GENEVA—Artificial intelligence (AI) governance has become remarkably good at producing principles. Every few months, governments, companies, and international organizations release new frameworks for AI safety, security, responsibility, or trust. In the past two years alone, governments and international organizations have released the Bletchley Declaration, the Hiroshima AI Process, the Group of Seven Code of Conduct, the Organisation for Economic Co-operation and Development’s updated AI Principles implementation work, and multiple national AI strategies. Each is valuable individually, but collectively they illustrate how guidance is outpacing implementation. Despite this growing body of guidance, one question remains surprisingly unanswered: Who is actually responsible for making trust in AI work in practice?
Meanwhile, AI capabilities continue to evolve at a pace that routinely outstrips the institutions meant to govern them. As new models become more capable, governance discussions often return to the same familiar debates over principles and regulation, while far less attention is paid to building the institutions needed to implement them.
This gap is becoming one of the defining governance challenges of the AI era. Developers, policymakers, and commentators are spending considerable time debating frontier models, existential risks, and regulatory approaches. Much less attention is being paid to the institutions that will have to operationalize these ambitions—to test systems, share information, coordinate responses, develop common practices, and create confidence across borders. Without this institutional layer, AI governance risks becoming little more than an accumulation of disconnected initiatives. Different governments, companies, and research organizations increasingly develop their own methodologies, making meaningful comparison and collective learning far more difficult.
This challenge was a recurring theme during the first UN Global Dialogue on AI Governance in Geneva. While discussions covered familiar topics, from international cooperation to the role of standards, a quieter but more consequential conversation emerged around the practical infrastructure needed to sustain trustworthy AI. The question is no longer simply what principles should govern AI, but how countries, organizations, researchers, and industry can work together to implement them.
That conversation points toward a broader realization. AI trust and safety cannot be reduced to a narrow model-testing and evaluation exercise. It has become fashionable to think of AI safety primarily through the lens of model evaluations or red-teaming for weaknesses. Those capabilities are undoubtedly essential. But trust encompasses far more than technical robustness. It includes data governance, privacy, cybersecurity, transparency, accountability, content authenticity, incident reporting, and mechanisms for public oversight. If these remain fragmented across many institutions and jurisdictions, confidence in AI systems will remain equally fragmented.
This is where the emerging ecosystem of AI safety, security, and trust institutes could play a transformative role. These institutes are emerging in different forms. Some are national bodies established or supported by governments, others are independent research organizations, while still others operate through university or public-private partnerships. Their institutional models differ, but their shared function should be to generate technical expertise, evaluate AI systems, develop common methodologies, and serve as trusted nodes within an international network rather than isolated national projects. Rather than simply serving as laboratories for testing frontier models, these institutes should become part of a broader public infrastructure for AI governance.
Public infrastructure does not imply centralized control. The internet itself offers a useful analogy. Its resilience stems not from a single authority but from interoperable standards, distributed institutions, and shared technical practices that enable independent actors to collaborate. AI governance requires something similar: a federated network of institutions capable of generating shared knowledge while respecting regional priorities and national contexts.
Such an approach would shift attention from creating ever more organizations toward improving coordination among those that already exist. The challenge is not the absence of actors but the absence of mechanisms that allow them to cooperate effectively.
That cooperation should rest on several foundations.
First, common reporting standards. Today, incidents involving AI systems are reported inconsistently, if at all. Without shared taxonomies for documenting failures, vulnerabilities, or misuse, it becomes nearly impossible to learn collectively or compare risks across jurisdictions.
Second, interoperable infrastructure. Institutes should be able to exchange methodologies, benchmarks, datasets where appropriate, evaluation techniques, and lessons learned without requiring every country to reinvent the wheel. Shared capabilities reduce duplication while strengthening collective resilience.
Third, continuous measurement of ecosystem maturity. Governance discussions often focus on whether countries have adopted strategies or regulations. Equally important is understanding whether they possess the institutional capacity to implement them. Mapping AI trust and safety capabilities should therefore become an ongoing process rather than a one-off report. As the technology evolves, so too must our understanding of who can perform which governance functions.
Perhaps most importantly, coordination should avoid reproducing existing geopolitical asymmetries. Much of today’s AI governance ecosystem remains concentrated in the Global North. Many of the world’s most influential models are developed, evaluated, and governed by institutions located in a handful of countries. Yet these same systems are increasingly deployed across societies that have little influence over the standards governing them. This lack of inclusion has become a significant threat to AI’s future, something that was also recognized earlier this month in Geneva.
This imbalance extends beyond representation. Languages remain underrepresented, local contexts are often overlooked, and evaluation methods developed in a handful of countries do not always capture how AI systems perform across diverse societies. Governance risks becoming an exported product rather than a genuinely global public good.
A federated model offers a more promising alternative. Regional networks could develop context-specific expertise while contributing to shared international baselines. Africa, for example, has already begun advancing regional AI cooperation through initiatives such as the Smart Africa AI Council. Similar regional trust and safety frameworks could create feedback loops between local experience and global governance, ensuring that knowledge flows in both directions rather than from a single center outward.
This also suggests that future coordination efforts should focus less on ranking individual countries and more on understanding regional ecosystems. Continents and regional organizations often provide more meaningful governance units than individual countries alone. Strengthening regional capacity may ultimately prove more valuable than establishing isolated national institutions. For instance, Europe’s network of AI Safety Institutes and Africa’s Smart Africa AI Council demonstrate how regional cooperation can complement national initiatives while feeding into broader international governance.
Underlying all of this is a simple but often neglected reality: AI governance begins with data. Discussions about frontier models frequently overshadow the foundational layer that enables them. Questions about data quality, provenance, governance, access, and protection shape the performance, fairness, and safety of every AI system. Any trust and safety architecture that overlooks data governance will inevitably remain incomplete.
The international community understandably continues to debate ambitious global governance arrangements. Political agreement, however, will remain difficult and slow. Technical cooperation need not wait. Building interoperable reporting standards, shared evaluation methodologies, regional networks, and mechanisms for knowledge exchange are practical steps that can begin today regardless of broader geopolitical tensions.
If the discussions in Geneva demonstrated anything, it is that AI governance is gradually moving beyond declarations toward implementation. The real test now is whether governments, industry, researchers, and civil society can resist creating yet another collection of disconnected institutions and instead invest in the connective tissue between them.