Few sights in modern economics are stranger than an industry asking to be regulated. In May 2023, OpenAI’s chief executive told a Senate subcommittee that Congress should create a new licensing agency to oversee safety standards and require independent audits. Executives have since signed open letters urging pauses and oversight. The usual explanation comes from public choice: incumbents want regulatory moats. That is plausible but incomplete. The Austrian tradition offers a deeper objection. Even a benevolent, incorruptible regulator could not do what AI licensing asks of it, because the knowledge it requires does not exist in any single (or centralized) mind.
Interventionism Does Not Stay Put
Mises argued that partial interventions in the market tend to produce results their authors did not intend, and that these results invite further interventions. AI licensing fits the pattern. Compliance costs are largely fixed, so they weigh most heavily on small entrants and least on firms with large legal departments. As a result, concentration rises. That concentration then becomes the justification for tighter oversight. What began as risk mitigation ends as a managed cartel.
There is also a problem of ends. In Mises’s praxeology, action is the use of means to reach ends, and the ends belong to acting individuals. In this context, an algorithm is only a means. When an agency decides which models may exist and for what purposes, it does not merely regulate a technology. It replaces the ends of millions of users and entrepreneurs with those of officials.
Mises’s calculation argument was aimed at socialism, and AI regulators do not abolish markets. Still, the logic carries over in miniature. To license a model, an agency must judge whether a requirement’s marginal safety benefit exceeds its cost in forgone innovation. The regulator faces no profit-and-loss test, and the costs are precisely the applications that never get built, which no one sees. Economist and political scientist Solis-Mullen makes a parallel point: the displaced worker is visible and immediate, while the jobs that innovation would have created are not, and this asymmetry pushes politicians to intervene.
Safety Is Discovered, Not Given
Hayek located the economic problem in the fact that knowledge exists only in dispersed, incomplete, often tacit fragments, and that prices coordinate it. In a later lecture, he described competition as a discovery procedure, valuable because its outcomes are unpredictable and different from what anyone could have consciously aimed for.
Apply this to “safe AI.” Regulators talk as if safety were a fixed specification that experts can write down in advance. But whether a system is safe depends on context: a triage tool in a hospital, a coding assistant, and a chatbot for teenagers carry different risks, different users, and different safeguards. That knowledge is contextual and revealed through use, through failures, reputational damage, lawsuits, and insurance premiums. Safety is an emergent property of decentralized experimentation, not a static variable that can be measured ex ante.
Hayek distinguished cosmos, a spontaneous order that grows from individual actions, from taxis, a deliberately arranged organization. Ex-ante licensing converts an adaptive cosmos into a taxis with a single hierarchy of ends. It also creates a chokepoint. Andreen observes that AI-risk commentators rarely name the state as a potential bad actor, even though the core technologies of today’s labs grew out of state-funded research and now feed government surveillance and defense work. A licensing agency would decide who may build, and that gatekeeping power is exactly what a state seeking to expand its reach would value.
Regulation in the Capital Structure
Garrison’s capital-based macroeconomics offers a useful lens, with one caveat: his framework was built to analyze credit-driven cycles, so what follows is an extension by analogy. Foundational models are high-order capital goods. They sit in the early stages of the Hayekian triangle, and downstream products such as medical tools, legal assistants, and tutoring systems depend on them.
Ex-ante licensing is a shock to those early stages. Compliance burdens lengthen the time before returns arrive and raise the opportunity cost of capital committed there. The economy’s intertemporal frontier shifts inward: less capability is built, and it arrives later. Resources also migrate. Talent and capital flow away from improving models toward legal, audit, and lobbying activity, and toward established architectures that already have the resources to pass approval. This is a form of malinvestment created by regulatory rules rather than by credit expansion, and it distorts the capital structure toward what regulators can certify instead of what consumers value.
The Strongest Objection
The strongest objection is that some AI risks are catastrophic, and that liability comes too late when the harm is irreversible. Austrian-school economists should take this seriously. Rather than “no rules” at all, their answer is “different rules.” Property rights, contracts, tort liability, and insurance already price risk and punish harm, and they do so using dispersed knowledge. An insurer that must pay for a model’s failures has every incentive to audit it, and an insurer’s premium is a price signal, which a licensing exam is not. I concede that liability is weakest at the extreme tail, where a defendant may be judgment-proof. But a monopoly licensing authority is a poor remedy for that weakness, since it concentrates precisely on the discretion and error that a competitive process would disperse.
The Latin American Double Distortion
For technology importers like Colombia and its neighbors, the case is stronger still. Several Latin American legislatures have debated bills that borrow the risk-tier architecture of the EU’s AI Act. That framework was designed for an economy with deep capital markets, large compliance industries, and firms able to absorb fixed costs. Importing it means paying rigid institutional costs before accumulating the capital that would make them bearable.
It also risks blocking convergence. Developing economies raise labor productivity largely through fast, cheap adoption of foreign technology, and open-source models are especially valuable there because local firms can adapt them without permission. Licensing barriers suppress exactly that adaptation.
Finally, centralized risk management is only as good as the institutions running it. Where they are weak, an AI regulator is likely to become a source of local rents, a new office to lobby and a new permit to sell. Solis-Mullen’s account of Beijing warning firms against AI-driven layoffs, and of California’s executive order on displaced workers, shows how quickly economic change becomes a political object.
Who Should Really Decide If AI Is Safe?
The clamor of AI executives is less mysterious than it looks. Incumbents may honestly fear technology, or may quietly welcome rules they can afford, but neither motive changes the epistemic problem: no agency can know in advance what safe AI looks like. That knowledge emerges from trial, error, liability, and competition. The task is not to abolish safety concerns but to place them where the knowledge is, with owners, insurers, courts, and users, and to resist the temptation to plan a process whose whole value lies in the unplanned.
Photo Credit: Photo by Steve A Johnson on Unsplash
We roast sacred cows medium-rare. Help fund the habit!
This piece reflects the author’s views, not necessarily the entire magazine. We welcome a range of pro-liberty perspectives. Send us your pitch or draft.
