By Ana PALACIO
One way to represent the age of artificial intelligence is with a steep curve and a gentle slope. The curve rises almost vertically: models are rapidly becoming cheaper, more powerful, more accessible, and more deeply integrated into manufacturing, science, education, finance, administration, security, and warfare. Below it, the slope moves at a much gentler angle. This is the capacity of society to absorb change.
The slope includes updates to regulatory regimes, legal structures, and the powers of public administrations. It also includes changes in school curricula, business models, workflows, and workplaces. And it encompasses the ability of citizens to function in this re-envisioned landscape—including their capacity to distinguish truth from simulation, persuasion from manipulation, and assistance from dependence. This is the social metabolism through which a technological shock is integrated into a functional order.
Of course, previous technological and industrial revolutions could also be represented by such a curve and slope. Technologies advanced and spread faster than regulations, business models, education systems, labor markets, and ways of life could adapt. The transitions were often painful and never automatic. But the gap between the curve and the slope was smaller: in terms of speed and scale, the transformation driven by artificial intelligence seems to be in a class by itself. The discussion about how to manage this transition usually focuses on the question of how much regulation strikes the “right balance” between innovation and security. We do not want to put a bureaucratic cap on progress, just as we do not want to leave societies to cope on their own and in a hurry with the consequences of a technology that is transforming our most fundamental understanding of work, knowledge, power, responsibility, and judgment.
But this question is too narrow. Regulation has a crucial role to play, but it does not in itself constitute governance. To do this, we need to define the conditions necessary to enable societies to live with artificial intelligence without being overwhelmed by it.
The first step is to stop thinking of AI as something intangible—an ethereal technology that exists in the “cloud” and delivers whatever demand demands. AI depends on vast amounts of capital, cutting-edge chips, cybersecurity systems, supply chains, and talent. It also relies on large data centers, which emit noise, light, and heat, and require land and permits, as well as vast amounts of electricity and water.
The communities that house the physical foundations of this supposedly weightless technology are now being forced to face major disruptions and bear high costs as their lands are acquired, their electrical grids are strained, their water supplies are depleted and polluted. Difficult decisions must be made about land allocation, electricity prices, and environmental protection.
But governing artificial intelligence is not just a community-level challenge. This technology is increasingly central to the infrastructure of national power: military targeting, cyber operations, intelligence analysis, scientific discovery, industrial automation, surveillance, financial modeling, and political influence. Whoever controls the most advanced models, chips, processing clusters, and talent holds not just a commercial advantage, but a strategic one.
This has not escaped the notice of the United States and China, which are fiercely competing for leadership in artificial intelligence: each believes that falling behind would leave it vulnerable militarily, economically, or politically. As with nuclear weapons, the stakes seem existential. This is a recipe for secrecy, mistrust, deterrence, and uncontrolled acceleration. The risks are exacerbated by the fact that artificial intelligence is more distributed than nuclear technology, more commercially integrated, and more widely deployed. Moreover, its products are difficult to verify, and once the capabilities are encoded in software, proliferation becomes easy.
Governance of artificial intelligence must consider all of these factors, from the material demands of the technology to its effects on individuals, communities, businesses, labor markets, the environment, national security, and international strategic competition. To this end, a comprehensive architecture of verification-based trust—broader than national regulation of artificial intelligence and more flexible than a classic treaty—is essential.
The 2015 Paris climate agreement offers a useful model. It created a living framework that includes voluntary commitments, mutual pressure, regular updates, credible measurements, and broad participation. But a similar agreement for artificial intelligence should also reflect a key lesson from nonproliferation regimes: where political trust is weak, it must be supplemented by testing, limitations, and verification. The resulting framework would combine the flexibility of climate diplomacy with the discipline of nuclear control. Emerging commitments would be accompanied by mechanisms for assessing the credibility and impact of those commitments. State actions would be accompanied by corporate accountability.
A “Paris Agreement on Artificial Intelligence” would provide common definitions for cutting-edge models; thresholds for in-depth review; mandatory incident reporting; common assessment standards; protection of model weights; rules for deployment in sensitive sectors; and independent testing of systems that pose risks to cybersecurity, biosecurity, critical infrastructure, or military decision-making. Each of these elements would be regularly reviewed and strengthened over time, by technical bodies that understand the evolving technology and political bodies capable of imposing consequences when commitments are ignored.
Of course, not every application requires the same treatment. A narrow tool used in low-risk environments requires limited oversight. Conversely, an advanced design that could impact cybersecurity, biological design, military operations, or the autonomy of critical systems should be subject to more stringent requirements and, in extreme cases, intervention.
Some might argue that this would stifle innovation. But no one would board a plane or take a new drug if safety depended solely on the goodwill of manufacturers. In these areas, regulation does not eliminate risk; it makes risk socially bearable. The same is true for artificial intelligence.
Such a framework cannot be presented as a European project, an American instrument, or an alliance aimed at countering China. Its legitimacy would depend on inclusion, where countries on the technological front, “middle powers,” and emerging economies would all have a voice. Universities, companies, and civil society organizations must also be involved. This curve will continue to get steeper. We cannot flatten it by decree, nor obscure it by speeches. But we can decide whether it will grow without governance or within an architecture that can meet the demands.
(Ana Palacio, former Foreign Minister of Spain and former Senior Vice President and General Counsel of the World Bank Group)