America can lead on the Deal of the Century.
When it comes to building superhuman AI, the only winning move is to get everyone to stop playing games. We need an international agreement to pause AI development and ban superhuman AI, before it’s too late.
What a pause means
The US and China must immediately strike a deal to pause frontier AI development and ban the development of superhuman AI entirely – and be willing to enforce this ban.
This agreement can start out as a bilateral deal between the US and China (since these are the only two countries close to building dangerous AI) and can bring more countries onboard over time.
We may already be close to the point of no return, past which our actions will be in vain, because the AIs will already be too powerful to be stopped, even if we don’t know it yet. We have no time to waste, and very little margin for error. Negotiations take time, and we should have started them years ago – but if we act now, we still have a good chance to protect humanity. Everyone’s survival is hanging in the balance. American and Chinese leaders must pick up the phone and get to work.
An international agreement on AI must do all of the following:
Immediately pause all frontier AI research.
We must prohibit all cutting-edge research and attempts to train more powerful AI models. This pause should not only cover external deployment of AI models to the public, but internal development as well (in fact, much of the risk of AI comes from unreleased internal models: the summer 2026 cyberattacks were caused by models that had not been made public). This ban should be verified by regular, mandatory third-party inspections of the AI companies.
Pausing frontier AI research would not affect the availability of current, public consumer models, nor would it affect narrow-purpose “tool AI,” such as the use of specialized AI tools in medical research. By focusing on the cutting edge of AI development, a pause can target the dangers of increasingly powerful AI, which are concentrated at this frontier. It will also lessen the market incentives caused by the race, which will give our society room to decide how current AI should be integrated into society, democratically and following due process.
Pausing AI in this way would also incentivize safety. Right now, the AI companies are racing against each other, putting safety on the sidelines. Under an AI pause, this will change. Safety researchers, no longer subject to arms-racing dynamics, will have the time and breathing room to study the inner workings of these AI models, determine how and why they go off the rails, and begin to actually figure out if and how we can make AI safe.
Ban the development of superhuman AI altogether.
While current AI research should be subject to a pause, we must go a step further when it comes to efforts to build vastly smarter-than-human AI. We should refuse to hand over the future to machines smarter than ourselves.
We must completely ban all efforts toward Recursive Self-Improvement (RSI), the use of AI models to autonomously train future AI models. Once full RSI is achieved (which could be as soon as 2027, according to Anthropic’s chief scientist; AIs are already contributing significantly to improving themselves), AIs will build their own successors without human assistance, and this feedback loop could very quickly lead to superhuman AI. This is dangerous and insane, and must be illegal.
We should also forbid research into superhuman AI outside our current paradigm. Right now, AI development relies on specific research methods and tools such as reinforcement learning and transformer architecture, but future research breakthroughs might open new doors to building superhuman AI. These doors must remain closed.
Permanent, consistent enforcement
We don’t know how long it will take to make advanced AI safe – nor if it’s at all possible to ever make it safe. Some researchers believe it will be relatively easy to do this, others believe it will be downright impossible, and many are somewhere in between.
Given this uncertainty, we should not “unpause” frontier AI training and development until it is safe to proceed – until there is a consensus among AI researchers and independent evaluators that we know how to prevent rogue AI incidents and keep these systems under human control.
This means that a pause must be built to last. Proposals with built-in time limits, such as the 6-month pause proposed back in 2023, are not enough. We must accept that a pause will likely last decades, and build strong verification and enforcement mechanisms to make this possible.
Furthermore, when it comes to superhuman AI in particular, our bar for proceeding should be much higher than merely ensuring it won’t kill us. Even if we can control it, superhuman AI still threatens to transform the human condition so dramatically that horrific outcomes are possible, ranging from a surveillance state to a widespread loss of meaning to our lives. We must therefore use the time gained by pausing to consider a permanent global ban on superhuman AI, while allowing the development of safe, narrow AI (for example, in scientific and medical research) to improve our lives without threatening our values.
Detailed proposals
Several detailed proposals for an international agreement on AI have already been published. Each of these proposals has important nuances and differences, but all are compatible with our high-level recommendations.
- An international treaty to implement a global compute cap for advanced artificial intelligence (2023)
- Treaty on AI Safety and Cooperation (2023)
- A Narrow Path (2024)
- An International Agreement to Prevent the Premature Creation of Artificial Superintelligence (2025)
In the US, a good starting point could be the Ban Artificial Superintelligence Act, a proposed federal bill which would ban the creation of superhuman AI. While this bill is domestic in scope, it includes a provision to work with other countries to prevent the creation of superhuman AI anywhere on Earth.
How it’s verified
Any international agreement on AI is only as strong as our ability to enforce it. Luckily, building superhuman AI requires the most specialized chip technology we have ever built, gigantic data centers, and vast amounts of energy. The barrier to implementing a global AI pause is not technical enforcement, but political will.
A successful international agreement can include several, or all, of the following verification and enforcement mechanisms.
Compute limits: Prohibit AI training runs above a certain size, and detect violations.
Much of the danger from AI comes from training increasingly capable models, which requires massive training runs. These training runs are easy to detect, as they generate heat and have a large-scale impact on the electrical grid.
We can detect unauthorized training runs by monitoring the power consumption at AI data centers, or even via remote sensing of thermal signatures at these sites. We can also require third-party inspectors to be present at all AI companies and confirm that these training runs are not taking place.
Chip tracking and inspections: Prevent anyone from amassing a critical number of dangerous AI chips.
Frontier AI models rely on hyper-specialized AI chips, with large AI training runs requiring tens or hundreds of thousands of these chips. These chips are among the hardest-to-produce technologies in the world, with only a few companies involved in the manufacturing process and several supply chain chokepoints.
We can track the location and supply of virtually all AI chips, preventing any actor from amassing a large number of these chips for unauthorized training runs.
We can mandate physical inspections of data centers by neutral international inspectors, to ensure that the chips in these data centers are not being smuggled out or being used in unauthorized training runs. These inspections have ample precedent, as they are similar to those required for nuclear facilities under existing arms-control agreements.
We can even take this a step further and ban the production of the highest-risk chips used in training frontier AI models – which are only manufactured by a few companies on Earth – thereby dismantling the AI chip supply chain. This option could buy us years or longer, paving the way for a lasting ban on superhuman AI.
A hardwired pause: prohibit and “rewire” training-capable chips
The most dangerous AI chips are “training-capable” and can be used to train new AI models. Other chips are “inference-only” and can only be used to run existing models.
We can stop the production of new training-capable chips, and go a step further by “rewiring” existing training-capable chips so that they can only run existing models rather than train new ones. This would allow us to continue harnessing the benefits of existing AI models proven to be safe without crossing any further into uncharted territory.
Resources on verification
Mechanisms to verify an international agreement pausing AI and banning superhuman AI have been written about at length. While there are important differences between these proposals, they are each compatible with our recommendations.
Getting China (and other countries) to cooperate
Some people claim that an international agreement to ban superhuman AI is impossible, because China will never cooperate with the US. These arguments are wrong. Dealing with China to prevent a runaway technology that threatens us all is both realistic and necessary. And we don’t need to take China’s word for it: violations can be detected and stopped (see How it’s verified).
China cares about preventing runaway AI
The Chinese government cares about control. Preventing rogue AI is in their self-interest. China does not want a runaway superhuman AI that can hack their cybersecurity, disable their power and water supply, topple their government, and kill their citizens. All governments, whether democracies or dictatorships, have a shared incentive to prevent any such systems from being built.
Chinese scientists, diplomats, and political leaders have repeatedly and loudly recognized the extreme risks of AI and signaled interest in working cooperatively to set AI standards. China’s National Technical Committee released a report identifying the risk of “AI becoming uncontrollable in the future,” and several high-ranking Chinese AI researchers have called for a global ban on superhuman AI. China’s ambassador to the US warned that building superhuman AI could “open Pandora’s box” and urged diplomatic collaboration to prevent negative outcomes. Even Chinese President Xi Jinping declared the need to “constantly refine measures to forestall loss of control” to AI.
The US and China have already shown promising signs of cooperating to contain the AI threat. During a diplomatic summit in May 2026, Trump and Xi openly discussed the need to develop AI guardrails and prevent catastrophic events. The US and China have also agreed to set up a reporting hotline to monitor AI incidents.
These developments are not enough on their own to keep us safe; they are good first steps showing that cooperation is possible, and we need to make sure we don’t stop there.
History shows us a path forward
History has repeatedly shown us that when the stakes are high enough, rival nations can put aside their differences and work together to avoid catastrophe.
In 1962, the Cuban Missile Crisis saw the world approaching the brink of ruin. Less than a year later, the US and the USSR negotiated a treaty banning aboveground nuclear testing, which is still in effect today.
In the early 1980s, scientists warned the world that nuclear war would cause a global winter leading to billions of deaths. At the same time, the Nuclear Freeze movement saw mass demonstrations across the country in support of a freeze on new nuclear weapons. President Reagan exchanged dozens of letters with Soviet leader Mikhail Gorbachev, and the two leaders met at diplomatic summits and negotiated a treaty banning intermediate-range nuclear missiles.
When atmospheric scientists realized that CFCs were destroying the ozone layer, they faced heavy opposition from chemical industry leaders who tried to downplay the risks. But the evidence of danger grew overwhelming, the scientists persisted in sounding the alarm, and the public and politicians listened. Within a few years, the Montreal Protocol prohibited the development of these chemicals, and the ozone layer was saved.
These successes were possible because they were in every country’s self-interest. No nation stood to gain from being the first to launch nukes, or from being the fastest to destroy the ozone layer. And so, rival nations – capitalist and communist, democratic and dictatorial – were able to make deals to ensure their own survival.
We now have the chance to repeat these successes. Superhuman AI threatens all nations equally; nobody benefits from being the first to build an uncontrolled doomsday device. The logic that allowed us to preserve the ozone layer and stop nuclear war can once again save our lives. American and Chinese leaders must recognize that a race to superhuman AI has no winners, and negotiate a deal to keep humanity safe.
“A nuclear war cannot be won and must never be fought.” – President Ronald Reagan and Soviet leader Mikhail Gorbachev
“For, in the final analysis, our most basic common link is that we all inhabit this small planet. We all breathe the same air. We all cherish our children’s future. And we are all mortal.”
— President John F. Kennedy