Generated by Codex with GPT 5.6 Sol XHigh

This summary covers The Economist’s July 4th, 2026 Leaders article listed in the contents as The government and AI and published under the headline Rules for supermodels.

Frontier artificial intelligence is becoming too capable to release casually and too useful to lock away indefinitely. The article argues that governments need a durable system for managing that tension before a model enables a major cyberattack, helps a bioterrorist or causes some other disaster. America has begun regulating releases, but its improvisational approach shows why clearer institutions are urgently needed.

The immediate example is the Trump administration’s treatment of Anthropic’s Fable and Mythos models. Officials first restricted Fable for non-Americans after concerns about its safeguards, prompting Anthropic to block access more broadly because it could not reliably separate users by nationality. The Commerce Department then lifted the restriction after Anthropic adjusted those protections. Alongside pressure on OpenAI over its Sol model, the episode made policy look improvised, politically influenced and tilted toward American users.

Why permanent containment will fail

Licensing frontier models is unlikely to disappear, because their strongest capabilities create genuine security risks. Yet an indefinite block would be both unworkable and harmful. Chinese labs are only months behind the American frontier, and many of their models are open-weight, allowing users to run and modify them. Chip shortages and tighter controls on model distillation may slow Chinese firms, but probably only temporarily. Even a broad American ban would leave them a vast domestic market.

Restrictions would also hurt American researchers and companies that use inexpensive, adaptable Chinese models. More broadly, the same intelligence that makes advanced systems dangerous also makes them valuable. If regulators suppress several generations of progress and then release them at once, society loses the chance to adapt gradually. A small group with privileged access gains unusual power, while everyone else faces a sudden and potentially chaotic leap in capability.

Separate political choices from technical judgments

The article supports an emerging practice: evaluate a frontier model before release, then introduce it gradually through trusted institutions. But that process should be formal rather than negotiated case by case between politicians and AI companies.

Elected leaders should decide the acceptable level of risk because that is a public choice. They should not personally manage model tests or bargain over individual releases. Technical evaluation should instead sit with a specialist structure that combines expertise from government institutes, AI labs, independent researchers and allied regulators. Industries such as finance and electricity offer a useful pattern: expert bodies conduct oversight within goals and limits set by government.

International co-operation would make this system stronger, but America may resist sharing control over technology that gives it immense strategic leverage. The article therefore advises other countries to build bargaining power and resilience of their own. They need credible AI sectors, workable regulations and contingency plans that let businesses move to non-American models running in non-American data centres if export controls suddenly cut access.

The conclusion is not that allies should trust China instead. Dependence on America remains far safer. But relying on any single government for access to frontier intelligence is now a strategic vulnerability. Sensible AI regulation must reduce catastrophic risk without turning model access into an arbitrary political weapon.