If You're Wondering What Can Be Done About AI, Start Here
Updated: Sep 29
A response to Bill Gates's call for an international AI framework, and a practical way for nations and AI companies to reach one
Millions of us now use AI every day. Many of us have started asking a harder question: is anyone steering this, and what could we actually do about it?
On August 26, Bill Gates gave a blunt answer to the first part. His verdict was that nobody has shown up. He called for new national institutions able to coordinate AI policy across security, employment, education, taxation and health, alongside an international body drawing on elements of nuclear inspections and international aviation regulation. He has since been sharper still, saying the AI industry cannot regulate itself, calling on the US government to lead, and pressing for binding international rules and mandatory testing. He said he is baffled that industry and the White House are relying on voluntary safeguards for what he considers the most dangerous tool ever built.
He also set a condition. If a credible plan existed for slowing AI globally, he wrote, he would likely support it, but he does not expect one, because the geopolitical environment makes coordinated action exceptionally difficult.
This post is about how that condition could be met. We are not claiming to be the plan Gates would sign. We are describing the mechanism by which such a plan could be agreed at all, which is the part that has been missing.
Why hasn't anyone fixed this already?
Every major government regulating frontier AI is doing so cautiously, and the reason has little to do with will. The first country to impose serious obligations bears the cost alone, while the development it constrains simply moves elsewhere. Each nation, and each AI company, fears being disadvantaged if it adopts costly safety measures while competitors do not. That is not timidity. It is a rational response to a genuine collective action problem, and good intentions alone won't break it.
Gates put the trap plainly. Most people he talks to tell him that if everybody else slowed down, maybe they would too.
A second obstacle is less obvious. Even when an agreement exists that would leave every party better off, negotiations over many competing interests often fail to find it. Bargaining issue by issue, with each side defending its own draft, tends to miss the trades that would make a deal work.
Why can't the AI companies just agree among themselves?
Because, as Gates says, an industry cannot regulate itself, and because the attempt is now running into the law. Four paying subscribers filed a proposed class action in the Northern District of California on September 18, alleging that Anthropic, OpenAI, SpaceXAI and Google agreed to slow how fast they improve competing products, in breach of Section 1 of the Sherman Act. The allegations are unproven, and a filed complaint is not a finding. But the complaint is instructive for what it does not attack: the plaintiffs expressly say they are not challenging any company's own decisions about safety, testing, or its own pace. What they target is coordination among competitors.
That is the distinction our proposal turns on. Safeguards agreed by governments and written into national law are not an arrangement between rivals. They are the ordinary way societies set rules for an industry, and they are what Gates is asking for.
What is the solution to the AI coordination problem?
Our proposal combines two tools, one for each obstacle.
The first is the Simultaneous Policy mechanism, known as Simpol. Participating governments commit to implement agreed measures together, on the same date. No one accepts a first-mover disadvantage, and purely domestic policy remains a national matter. Simpol has spent more than two decades building this mechanism and applying it to tax cooperation and climate policy, and has built a base of legislators who have committed in principle to simultaneous implementation.
The second is Smartsettle Infinity, a multi-issue negotiation system. It places the whole package into a single negotiating framework and elicits each party's priorities privately. It then searches for combinations that leave every party better off than the status quo, using a criterion called Maximize the Minimum Gain, which raises the least-satisfied party as far as possible. This lets a party concede something it barely values in exchange for something it values greatly, and trades that haggling never reaches become visible.
In short, one tool removes the fear of going first by ensuring that nobody does, and the other finds the agreement worth moving together for.
What would a global AI safety agreement actually contain?
Our working document sets out a candidate package built around a levy on frontier AI compute. Compute is the one part of AI development that can be observed from outside. Training a frontier model requires physically identifiable chips, data centres and electricity, which can be metered and audited without seeing a line of code.
The levy would fund two things. The first is independent verification of safety claims, because today almost every framework relies on companies reporting on themselves. The second is capacity-building and affordable AI access for countries that do not host frontier development. For scale: the main existing industry fund for safety work has raised just over $10 million. By contrast, independent analysts estimate AI infrastructure spending by the largest US cloud providers at $550 to $800 billion in 2026. Even a 0.1% levy on that base would be 55 to 80 times larger than today's fund.
This is a different instrument from the token tax Gates has floated, which would fall on AI usage and help fund support for displaced workers. Ours falls on training compute and funds verification. The two are complementary rather than competing, and a mature package could carry both.
What each party gets matters as much as what it pays. The United States and China each want verified assurance that the other is subject to the same testing regime, and neither can get that by acting alone. AI companies could receive mutual recognition of audits across jurisdictions, or a cap on liability in a market where insurers are withdrawing cover. Smaller nations gain access and a seat in governance.
Whether the instrument should go further is itself on the table. The options governments would weigh run from funding verification alone to empowering a recommendation to pause development in extreme cases. Treating the purpose of the instrument as something to be negotiated, rather than assumed, is how you find out which version every indispensable participant can accept.
We are explicit that every figure here is a starting point for negotiation, not a conclusion. The proposal does not ask anyone to accept our numbers. It proposes the process by which the numbers get determined.
Has this approach worked before?
In a simulated global climate negotiation, six stakeholder groups used Smartsettle to reach an optimized consensus. That consensus left every bloc better off than the status quo, which the consensus procedure then in use could not find. The results were published by the Toda Peace Institute as Policy Brief No. 287 in March 2026, and you can watch the 20-minute video here. The conclusion was that the "multipolar trap" in climate comes from the negotiating procedure, not from the underlying interests. We think the same question is worth asking of AI.
What can I do about AI risk?
If you work inside frontier AI, we most need you to test the open questions in our working document (downloadable below), because they cannot be answered from outside. If you are a legislator, we ask for support in principle for the Simultaneous Policy pledge, which commits you to a process, not to any specific policy. If you represent a foundation or an institution with standing in multilateral affairs, we invite you to support or host a simulation that demonstrates what we are proposing. And if you are simply someone who uses AI and wants it to go well, join the Simpol campaign and share this with someone who can do one of those things.
What happens next?
A bounded demonstration is underway. Five government groups are being modelled as the parties: the United States, China, the European Union, the United Kingdom, and a coalition of countries hosting no frontier development. The frontier laboratories and the cloud and chip providers who would collect the levy will be modelled as interests, so that their real preferences shape the framework and its ranges without giving any company a seat alongside governments. Preferences will be modelled from public positions and labelled as such. The results will be published whatever they show, including a finding that no package works for some parties.
What comes after the five-party demonstration depends on how it goes and on the discussion it prompts. In the first run, each government's preferences include its own estimate of what its industry will bear. A second run would add the frontier laboratories and the compute providers as parties in their own right, and comparing the two would tell us something no argument can: whether governments can represent their industries faithfully, or whether the constraint only becomes visible when the firms speak for themselves.
Our full proposal, which you can download below, sets out that wider version — the one illustrated by the schematic above, with all six groups at the table.
Correspondence is welcome at info@smartsettle.com or through simpol.org.
Comments