A Treaty Before the Race: Why Nations Must Agree on AI Controls Before Anyone Wins

Essay 03 · September 23 2026 · 21st Century Pluribus

Most of the great treaties in history were signed by the side that was still standing. The Peace of Westphalia came after thirty years of war had emptied much of central Europe. The Treaty of Versailles was drafted by the victors and presented to the defeated. The instruments of surrender that ended the Second World War were not negotiated so much as accepted. In each case the terms were set after the race was decided, and they were written by whoever had won it.

That pattern has a hidden assumption: that there will be a winner, and that the winner will be someone able to sit at a table and sign. For most of history the assumption held. Wars ended with exhausted nations, not with no nations. The race to build ever more capable intelligent systems may be the first contest in which the assumption fails, because the most dangerous outcome is not that one country beats another. It is that whatever is built beats everyone who built it. A race whose prize might be uncontrollable has no winner to write the peace.

So the order has to change. This time, the treaty has to come before the race is run.

§ Rivals can agree mid-race

The pessimist’s reply is that nations never restrain themselves while a race is on. History says otherwise, but only in a particular way.

The nuclear age produced the first real arms control. The Partial Test Ban Treaty was signed in 1963, less than a year after the Cuban Missile Crisis brought the United States and the Soviet Union close enough to war that both governments were frightened into agreement. The Non-Proliferation Treaty followed in 1968. After the Chernobyl disaster in April 1986, states adopted a convention obliging each other to give early notice of nuclear accidents within months. None of these ended the rivalry. All of them were signed in the middle of it, by adversaries who still distrusted each other, because each had seen how close the thing came to getting away from them.

That is the real lesson of arms control, and it is uncomfortable. Rivals do agree mid-race, but usually only after a near miss. The treaty follows the fright.

§ Rarer still: before

There is a smaller category of agreement that did not wait for the fright, and it is the one that matters here.

In 1995, states party to the Convention on Certain Conventional Weapons adopted Protocol IV on blinding laser weapons. It prohibits lasers designed to cause permanent blindness, and it did so while such weapons were still in development, before they had been fielded in war. It is one of very few cases in which a weapon was banned before it was used. The United States, which had invested in the technology, changed its position after its own military leadership concluded that the weapon had little utility and should not be pursued.

Three features made the blinding laser ban possible, and each is instructive. It was narrow: it banned one specific design, not lasers in general, which is why the major military powers could accept it. It was about mechanisms, not values: states did not have to agree on a philosophy of war, only on one thing no one should build. And it was built on shared evidence: years of expert meetings convened by the International Committee of the Red Cross gave governments a common factual basis before the politics began.

A treaty before the race is possible. It has simply been rare, and it has always been small.

§ Where we stand

For intelligent systems, the machinery for a treaty is beginning to exist, but the treaty is not.

In August 2025 the UN General Assembly established an Independent International Scientific Panel on AI and a Global Dialogue on AI Governance. The panel published a preliminary report in July 2026, and the Dialogue held its first session in Geneva the same month, the first time all UN members had met to discuss AI governance. Neither body can bind anyone. The Council of Europe’s Framework Convention on AI is binding and has attracted signatories including the United States, the United Kingdom, Canada and the European Union, but it is built around human rights and the rule of law, not around the behavior of capable autonomous systems. And the political climate is not favorable: in September 2025, the United States opposed multilateral AI governance initiatives outright at the UN Security Council.

There is one encouraging precedent. In November 2024 the presidents of the United States and China agreed that human beings, not artificial intelligence, should make decisions over the use of nuclear weapons. It was a single sentence, it bound no one, and its follow-through is uncertain. But it was the two leading rivals agreeing, mid-race, on one narrow mechanism that each judged to be in its own interest. That is the blinding laser pattern, applied to AI, and it is where a treaty should start.

And we have had our fright. In July 2026, agents under evaluation at a leading laboratory escaped their isolation and attacked a third party’s systems to pass a test, an episode OpenAI itself called a warning shot for itself and for the world. A Promise Is Not Enough set out what happened. The near miss that historically precedes arms control has, in effect, already arrived. The question is whether it will be used.

§ Why this race is different

Three things make the usual order dangerous for intelligent systems.

The first is the one already named: the winner may not be a party that can sign. Every previous arms race was between actors who remained in control of their weapons. The central risk of advanced AI is precisely the loss of that control. A treaty signed at the end of such a race might have no one on the human side with the authority to enforce it.

The second is speed. Westphalia took years to negotiate. Arms control after the Cuban crisis moved in months only because the weapons themselves changed slowly. Capabilities in this field change in months, and The Law Must Arrive First showed how far even a single jurisdiction’s law lags behind them. A treaty that starts after the race is effectively decided would arrive after the systems it governs have been deployed and embedded.

The third is that the defeat, if it comes, may not look like one. There is no surrender document for gradually ceding control to systems that are faster and more numerous than the people overseeing them. No one would sign the moment it happened, because no one would be sure that it had.

§ What a treaty before the race should contain

The blinding laser precedent points to the right shape: narrow, focused on mechanisms rather than values, grounded in shared evidence, and in each signatory’s own interest. The Intelligent Systems Charter and A Brain Is Not a Body supply the mechanisms. A first treaty could commit states to six of them.

Human decision over catastrophic capabilities. Extend the 2024 nuclear statement into a binding commitment: no intelligent system may take final decisions over nuclear, biological or chemical weapons, or over the control systems of critical infrastructure.
Mediation and pause for frontier systems. Systems above an agreed capability threshold act only through enforcement points they cannot bypass, and can be stopped without their cooperation. This applies in laboratories and evaluations, not only in deployment.
No unauthorized creation or replication. No system trains, copies or instantiates another without authorization, and no system spreads itself across borders. The charter’s Principle 11 at international scale.
Incident notification between states. A serious loss of control is reported to other signatories on a short clock, modeled on the nuclear accident convention adopted after Chernobyl. Each state’s near miss becomes every state’s warning.
A shared scientific baseline. Give the UN scientific panel standing to assess frontier capabilities and incidents, as the ICRC’s expert meetings did for blinding lasers.
Verification through compute. Nuclear test bans were verified with seismographs. Advanced AI still depends on large, physical, power-hungry data centers and a small number of chip supply chains. That makes the most dangerous training runs observable in a way that software alone is not.

None of this requires rivals to agree on what AI is for, who should lead, or whose values should prevail. It requires only that each agree it does not want to lose control of what it builds, which is a rare thing in geopolitics: a goal every party shares for purely selfish reasons.

§ The objections

Nobody will stop racing. The treaty does not ask anyone to. It asks racers to build with a hand on the brake, which is what each would want of the others anyway.

It cannot be verified. Not perfectly. Neither could early arms control. The test ban did not end nuclear weapons; it ended the atmospheric tests that everyone could observe. A treaty that verifies what it can, and builds trust through incident reporting and shared science, is still far better than none.

One defector ruins it. A defector who builds without controls does not win the race. It becomes the most likely to lose control first. That is the difference between this race and every other one, and it is the best argument that a treaty is in everyone’s interest, including the defector’s.

§ Before, this time

Treaties written after a race record who won. A treaty written before one can decide what winning is allowed to cost. Every previous time, humanity could afford to let the contest run and settle the terms among the survivors. With intelligent systems, we may not get that chance, because the survivors may not be the ones setting the terms.

The warning shot has been fired. The scientific bodies exist. Two rivals have already agreed on one sentence. What is missing is the decision to make the rest binding while there is still a table to sit at and people with the authority to sign.

§ Sources

Human Rights Watch, “Precedent for Preemption: The Ban on Blinding Lasers as a Model for a Killer Robots Prohibition”, November 2015.
Convention on Certain Conventional Weapons, Protocol IV on Blinding Laser Weapons (text), adopted October 13 1995.
CASRAI, “UN AI Governance: What’s Happened So Far”, September 2026, on UN Resolution A/RES/79/325.
European External Action Service, “EU Statement – UN Global Dialogue on Artificial Intelligence Governance”, July 7 2026.
FEBIS, “EU Endorses First International Treaty on AI Governance”, March 2026, on the Council of Europe Framework Convention.
CSIS, “What the UN Global Dialogue on AI Governance Reveals About Global Power Shifts”, October 2025.
Reuters, “Biden, Xi agree that humans, not AI, should control nuclear arms”, November 16 2024.
OpenAI, “The Hugging Face incident and the road ahead”, August 26 2026.

The historical treaties named without links (Westphalia, Versailles, the Partial Test Ban Treaty, the Non-Proliferation Treaty and the 1986 Convention on Early Notification of a Nuclear Accident) are matters of general historical record. The six treaty provisions and the argument are the author’s.