Generated by Codex with GPT-5

This summary covers The Economist’s June 20th, 2026 Briefing article listed in the contents under Access to AI and published under the headline A bouncer gone berserk.

A Ban That Shows Power And Weakness

The article argues that Donald Trump’s order forcing Anthropic to block foreigners from its most advanced Fable 5 and Mythos 5 models is less a coherent AI strategy than a dramatic proof of American control. Washington may have been trying to punish Anthropic after security concerns about its models, but the result was much larger: allies, companies and governments discovered that access to frontier AI could vanish overnight by political decree.

That makes AI look like earlier strategic technologies, but only up to a point. In the 1970s and 1980s America treated strong encryption like a munition and tried to restrict its export. Civil-liberties advocates and software’s natural spread eventually made that posture untenable. Nuclear technology offers another analogy: after the second world war, America cut Britain and other allies off from nuclear co-operation until Britain proved it could build its own bomb. Frontier AI sits awkwardly between the two. It may be dangerous enough to matter for cyber-attacks, weapons design and national security, but it is also software, talent, chips, cloud capacity and commercial service delivery all at once.

Allies Learn Their Dependence

The immediate diplomatic damage is the article’s sharpest point. The ban did not spare even the Five Eyes intelligence partners, and it locked out Britain’s AI Security Institute, one of the leading bodies for testing powerful models. Governments and big firms that had spent months arranging access to Anthropic’s systems were suddenly told that their permissions no longer counted.

For middle powers, this is a brutal lesson in dependence. Europe lacks America’s computing base and cannot easily train a frontier model clone, especially if Washington also controls access to the chips needed to do so. The article notes that some officials now talk more urgently about AI sovereignty, but sovereignty is hard to build when the supply chain runs through American companies, American cloud infrastructure and American export rules.

Controls That Leak

The ban also illustrates why crude access controls may fail. Foreign users can buy American identities, route around restrictions or use jailbreak tools. Anthropic already blocks Claude in China, yet some Chinese users still reach it. Meanwhile, barring non-Americans from the newest models is hard to reconcile with how American AI labs actually work: they employ many foreign nationals and rely on global customers, researchers and partners.

That tension matters because the same administration has been inconsistent. It rolled back much of the previous AI regulatory regime, allowed some advanced chip sales to China, then abruptly restricted access to Anthropic’s models. It also asked labs to give the government early access to frontier models, but without a transparent process for judging what should be released, restricted or banned.

The article’s closing warning is that AI access is becoming a national-security question before America has built institutions equal to the task. If the United States is going to act as gatekeeper to the world’s most capable models, it needs rules that allies can understand and firms can plan around. Otherwise the gatekeeper starts to look less like a strategist and more like a bouncer acting on impulse.