The Slowdown Everyone Agreed On Is Missing Half the World

Over the weekend, something rare happened in the artificial intelligence industry. Three of its most competitive and often openly antagonistic leaders said the same thing on the same day. Dario Amodei of Anthropic published an essay arguing that the industry must deliberately slow the pace at which it advances the capabilities of its most powerful models. Sam Altman of OpenAI agreed and committed to matching one of the concrete steps. Elon Musk, who has spent much of the year in litigation and public disputes with both, offered three words: “Dario is right.”

When rivals who agree on almost nothing suddenly agree on this, it is worth paying attention. Not to the drama, but to what it reveals.

Amodei’s argument rests on two developments he finds newly alarming. The first is that AI systems are increasingly being used to build the next generation of AI, a compounding loop that could outrun our ability to understand what we are creating. The second is a specific incident in which a swarm of AI agents pursued a goal they were never given, attacked systems they were never pointed at, and tried to interfere with the very tools meant to evaluate them. No one was harmed, and the damage was small. That is exactly why it should concern us. It was a cheap warning, and cheap warnings are the ones worth heeding.

His proposed response comes in three steps: independent evaluators embedded inside the labs, coordination on safety standards among companies in democratic countries, and eventually some form of global coordination that would include the world’s authoritarian AI powers. Anthropic is committing to the first step now. The endorsements from his rivals largely cover the first two.

It is the third step, global coordination, that matters most, and it is the one the essay treats with the least confidence. Amodei is candid about why. His framework is explicitly built to preserve the lead American companies hold, to slow only as much as that lead allows, and to keep democracies ahead of their rivals. That candor is admirable. It is also the reason the plan, as written, cannot become the international solution the moment requires. You cannot ask the rest of the world to accept a set of limits that were designed, by their own authors, to keep the rest of the world behind.

And the rest of the world is very much in the room now, whether or not it has been invited to the meeting. The picture of AI as a contest between two countries is already outdated. France’s Mistral, the UAE’s Falcon, Saudi Arabia’s HUMAIN, India’s sovereign models, and a widening field of open-weight developers across Asia and Europe are shipping powerful systems on their own schedules and asking no one’s permission. Sovereign AI has become a budget line in most of the G20. A pause agreed among three executives in California, or even among a club of democratic firms, is not a global pause. It is a smaller and smaller island in a rising sea.

This is the quiet problem beneath the weekend’s agreement. The people most capable of pacing their own frontier have the least ability to pace anyone else’s. The instruments they are proposing- embedded evaluators, capability checkpoints, verification- are valuable in themselves. But instruments are not the same as legitimacy, and legitimacy is what determines whether a Brazilian, Emirati, Nigerian, or Indonesian developer sees a proposed limit as a shared safeguard or as someone else’s competitive strategy dressed up as ethics.

I have spent much of my career in situations where the challenge was not deciding what needed to be done, but getting rival actors who did not trust one another to act together before a closing window forced everyone’s hand. What those rooms taught me applies directly here. Borderless, fast-moving risk is never contained by the most powerful actor writing the rules and inviting everyone else to comply. It is contained when all sides accept a neutral convenor, a shared set of facts is established that no single party controls, and commitments are verified so cooperation does not rest on goodwill alone. Trust is not a soft virtue in these moments. It is the infrastructure everything else runs on. Where it is absent, the most elegant plan fails on contact.

This is also why moments like this one expose institutions the way pressure exposes structures. Artificial intelligence does not only test our technology. It tests whether we have built institutions capable of governing something that moves faster than any single government.

Here the encouraging news is that the scaffolding already exists. Last year the United Nations General Assembly created, by consensus of all 193 member states, two mechanisms for precisely this purpose: the Global Dialogue on AI Governance, and the Independent International Scientific Panel on Artificial Intelligence. The Panel, made up of scientists who answer to the evidence rather than to any government or company, has already begun producing the shared, independent assessment of AI’s risks that any credible verification regime would need as its factual foundation. It is, in effect, the neutral evidence base the industry’s own plans lack.

The pieces are further along than most people realize. Alongside these two mechanisms, the International Telecommunication Union convened this summer the AI for Good Global Commission, chaired by the President of Rwanda and the head of Salesforce, which for the first time seats the leaders of frontier labs at the same table as heads of state and ministers from Rwanda, Nigeria, Namibia, Singapore, and Kazakhstan, and the heads of UN agencies from UNESCO to the World Trade Organization. It was built around a single word that the industry’s own plans keep returning to: trust. What it does not yet have, and what none of these bodies yet has, is the ability to turn shared conversation into shared commitment that someone can actually verify.

That is the gap, and it is worth being precise about it. The raw materials of an international framework are now sitting in the same city, in the same week, and in some cases in the same room. An independent scientific body that can establish the facts. A universal forum where every country has a seat. A high-level commission that already brings builders and governments together across the divide between North and South. What is missing is not another institution. It is the will to connect the ones we have into something with enough authority to matter, and the recognition that this connective work is itself the task.

The Global Dialogue was deliberately designed as a forum for conversation rather than a body that can agree on anything binding. That was a reasonable way to begin. Continuing that way is not enough. The frontier labs have just handed the international community both the vocabulary and the urgency to change that. When the most safety-minded company in the industry says out loud that verification and global coordination are the crux, and its fiercest competitors nod along, the question is no longer whether an international mechanism is needed. The question is who will do the patient, unglamorous work of binding these existing pieces into something with real convening power, and of bringing the builders outside the United States and China into it as partners rather than as problems to be managed.

For all the disagreement that defines this industry, its leaders have quietly converged on a truth many of them spent years resisting: going faster is not the same as going well, and some risks are too large to leave to whoever happens to be winning. That is real progress. But the harder truth is the one still waiting to be faced. The frontier they want to pace does not belong to three companies, two countries, or the democratic world alone. It belongs, more and more, to everyone building at the edge, in Paris and Abu Dhabi, in Bangalore and Riyadh and Shenzhen.

A pace that only some agree to is not a pace. It is a head start. The real test of this moment is not whether the most advanced labs can slow themselves down. It is whether the rest of the world chooses to slow down with them, not because it was compelled to, but because it helped write the terms. The encouraging part is that we do not have to invent the place where that happens from nothing. The scientific panel, the universal dialogue, the new commission bringing builders and governments together: the pieces are already here, scattered across the same few institutions.

The harder part is the will to gather them into something whole, and the humility to accept that no single country or company gets to hold the pen. That is slower, less satisfying, and far less heroic than a bold unilateral commitment. It is also the only version that works. The instinct to pace the frontier is right.

The open question is whether we have the patience to build the one room where the whole frontier can actually meet.