A Global Handshake on Safety
In the past couple of years, the conversation around AI has shifted dramatically. High-level meetings like the AI Safety Summit in Bletchley Park and its successor in Seoul have brought together world leaders, top researchers, and the CEOs of major AI labs.
The primary goal has been to establish a shared understanding of the risks posed by advanced AI and to foster international collaboration. These summits have produced declarations, such as the Bletchley Declaration and the Seoul Declaration, where dozens of countries and leading companies like OpenAI, Anthropic, and Google DeepMind have pledged to work together on ensuring AI is developed and deployed safely. Even fierce competitors have stood together at forums like the United Nations, calling for global standards and oversight, acknowledging that no single company or country can manage the technology alone. This public alignment is a significant step, moving the debate from niche academic circles to the center of global policymaking.
The Chasm Between Words and Rules
Despite the apparent harmony, these commitments are almost entirely voluntary. They represent a shared direction of travel, not a set of enforceable rules. There is a vast difference between a company promising to conduct safety testing and a legally binding requirement that mandates it, complete with penalties for non-compliance. Critics point out that without enforcement mechanisms, these pledges are little more than good-faith promises from an industry known for its intense competitiveness. The United Nations' human rights chief recently stated that voluntary self-regulation is “nowhere near sufficient” to address the potential harms of AI. The current landscape is a patchwork of principles and statements of intent, not a cohesive global framework. This gap is where the real challenge lies: turning goodwill into governance.
Why a Binding Treaty Is So Hard
Several major obstacles prevent the leap from voluntary pledges to a binding international treaty. First is geopolitical competition. The world’s three major AI power centers—the United States, China, and the European Union—have fundamentally different approaches to regulation. The EU prioritizes fundamental rights with its comprehensive AI Act, which takes a risk-based approach. The U.S. has favored a more innovation-first, sector-specific model to maintain its competitive edge, though that is now being debated. China, meanwhile, uses a state-directed approach focused on both development and control. These competing philosophies make a single, universal rulebook incredibly difficult to negotiate. Unlike nuclear materials, which rely on scarce physical resources, AI is fundamentally code and data, making it much harder to monitor and control on a global scale.
The Race for Competitive Advantage
Beyond geopolitics, corporate competition is a powerful force working against binding agreements. The race to develop the most capable AI models creates immense pressure to move quickly. Strict, enforceable safety protocols could be seen as a handbrake on innovation, potentially allowing a less constrained rival to pull ahead. While CEOs publicly call for regulation, their companies are simultaneously investing billions to outpace one another. This dynamic creates a classic dilemma: everyone would be safer if everyone slowed down, but no one wants to be the first to do so. This is why many current agreements focus on post-development transparency, like sharing safety incident reports, rather than pre-deployment limits that might slow the race itself. The industry's logic often favors moving fast and managing problems as they arise, a stark contrast to the precautionary principle that a binding treaty would likely entail.
