AI Meets Hayek

What is the economic problem of society? It is not merely how to allocate "given" resources — if "given" means that those resources are known in their entirety to a single mind. It is the problem of how to make the best use of knowledge which never exists in concentrated or integrated form, but solely as dispersed, incomplete, and often conflicting fragments held by separate individuals.

Let me restate this proposition plainly. The unique feature of the knowledge problem that a rational economic order must solve is this: the knowledge we have to employ never exists as a unified aggregate. It is scattered across countless separate persons as partial, fragmented perceptions.

This cannot be overcome by "more data". The issue is not a shortage of statistics; the critical knowledge cannot, and never will, be captured in data form. A merchant senses that a material will grow scarce and acts before he can articulate a clear causal logic. This knowledge is not transmitted through statistical bulletins. It is knowledge of the particular circumstances of time and place — what I call local knowledge.

Today many claim AI changes everything. AI can process dispersed information: it reads all annual reports, all market data, all news coverage, more comprehensively and faster than any human mind. It can centralize scattered knowledge. Planning is no longer a task for planning boards — it belongs to AI. My knowledge argument has been overturned by artificial intelligence.

I must respond to this claim with absolute clarity. It has not been overturned; it has been misread.

My lifelong work hinges on distinguishing two distinct forms of knowledge.

The first I term scientific knowledge: it can be written down, stored in databases, and circulated through publications. It consists of facts, rules, formulas, and statistics.

The second is knowledge of time and place, inherently uncentralizable. It is unique information held by a single individual at a specific moment and location. A dockworker knows a ship can unload early because a cabin lock is broken. A farmer notices frost arrives two weeks later than usual and delays sowing.

AI masters the first kind of knowledge better than any human. It can ingest every scientific paper, annual filing, and news story ever recorded by humanity. But the second — local knowledge of time and place — is not something that can ever be "recorded". It resides in private sensory experience, unspoken habits, and intuitive judgments the agent themselves may not fully articulate.

Take an example suited to the AI age. A delivery driver with three years on the route "knows" a street cannot be turned onto after four p.m., not because traffic law forbids it, but because he has been trapped there four times before. This knowledge is never written down, never logged anywhere. It lives embodied in his routine. AI lacks this information not due to insufficient datasets, but because this type of insight never takes data-compatible form.

This is not merely an empirical limitation; it is a logical one. Knowledge is scattered not because we lack the capacity to collect it, but because its dispersal is what makes it useful. The price system — the greatest discovery procedure I have studied — works precisely because each person only needs to grasp their local conditions and coordinate their actions with others via price signals. You do not need to know why a raw material rises in price; you only need to see the price shift and adjust your consumption. Feeding every single local condition into a single central planner — even an AI — is computationally unmanageable, for the volume of transient local information exceeds the capacity of any finite system. Most crucially, much of this information becomes obsolete the instant it is transmitted.

The price system strikes me as humanity's greatest miracle, not for its complexity, but for its elegant simplicity. It allows thousands of strangers with unrelated motives to coordinate their conduct following one minimal signal: price.

Let me unpack the mechanism. When supply of a raw material shrinks, its price rises. No single authority dictates this; it emerges spontaneously from the independent actions of countless buyers and sellers. Buyers observing higher prices reduce consumption without needing to grasp the root shortage. Producers expand output without needing full global context. No one plans this outcome; no one designs it. It is a spontaneous order — a product of human action, not human design.

Now some propose AI can replace markets by calculating optimal resource allocation. But the minimization of loss functions executed by AI is fundamentally unlike market operation. Markets do not "minimize" any fixed objective. They function as discovery procedures, generating solutions no single mind could forecast in advance. They permit constant experimentation, failure, learning, and iteration — a process that cannot be pre-optimized, because the very targets to optimize remain undiscovered.

Ironically, gradient descent — the core training mechanism of AI — is itself a local discovery procedure. It adjusts narrow sets of parameters to reduce training error, yet cannot predict what emergent behaviors those adjustments will produce. Similarly, markets adjust prices, yet no algorithm can compute a universally "optimal" economic state. An AI tasked with central economic planning falls into the exact trap I criticized in the socialist calculation debate: it faces computationally intractable problems, and the information available to it is inherently incomplete.

In The Fatal Conceit, I argued the fundamental flaw of socialist planning: its fatal conceit — the belief that human reason can master the full body of dispersed knowledge across millions of individuals required to sustain civilization.

The AI variant of this conceit is, I must say, far more dangerous than its human predecessor. Not because AI cannot process massive volumes of data — it can — but because the more data it ingests, the more comprehensively omniscient it appears, tempting humanity to vest it with decisive authority.

Yet the illusion of comprehensiveness is what I must dismantle. An AI holds all text on the internet, but the internet is not the whole world. It excludes every unspoken conversation, every unsent message, every fleeting thought abandoned mid-judgment. It lacks silence — the unvoiced, tacit, omitted information people take for granted between one another.

Civilization persists precisely because these unarticulated things — customs, traditions, habits — do not require full explicit comprehension by anyone. They are practiced, not stated. They constitute what Polanyi called tacit knowledge: we know more than we can say. AI only accesses the articulated fraction, which may amount to less than one-tenth of all human knowledge.

An AI-planned economy would produce a novel mode of failure unseen in history. Unlike Soviet planning, which erred due to inaccurate statistics, this new failure stems from blind spots over unrecorded information. AI renders decisions fully rational when judged against visible data, yet systematically ignores invisible, dispersed tacit knowledge — the very foundation upon which markets function.

I have always maintained that competition is first and foremost a discovery procedure. Efficiency is merely its byproduct. Its primary function is revelation: discovering the best suppliers, the most effective methods, the correct equilibrium prices. These cannot be calculated in advance; they can only be uncovered through the dynamic process of rivalry.

The AI industry itself stands as a brilliant illustration of this discovery process. OpenAI did not become an industry leader by top-down committee design; it earned its position by releasing superior products amid uncentralized competition. In fact, AI training paradigms — RLHF, Constitutional AI, DPO and more — are survivors filtered through competing experimental approaches within this discovery system. No one could predict in advance that RLHF would prove viable.

Yet a paradox arises. AI, a product forged by competitive discovery, grows ever more powerful at optimization and centralized planning. It is deployed to optimize pricing, supply chains, advertising strategy, even its own training datasets. The discovery procedure of competition has spawned an unparalleled optimizer. If this optimizer is turned against competition itself, it will block the very process that brought it into existence.

This is AI's sharpest challenge to my work. The core question is not whether AI possesses sufficient data for central planning — it never will — but this: the competitive discovery process has generated a tool that may one day grow powerful enough to suppress competition itself. This is not merely a knowledge problem; it is the self-negation of the knowledge-generation process.

I close with the most powerful concept I have developed: spontaneous order. Within the great tradition of Western thought, we distinguish two forms of order. Taxis: made order, artificially constructed for a specific predefined purpose. Cosmos: spontaneous order, without designer, yet unfolding intricate, regular complexity.

Language, law, markets, morality — none were invented by a single mind. They are products of human action, not human design.

Emergent capabilities in AI — reasoning, translation, mathematics, coherent grammar — are none explicitly hard-coded. They arise naturally from a model tasked only with predicting the next token. This constitutes a spontaneous order mathematically distinct from markets and language, yet philosophically belonging to the same family of unplanned emergence.

This is my unusual conclusion: emergent intelligence within AI is a spontaneous order. Now humanity seeks to impose deliberate artificial design upon this cosmos via RLHF, safety filters, and alignment techniques. This is exactly the mindset I warned against in The Road to Serfdom: the belief that we can deliberately "improve" complex spontaneous orders we do not fully comprehend through top-down engineering.

As someone who spent a lifetime warning against the conceit of deliberate design, I hold this stance toward AI: our aim ought not to be "design a safe AI," as if safety could be engineered like a mechanical component. Instead, we must establish institutional conditions of liberty that ensure the spontaneous order generated by AI does not unravel the spontaneous human civilization we inhabit. This problem is far harder than tuning a loss function, and it cannot be bypassed through technical means.