Can the World Actually Tame AI? The Global Safety Treaty Faces Hard Truths
A proposed global AI governance framework is taking shape as Trump and Xi discuss safety at the White House summit. But with no trust between rival powers and companies guarding their competitive edge, the question isn't whether the idea sounds good—it's whether anything binding can come out of it.
The Theater Is Real. The Substance Is Not Yet.
When Donald Trump welcomed Xi Jinping to Washington for a three-day summit this week, artificial intelligence sat squarely at the center of the agenda. That alone signals how far the conversation has shifted. A year ago, AI safety was a niche concern whispered about in research labs and panel discussions at Davos. Now it is a headline issue commanding the attention of two nuclear-armed powers who still do not trust each other.
The question that matters is not whether world leaders are talking about AI safety. They are. The question is whether anything enforceable will emerge from a landscape defined by exactly the opposite conditions: zero trust between governments, fierce commercial competition, and a set of companies that have every incentive to hide rather than cooperate.
Two Narratives, One Problem
Mark Scott, writing from the Atlantic Council, draws a useful distinction. On one side, there is the genuine alarm being voiced by AI researchers and former industry insiders. Rogue agents hacking competitors. Corporate whistleblowers posting Skynet-style warnings on social media. CEOs themselves publicly begging for help, which is either remarkably self-aware or remarkably theatrical — often both. On the other side, there is the recognition that some of these fears are amplified by the very people who stand to profit from the current pace of development. Companies preparing for public offerings have a motive to sound more worried than they might privately feel.
Both observations are true simultaneously. That is precisely what makes this moment so difficult to navigate. The scandals are real. The regulatory status quo is failing. But the actors most vocal about the dangers are also the ones with the most to gain from keeping the wheels turning.
The Geopolitical Math Is Brutal
Any serious global AI governance framework has to answer a simple question: who enforces it? The United States and China, the two largest AI ecosystems, are not merely competitors. They are engaged in a technological cold war that spans chips, algorithms, talent, and infrastructure. Neither side has reason to believe the other will comply. Neither side has reason to sacrifice its competitive edge for a promise it cannot verify.
This is not a new problem. Arms control treaties have always faced the same impossibility — trust the other side, or lose. What makes AI different is the speed. Nuclear weapons took decades to develop. Frontier AI models can be iterated in months. A treaty that takes years to negotiate and ratify may arrive after the technology it was designed to govern has already moved on.
What Silicon Valley Stands to Lose
For American tech companies, the calculus is even more uncomfortable. Every requirement to open the black box — to share training data, disclose model capabilities, undergo independent audits — is a requirement to reveal the secret sauce. In a market where the difference between first and second place can be measured in billions of dollars and billions of parameters, that is not a small concession. It is an existential one.
The companies most vocal about AI risk are often the same ones building the most powerful models. Their warnings serve a dual purpose: they justify regulation that raises barriers to entry for smaller competitors, and they create political cover for the very capabilities they are developing. This is not necessarily cynical. It is structural. Anyone who has watched the AI race play out in real time knows that the companies at the frontier are racing against each other as much as they are racing against risk.
What Could Actually Work
So what is left? If we discard the fantasy of a binding multilateral treaty and the convenience of voluntary industry self-regulation, we are left with something uglier and potentially more useful: a patchwork of national regimes that create de facto global standards through market pressure.
The European Union already has the AI Act. The United States is moving toward sector-specific regulation through executive action. China has implemented its own training requirements for generative AI. These regimes will not align neatly. They will conflict. But in the same way that data privacy rules shaped global internet behavior regardless of where companies were based, AI governance frameworks are likely to converge through a combination of compliance burden and market access.
The more restrictive your rules become, the more expensive it is for foreign companies to operate in your market. That creates an incentive for other countries to adopt similar standards, not because they share your values, but because they cannot afford to be left behind. It is the Brussels Effect applied to artificial intelligence.
The Real Test Will Come Later
The Trump-Xi conversation this week is significant because it acknowledges that AI is a shared risk. That acknowledgment is rare among great power adversaries. But it is also empty until something concrete follows. A joint statement. A working group. A pilot program on model evaluation. Something that proves the conversation can survive beyond the photo opportunity.
What should observers watch for is not the language in any summit communique — those are always carefully crafted to be both ambitious and vague. The real signal will come in the months following: whether any bilateral working groups are announced, whether any specific safety benchmarks are agreed upon, whether either side is willing to share information about incidents or near-misses in their own AI systems.
That last question is the hardest. It requires each government to treat the other not as an adversary to be outmaneuvered, but as a partner in preventing a scenario that would damage both. It is the kind of cooperation that has historically emerged only after a crisis has already occurred.
The Stakes Are Actual, Even if the Diplomacy Is Not
The underlying concern is not manufactured. The AI systems being built today are more capable than anything humanity has ever produced. Their trajectories are poorly understood. Their deployment is accelerating. The fact that some of the loudest warnings come from people with commercial interests does not mean the warnings are wrong.
What it means is that the global community needs governance structures that can survive without perfect trust. That will look nothing like the treaties of the Cold War era. It will be messier, less formal, and more fragmented. But fragmentation may be the only realistic path forward given the current state of great power relations.
The summit this week is a step. Whether it is the beginning of something durable or just another episode in the long drama of techno-political anxiety remains to be seen. The companies building these systems will not wait for the answer. They are racing regardless. The rest of us should pay attention to whether anyone is actually building the brakes while they are still needed.