Asha Jadeja is a Silicon Valley venture capitalist.
The most dramatic global competition of our era isn’t simply the race between the United States and China to dominate artificial intelligence. It is the race to develop safeguards before AI capabilities outrun them.
The U.S. cannot and should not abandon the AI race. Nor should it allow China to set the rules for the world’s most important technology. But Washington and Beijing need to recognize that they face shared risks and vulnerabilities, and act accordingly.
We have done this before — and it worked.
During the Cold War, the U.S. and the Soviet Union were adversaries armed with weapons capable of destroying one another. They didn’t need to trust each other to understand that accidental nuclear war would be catastrophic for both. Instead, they established hotlines, negotiated arms-control agreements and created protocols intended to prevent miscalculation.
Set up in 1963 in the wake of the Cuban missile crisis — the closest call of the Cold War — the so-called red telephone wasn’t in the White House. Nor was it red. The system relied on a teleprinter in the Pentagon that could directly send and receive messages from the Kremlin.
President Richard M. Nixon used it to communicate with Soviet leader Leonid Brezhnev during such perilous moments as the 1973 Yom Kippur War.
AI now requires a 21st-century version of that arrangement — and the encouraging news is that dialogue has begun. American and Chinese security experts have proposed safeguards that include keeping AI out of nuclear launch decisions, maintaining meaningful human control over critical military systems and establishing a bilateral hotline for AI-related crises.
It’s imperative that, as the two global AI superpowers, Washington and Beijing have ongoing direct communications at the highest level, much as Washington and Moscow did.
But the two governments now need to go further, despite the obvious challenges and mutual distrust.
The U.S. and China should create a narrow agreement governing only the most powerful frontier AI systems. It should require rigorous testing before deployment for capabilities such as autonomous cyberattacks, biological-weapons assistance and attempts to evade human control. The two countries should agree never to delegate nuclear-launch authority to AI and should establish rapid notification procedures when developers discover capabilities that could pose a catastrophic threat.
In addition, the AI link between the two countries should go to the very top of any administration. If messages become mired in State Department or Pentagon bureaucracy, the system loses the very quality that makes AI a potential threat: speed.
China is not an ally any more than the Soviet Union was. That is why such an agreement shouldn’t require sharing models, intellectual property or national security secrets. Nor would it prevent either country from competing vigorously in commercial AI.
At the same time, an AI security agreement between the U.S. and China need not address other thorny bilateral topics, such as sensitive chip exports. To avoid any delay in making sure a dangerous AI scenario doesn’t become reality, those issues should be the focus of separate negotiations.
What would danger look like? The development of an AI-created bioweapon or unauthorized agentic lab infiltration would require an urgent response, with coordination at the highest level of both governments to contain it. A similar threat to electric grids or water supplies would also constitute a “red phone” moment.
Most important, we should recognize that artificial intelligence itself isn’t independent of human control. The most powerful systems depend on physical infrastructure: enormous data centers, electricity, cooling and immense computing capacity. These factors give governments and companies points of leverage.
Whatever system is developed, a bilateral AI arms control accord would also require close coordination between the White House and the heads of American frontier labs, including Anthropic, OpenAI and Google. Beijing, in turn, would have to partner with Chinese AI giants such as DeepSeek and Alibaba’s Qwen.
DeepSeek’s powerful open-weight releases are a concern that negotiators would have to tackle, along with ensuring that the incredible speed of American AI development does not quickly make proposed high-level safeguards obsolete.
President Donald Trump’s Sept. 29 AI summit at the White House with leading tech executives was a constructive first step. Now is the time to build on that momentum with his Chinese counterparts.
There is a tendency in Washington to view every discussion of AI safety through the lens of competition with China. Beijing likewise suspects that American safety proposals are merely an attempt to preserve U.S. technological advantage. Recent Chinese criticism of American executives’ proposals to slow some AI development demonstrates how quickly safety can become entangled with geopolitics.
That is precisely why a narrow bilateral framework is needed now, backed up by international monitors overseen by Washington as well as Beijing.
The U.S. should compete to lead the world in artificial intelligence, despite the addressable fears that now threaten to hold us back. It should build the infrastructure, attract the scientists and create the companies that will capture the enormous economic benefits that lie ahead.
The U.S. should race to develop AI. It should race even faster to make sure humanity remains in control of it.
The post The U.S. and China need an arms control pact — for AI appeared first on Washington Post.
