The European Union’s Artificial Intelligence Act requires developers to obtain government permission before releasing many AI models. Similar licensing and pre-approval regimes are spreading in other jurisdictions. Supporters claim these rules will make AI safer. In reality, mandatory pre-approval is epistemically impossible, destroys the entrepreneurial discovery process that generates both innovation and genuine safety knowledge, and systematically favors large incumbents.
This is not a minor bureaucratic inconvenience. It is a textbook illustration of the knowledge problem that F. A. Hayek identified long ago.
The Knowledge Problem
Hayek demonstrated that the knowledge required for rational economic coordination is dispersed among millions of individuals. It is local, often tacit, and constantly changing. No central authority can gather and process it effectively. Applied to artificial intelligence, the problem becomes especially severe.
Regulators must predict the behavior of systems whose possible inputs include the full range of human conversation and interaction; a set that is combinatorially explosive. An AI model deployed across hospitals, schools, or factories will encounter unique combinations of local data, human judgment, and unforeseen prompts. No regulatory test suite can simulate the creativity of a clinician who repurposes a diagnostic tool or the ingenuity of an adversarial user who discovers a novel way around safety filters. Even the model’s own developers cannot fully anticipate these outcomes.
The most important capabilities and the most important risks are emergent. They appear only through real-world use. Pre-approval forces regulators to certify a fixed list of “foreseeable” risks before the model ever meets the world. That list is necessarily incomplete. The result is a regulatory judgment frozen in time while the system continues to evolve with every new interaction.
Destroying Dynamic Efficiency
Jesús Huerta de Soto distinguishes static efficiency (rearranging known resources) from dynamic efficiency, the market’s ongoing capacity to discover new ends, new means, and entirely new knowledge. Dynamic efficiency cannot be planned or bureaucratized. AI pre-approval operates on a static model: compile a list of high-risk capabilities, require testing against that list, and block everything that fails to receive official approval.
The European Commission’s own impact assessment estimated average conformity assessment costs in the range of hundreds of thousands of euros and market-entry delays measured in many months for high-risk systems. For a small research team developing a tool to detect antibiotic resistance from clinical notes, that sum can exceed an entire annual budget. Bacterial strains and epidemiological patterns shift on a scale of weeks, not years. The opportunity is killed not because any concrete harm was demonstrated, but because the regulator’s checklist could not
recognise a useful application that did not fit its predefined categories.
Consider a clinician working in a rural clinic in Kisumu, Kenya, where antibiotic-resistant typhoid is evolving faster than the pharmaceutical pipeline. A research team in Bengaluru has built a promising prototype. Because the model uses a general-purpose architecture subject to the EU rules, the researchers cannot absorb the compliance burden for a tool aimed at a disease largely neglected by commercial markets. The model is never deployed. The clinician never receives the early warning. A child dies of an infection that open deployment and real-world learning might have helped flag. The visible result is a stamped compliance file. The invisible cost is a life.
Liberty Requires Permissionless Innovation
To require an AI developer to seek official permission before releasing code is to assert that a central authority may decide in advance which peaceful, voluntary acts are allowed. Applied consistently, this principle generates an infinite regress of permissions or terminates in an unaccountable final authority. The presumption of liberty is not one value among many to be balanced against safety. It is the only principle that does not collapse into tyranny.
The catastrophic-risk objection that even a tiny probability of existential harm justifies prior restraint does not escape the knowledge problem. To prevent a dangerous capability, regulators must first be able to identify it. That identification itself depends on the dispersed knowledge generated by open experimentation. Acting on an unverifiable prediction about a system whose future behavior will be co-determined by millions of users, in contexts not yet imagined, is not prudence. It is the formalisation of ignorance enforced by law.
AI pre-approval regimes do not produce safety. They produce unknown human costs that remain invisible precisely because the causal chain runs through bureaucratic non-deployment rather than through observable failure. Liberty requires the freedom to innovate without prior permission. Licensing schemes invert that presumption while guaranteeing the very ignorance they claim to prevent.
The Interventionist Spiral
Ludwig von Mises showed that intervention rarely stops at the first step. It disrupts the market, creates new distortions, and thereby generates apparent justification for further controls. AI licensing follows this pattern with precision. An initial compute threshold or risk classification is inevitably gamed or circumvented by entrepreneurs. Regulators respond by lowering the threshold or expanding the scope of control. The rising compliance costs systematically favor large incumbent firms that can absorb them. Those incumbents—now sheltered from competition—have every incentive to lobby for still stricter rules that raise barriers against the next generation of start-ups. What begins as a safety measure becomes a regulatory moat protecting the status quo.
The Safety Paradox
There is a deeper contradiction. Pre-approval delays the very process that generates real safety knowledge. Most harmful AI behaviors documented in incident databases appear only after systems have been released into actual use. Controlled testing environments, however elaborate, cannot replicate the full diversity of human creativity, malice, and mismatched expectations.
In cybersecurity, open and competitive deployment has long accelerated safety. Independent researchers discover flaws and work with developers to patch them before malicious actors can exploit them at scale. Closed, permissioned environments keep those flaws invisible and unaddressed. The same logic applies to AI. Real-world interaction is a source of safety information, not merely a risk to be managed. By requiring extensive paperwork and certification before deployment, pre-approval substitutes a static bureaucratic file for dynamic fieldwork.
