Progress on global AI governance may only be possible when the costs of inaction become too great to ignore.
International AI governance is at risk of failure. Not through a lack of foresight, but because the present-day political economy of AI makes binding, enforceable coordination on many aspects of AI governance close to impossible.
Significant power imbalances have emerged in AI. Private corporations, rather than states, increasingly control access to cutting-edge computational power, frontier models and the research trajectories that define future capabilities. This remains the case even as lower-capability AI diffuses globally.
The US and China are engaged in a contest for technological supremacy, with both nations seeing AI as essential to establishing or extending global geopolitical dominance. The so-called middle powers are betting growth strategies on AI development, with countries like the UK explicitly prioritizing innovation over regulation to maintain their position in a changing global economy.
But while AI governance activity has accelerated as a result of these imperatives, most existing efforts are limited to transparency, risk classification or voluntary restraint. Few seek to constrain the development of frontier capabilities, cross-border deployment or military and defence integration.
While limited coordination is possible in narrow domains, the problem of global AI governance is not going to be solved by better-designed summits or clearer sets of principles alone. At the frontier, there is a fundamental misalignment among the main players. Nations chasing growth are unwilling to countenance regulatory friction. Geopolitical rivals cannot trust mutual constraints on military capabilities. Recent governance efforts – ranging from legislative developments to non-binding codes – risk documenting this reality, instead of shaping it.
Proponents of inclusive, effective and global AI governance must confront a difficult truth: that rapid progress towards global AI governance may only become politically feasible when the costs of inaction become too great.
An AI crisis could take one of several forms. Financial markets frozen by cascading AI failures. Autonomous weapons systems crossing ‘red lines’, forcing a deadly response. Model failures demonstrating a clear gap between assumed and actual human control.
Such damaging, high-visibility moments may create a situation in which resistance to change collapses, and the type of coordination previously thought impossible becomes both feasible and essential. Not because decision-makers suddenly develop foresight and flexibility, but because crisis conditions compress decision timelines, weaken vetoes and elevate coordination as a solution.
International governance regimes have rarely emerged from foresight alone. Instead, they have been forged as a response to systemic failure. For example, the Chernobyl disaster catalysed the development of international nuclear safety norms and transparency obligations. The 2008 global financial crisis resulted in new macroprudential regulation, stronger capital requirements and coordinated oversight mechanisms. In both cases, risks long understood by experts became governable only once a tangible failure collapsed political resistance and recast inaction as unacceptable.
History suggests that nations, institutions and companies that position themselves to respond strategically to a global AI crisis of any nature will be better prepared to shape the development of post-crisis governance. Those who are unprepared, ill-equipped or excluded will have policy change imposed on them. In the past, crises have not reliably produced good governance. But they do frequently produce governance of some kind. Whether that governance is effective, equitable or durable depends heavily on the institutional and technical preparation that precedes the shock. If the world waits for perfect alignment or voluntary restraint, AI governance will arrive too late – if at all. If it waits for crisis without preparing, a system of governance will emerge, but may be badly designed and prone to failure.
