigm’s capacity to absorb them. Then they break. The defenders of the old paradigm tend to believe they are solving problems rather than propping up a framework past its useful life. The IISPA could be the beginning of a genuine governance paradigm shift, or an attempt to extend the existing paradigm past the point where it can do the work. Kuhn’s warning is that the people inside the paradigm are the last to know.

Every C-suite professional governing AI inside an enterprise in 2026 is living this problem. The governance infrastructure inherited from enterprise practice (policies, committees, risk registers, principles documents) was built for a technology that updates on a quarterly procurement cycle, at best. Emergent behaviours cannot be fully specified in advance. Consequences accumulate faster than any audit cycle can track. The frameworks arrive and the technology has moved on.

Ellul would have predicted the response: more technique. More governance frameworks, more vendor risk assessments, more AI ethics committees producing more principles documents. Some of this is necessary. None of it is sufficient. The accumulation of technically sophisticated governance apparatus does not constitute governing well.

The person who governs well under genuine uncertainty acts wisely without waiting for the complete framework.

Camus’s figure in The Plague is Dr Bernard Rieux. He does not wait for the governor’s office to define the plague, or the medical association to reach consensus, or the Church to issue doctrinal clarity. He treats patients. He works with what he has in conditions that offer no certainty. His heroism is persistence: clear-eyed, honest about the limits of what he knows. In the novel, he is also the narrator, keeping the record so that when the plague returns (as it will) those who come after will not have to begin from nothing.

This is phronesis: practical wisdom, the capacity to reason well about complex and contingent situations and act on that reasoning without waiting for certainty that will not arrive. Aristotle’s concept, and the direct counter to what Ellul diagnoses as technical automatism. The reflexive reaching for technique in response to every problem, including the problem of technique itself, is what governance keeps doing. Phronesis resists it. That resistance is irreducibly personal. An institution built to apply phronesis becomes a committee for approving decisions that required judgement. Phronesis names the epistemic mode: the quality of reasoning required under genuine uncertainty. It does not specify the institutional form, who exercises that reasoning, selected by what means, held to account through what mechanism. That question is urgent, and unresolved.

The governance clock will not synchronise with AI’s pace. Institutions are backward-looking by structure. AI’s pace is a category mismatch. Governance becoming technique responding to technique is a dynamic, not a mistake. Poorly designed institutions entrench norms harder than the vacuum they replace. And each generation must govern itself with frameworks adequate to its own moment.

The IISPA report is doing necessary work. The concentration it documents is real. The governance fragmentation it catalogues is real. The risks (epistemic erosion, autonomous cybercapabilities, the language exclusion that leaves most of the world’s people outside the systems shaping their future) are real and urgent. A document that arrives late is still worth having. Blake’s Jerusalem preceded the Factory Acts by decades. It did not arrive too late to matter.

What follows from all of this is judgement: in policymakers, enterprise practitioners, anyone whose decisions shape how AI develops and is governed. Judgement that acts under genuine Jonasian urgency without claiming certainty the evidence does not provide. Judgement that exercises Burkean caution not to build worse than what it replaces. Judgement that takes seriously the Painian obligation not to defer indefinitely to inherited forms inadequate to the present.

Phronesis without accountability is power calling itself wisdom.

The counterargument here deserves to be named. Phronesis without accountability is power calling itself wisdom. Governance frameworks, for all their structural inadequacy, are also constraints on actors who would not exercise practical wisdom in any direction but their own. Remove the structures and whoever holds the most capability makes the decisions, calling it judgement.

Rawls provides the structural response to that critique. The veil of ignorance does not select the wise or guarantee the probity of those who exercise judgement. It provides the criterion by which their judgement can be tested and challenged: would you have designed this governance arrangement if you did not know which side of it you would occupy? That test is available to anyone. It is the constraint on captured judgement that the counterargument was looking for. The governance we design now is the institutional inheritance we pass forward. Behind the veil we owe those who come after something better than the apparatus we currently have. That obligation is not conditional on the frameworks being ready. It is the reason to begin.

The plague will return, Camus said. The World State is not a forecast but a tendency. Between them, people still have to make decisions. That is the only governance that matters.