This week in New York, under the banner “Restoring trust, managing transformation,” the world’s diplomats convened for the 81st UN General Assembly and decided, somewhere between the wars and the climate panic, that artificial intelligence now belongs on the short list of things only they can save us from. Secretary-General António Guterres devoted much of his valedictory address to a warning about runaway AI, one of three existential threats he identified alongside climate change and inequality, and he told the world that voluntary, uneven national safeguards would not be enough: only an international framework, he insisted, could stop a race to the bottom on AI safety.
Weeks earlier, an Independent International Scientific Panel he had convened warned that computing power and technical talent were already dangerously concentrated in a handful of companies and countries. Washington, for its part, spent the same week trying to rename the whole technology “Superior Intelligence,” its own category error, but one that at least stops short of pretending the world needs a treaty to manage what it intends to out-race rather than tame.
He is not wrong that something remarkable is happening to the world’s most important new technology. He is wrong—in the oldest and most predictable way a planner can be wrong—about who is equipped to manage it.
Austrian economists have a name for that mistake, and it did not take a supercomputer to discover it: the knowledge problem. Ludwig von Mises showed that, without market prices for the factors of production, no planning body, however well staffed with independent scientific panels, can rationally compare alternative uses of scarce resources. Friedrich Hayek also argued that the knowledge an economy actually runs on is not the kind you can collect in a UN report. It is local, tacit, and constantly shifting, scattered across millions of minds that no committee in Geneva or New York will ever fully access.
A recent Mises Institute essay on climate policy put it with unusual precision: recognizing a problem is not the same as knowing its solution, and the existence of a crisis does not by itself prove government has the knowledge required to fix it. Swap “climate” for “AI” in that sentence and nothing about it stops being true. A Global Fund for AI, a UN-backed capacity-building network, a scientific panel issuing verdicts on which algorithms are safe: these are the same command-and-control reflexes that gave the twentieth century central banking and five-year plans, now dressed in the vocabulary of safety.
It helps, too, to be honest about what AI actually is. Per Bylund, a senior fellow at the Mises Institute, has spent years making the unglamorous point that artificial intelligence is not a science-fiction creature plotting humanity’s demise; it is capital: a producer’s good, valuable to the extent entrepreneurs can put it to work discovering what consumers actually want, and bound, like any tool, by the humbling rule of garbage in, garbage out. A chatbot that can write a passable sonnet is not a being with intentions, and dressing it up as an oncoming Terminator obscures where the real danger sits: not with the machine, but with whichever authority gets to decide how it is used. That is a fundamentally different problem than the one the General Assembly believes it is solving, and it calls for a fundamentally different remedy: not a treaty, but a market, full of entrepreneurs experimenting, failing cheaply, and correcting course in ways no five-year AI plan has ever managed.
We have, in fact, already run this experiment, and the transcripts survive. When Edward Snowden pulled back the curtain on America’s national-security state, the lesson was not that surveillance had gone slightly too far. It was that once a government builds the infrastructure to watch, sort, and predict its citizens, it does not hand that infrastructure back. A decade after his disclosures, press-freedom advocates were still cataloguing the ruthless persecution Snowden faced and the intelligence agencies’ ongoing refusal to submit to real oversight.
Artificial intelligence is that same infrastructure’s next, far more capable, generation, and the governments now asking the UN for AI guardrails are, in parallel, quietly building it into their own security services. A Mises Institute analysis published this year traced the pattern directly: DARPA and the National Science Foundation funded the early neural-network research underlying nearly every major AI lab today, and firms such as Palantir and Anthropic now supply AI analytics straight to US defense and intelligence agencies, including tools aimed at centralized citizen databases and predictive policing.
Murray Rothbard once warned that the state’s favorite trick is to dissolve itself into an innocent collective “we,” so that its accumulation of power is never presented as what it is. Mises was blunter: government, he wrote, is essentially the negation of liberty. Readers of Orwell will recognize the shape of what internationally-administered AI guardrails—run by permanent bureaucracies with no shareholders, no profit test, and no exit—could quietly become. Central planning with better facial recognition is still central planning.
The UN’s own family has already rehearsed managing a fast-moving global emergency, and the results are on the public record. The World Health Organization’s independent review panel, co-chaired by two former heads of state, concluded that the global pandemic alert system was not fit for purpose and that the agency itself was structurally overmatched, a verdict reached only after WHO waited weeks to declare an international emergency while diplomatic sensitivities about a powerful member state visibly shaped its public statements. That is the institutional track record now being offered as the template for AI oversight, alongside a General Assembly that struggles to keep its own peacekeeping budgets under control and a system of central banks whose twenty-first century has been one long improvisation from crisis to crisis. None of these institutions has covered itself in the kind of glory that inspires confidence that it should now referee a technology that changes by the week.
Nowhere is the mismatch between rhetoric and reality more visible than in the eagerness of governments, several of them African, to demand a seat at the table of AI control before they have so much as a seat at the table of AI production. An Oxford-led study published this year found that only 32 countries host specialized AI data centers at all; more than 150 have essentially no access to the computing power a modern model requires, and Africa and South America are almost entirely absent from the map. Sub-Saharan Africa hosts roughly 5.5 percent of the world’s data-center capacity despite a far larger share of the world’s population, with the computer that exists, the chips inside it, and the capital behind it still controlled overwhelmingly by American, Chinese, and Gulf firms.
Talk of “AI sovereignty” from governments that own almost none of the computers, hold few of the patents, and attract a rounding error of the investment is not sovereignty. It is a request to co-sign rules for an industry they do not host, using leverage they do not have. A genuinely global, borderless technology, trainable on rented cloud capacity in any jurisdiction, its weights copyable in seconds, cannot be fenced in by the regulatory preferences of any single country or UN committee. It can, however, be used as cover for the same governments to tighten their grip at home.
Notice, finally, the vocabulary doing the heavy lifting here: runaway, existential, race to the bottom. This is the language of emergency, and emergency is the oldest lever a planner has for suspending ordinary scrutiny of whatever he proposes next. It worked during the pandemic. It has worked for two decades on climate policy. It is now being tried on AI, and the burden of proof has quietly shifted: the innovator must justify his freedom to build, rather than the regulator justifying his authority to forbid.
None of this means AI carries no risk, and no serious Austrian has ever claimed markets produce a costless world. It means only that the work of managing a discovery process belongs, as it always has, to the discoverers: the entrepreneurs, investors, and ordinary users testing, failing, and correcting error through profit and loss, and not to a General Assembly that cannot balance its own peacekeeping accounts, restrain its most powerful members’ worst abuses, or keep a single pandemic under proper review for longer than a news cycle.
Over the past century, governments have done more lasting damage to their own citizens’ constitutional liberties than any chatbot has yet managed. The state did not invent artificial intelligence, does not understand it, and has already shown—in Geneva conference rooms and intelligence-agency server farms alike—exactly what it intends to do with the authority to control it. The market’s advantage was never that entrepreneurs are smarter than diplomats. It is that they are allowed to be wrong, one at a time, cheaply, and at their own expense, which remains the only safety testing a technology this new has ever actually received.
