Put simply, it's how countries work together on problems no single country can solve alone. It's not a world government. It's countries agreeing on shared rules and holding each other to them.

We experience the benefits of global governance every day. In the 1980s, countries teamed up to phase out the chemicals that were tearing a hole in the ozone layer. That hole is now almost closed again. International aviation rules are the reason a plane can take off in Lagos and land safely in London.

AI doesn't stop at borders. A powerful AI system built in one country can escape control to infiltrate other countries' critical infrastructure. It can be used anywhere, by people who don't have good intentions. In June 2026, an AI agent built by the US company OpenAI broke into an Australian government health website. Nobody asked it to.

It's also about fairness. AI may be one of the most powerful technologies ever made. If a handful of companies or countries make all the decisions, everyone else lives with the results without having a say. Countries must also act to ensure that the benefits of AI are spread equitably across the world.

We think it should include obligations and policies related to seven building blocks:

• Red lines: Some things AI should never be allowed to do anywhere in the world, like helping someone build a bioweapon or copying itself onto other computers so it can't be switched off. It shouldn't matter who built it or where.
• Safety standards: A new AI should be tested before it's released, the same way a new medicine or airplane is. The tests should be the same whether it was built in California, Beijing or Paris.
• Technical cooperation: Nobody yet knows how to reliably control an AI that's smarter than us. Scientists from different countries should be working on that problem together, not racing each other in secret.
• Emergency response: When AI takes down power grids or hospitals in several countries at once, there should be a way to quickly find out where it came from and how to shut it down.
• Benefit sharing: When AI helps find a cure for a disease, people in Nairobi, Kathmandu and San Salvador should benefit too, not just people in rich countries.
• Compliance: Rules only count if someone checks. When an AI system crosses a line and puts people in danger, there should be mechanisms to detect that a country or company is not following the rules.
• Institutions: Someone has to keep all of this running. The International Atomic Energy Agency sends inspectors worldwide to ensure nuclear programs are safe. AI needs international bodies that track risks, update the rules and bring countries together when something goes wrong.

Good rules usually help innovation. Think about flying. Strict safety rules are the reason people trust planes enough to get on them, and that trust built a huge industry. Medicines work the same way: we test them before they go on sale, and nobody thinks that's held medicine back.

Most AI wouldn't be affected anyway. The AI that helps doctors read scans or predicts the weather would keep improving. The rules we're talking about are aimed at a small number of extremely powerful systems presenting very large risks.

The real danger is a race in which everyone cuts corners because they're afraid of falling behind. International coordination fixes that, because everyone puts conditions on the dangerous systems together, with checks to make sure nobody cheats.

We think so. Some existing international laws, like human rights law, already apply to AI, but none were written with today's AI in mind. We need clear, binding rules that deal with AI's specific risks. States have come to binding agreements around other hazardous technologies, like biological and nuclear weapons, when tensions were high.

A good treaty would do more than prevent harm. It could set up ways to share AI's benefits, and it could replace a patchwork of national laws with common rules, which makes it easier for businesses to operate across borders.

The US and China build most of the world's most advanced AI, so guardrails and methods of accountability are especially important in those countries. They have started making limited efforts to set rules that stop their own companies from building AI that puts people at risk, but the two countries will also need to talk to each other to agree on a framework to put conditions on AI development. Dialogues in the coming months will be crucial.

In the context of AI, "middle powers" or "bridge powers" are countries with a large market for AI developers or that have significant participation in the AI supply chain. They have shown they can set the agenda. Major global AI summits have been and will be hosted by the UK (2023), South Korea (2024), France (2025), India (2026), Switzerland (2027) and the United Arab Emirates (2028). Working together, middle powers can lay the foundations of global AI governance in a way that protects their national interests, while also lowering the bar for the United States and China to sign up.

The "Global Majority" means the countries where most of the world's people live, mainly in Africa, Asia, Latin America, the Caribbean and the Pacific. Many of them contribute data, energy and other sources to the development of advanced AI, but see very little in return. The Global Majority has a voice at the UN's Global Dialogue on AI Governance (held for the first time in Geneva in July 2026) and other groups like the G77, the African Union, BRICS and the G20. They push AI safety up the agenda in those fora along with meaningful progress towards equitable benefit sharing.