Dozens of startups now offer Artificial Intelligence tools to help businesses set market prices. Assuming unlimited computing power to run such models and comprehensive data sets to train them, can AI replicate the way human actors make decisions in the marketplace? Socialists have argued for more than a century that enlightened bureaucrats can set prices as well as the myriad of private actors in the marketplace. Ludwig von Mises offered a celebrated refutation of the socialist case. Does the vast computing power behind Large Language Models give new life to the socialist argument? The answer is no, but Mises’ argument needs to be updated and sharpened.

　　We think of the marketplace as a realm of free personal choice, a venue that makes possible the individual’s pursuit of happiness (or at least money, which is the next best thing). What if machines could run a market as well as humans? What would that imply about our notion of freedom? It depends on what we mean by freedom. Computers could replicate the arbitrary exercise of individual caprice. But they can never innovate—that is, discover new ways of understanding nature and changing it. AI can work out the supply chain for a supermarket. It will never replicate a Jeff Bezos or Elon Musk.

　　As a starting point for this inquiry, I asked ChatGPT whether AI could duplicate the market mechanism. Here is what it told me:

　　How AI Can Duplicate Market Pricing Mechanisms:

　　1. Data Analysis and Pattern Recognition: AI can analyze vast amounts of data, recognizing patterns and trends that influence pricing, such as supply and demand fluctuations, competitor prices, consumer behavior, and historical data. Machine learning algorithms can predict future prices based on past trends and current market conditions.

　　2. Dynamic Pricing: AI can implement dynamic pricing strategies in real-time, adjusting prices based on current demand, inventory levels, and competitor prices. This is already used in industries like e-commerce, airlines, and ride-sharing.

　　3. Efficiency: AI can process information faster and more accurately than humans, leading to more efficient price setting and adjustment. This efficiency can help balance supply and demand more effectively.

　　4. Personalization: AI can personalize prices for individual consumers based on their purchasing history, preferences, and willingness to pay. This level of personalization is difficult to achieve through traditional market mechanisms.

　　This argument sounds convincing, and in a way it is convincing; it has the sole defect of being wrong. But why it is wrong is not self-evident. Some libertarians identify free choice in the day-to-day operations of the marketplace with freedom itself; that was the nub of Milton Friedman’s celebrated 1980 book Free to Choose. If AI could reproduce all the functions of the marketplace, it would be ludicrous to associate human freedom with day-to-day choice in the marketplace. Our behavior would resemble the random movement of particles in Brownian motion. Each particle might appear to move freely at the micro level, but the behavior of the whole system would be determined by scientific laws. Our sense of freedom would be an illusion.

　　The question of free will vs. determinism has tormented philosophers for millennia. If nature is a machine, that is, entirely explainable by scientific laws which by definition are deterministic (and that includes quantum probability, which is another form of determinism), free will must be an illusion. After all, the brain is made of chemicals we can analyze and uses physical means of transmitting information that we can monitor. If the brain simply is a machine (albeit a very complex one), then whatever happens in our brain must follow deterministic laws. Free will must be an illusion.

　　Attempts to replace the marketplace with something like a computer have been circulating since the Austrian economist Otto Neurath proposed an “administrative economy” in which planners would produce results better than market decisions. In 1920, Ludwig von Mises refuted Neurath’s half-baked proposal, arguing that no central system could reproduce the vast amount of information generated by price discovery in the marketplace. Mises and, a generation later, Friedrich von Hayek, argued that the marketplace is simply too complex for central planners to navigate; the number of individual bids and offers for goods and services is too large to be modeled.

　　If we assume that today’s decisions will be based on the same technology and other premises as yesterday’s, though, Artificial Intelligence could replicate an arbitrarily large number of individual decisions, by training machine-learning models on previous data. Without changes in technology and industrial organization, today’s output of goods and services will be the same as yesterday’s, and so will the investment opportunity set. Something like what Otto Neurath had in mind is embodied in the enterprise resource software that large firms use to source thousands of parts and myriad raw materials as well as hundreds of kinds of labor inputs.

　　AI can’t replace the innovative creativity of entrepreneurs. On the contrary: AI itself is an innovation whose outcome is uncertain.

　　But the assumption that today’s economic activity will stem from yesterday’s technology is the problem. The real world does not remain fixed. Innovation transforms economic life every day. The world doesn’t remain fixed in place, and deterministic models can’t capture its dynamism. Freedom is not an illusion; it manifests itself in our capacity to innovate. Scientists discover new physical properties, inventors turn them into applications, investors raise funds to commercialize them, entrepreneurs start companies to exploit them, workers learn new skills, and consumers embrace new products. Edmund Phelps, the 2005 Nobelist in economics, observed in his 2013 book, Mass Flourishing, that in periods of extraordinary growth, innovation embraces the whole grassroots of society. Freedom can’t be reduced to the choice between Coke and Pepsi; it is expressed in human creativity. And creativity isn’t the property of a few geniuses; it is a spirit that uplifts every participant in the modern economy.

　　Even if machine-learning models trained on every past transaction in the marketplace could duplicate the market mechanism as of yesterday, no model trained on past data could possibly make decisions about innovation. As I will explain, Modern Portfolio Theory, specifically Robert Merton’s Intertemporal Capital Asset Pricing Model, provides an important supporting argument.

　　In light of AI applications to complex systems, Mises’ complexity argument needs revision and updating. He distinguished between a “lower order” and a “higher order” of economic decision-making. His translator explains, “By ‘lower order’ Mises refers to those goods made for final consumption, and by ‘higher order’ those used in production.” As Mises wrote:

　　As a rule, the man who knows his own mind is in a position to value goods of a lower order. Under simple conditions, it is also possible for him without much ado to form some judgment of the significance to him of goods of a higher order. But where the state of affairs is more involved and their interconnections not so easily discernible, subtler means must be employed to accomplish a correct valuation of the means of production. It would not be difficult for a farmer in economic isolation to come by a distinction between the expansion of pasture-farming and the development of activity in the hunting field. In such a case the processes of production involved are relatively short and the expense and income entailed can be easily gauged. But it is quite a different matter when the choice lies between the utilization of a watercourse for the manufacture of electricity or the extension of a coal mine or the drawing up of plans for the better employment of the energies latent in raw coal. Here the roundabout processes of production are many and each is very lengthy; here the conditions necessary for the success of the enterprises which are to be initiated are diverse, so that one cannot apply merely vague valuations, but requires rather more exact estimates and some judgment of the economic issues actually involved.

　　Assuming unchanged technology, Mises’ argument comes down to the assertion that investing in capital goods is more complex than pasturing a cow. Complexity as such melts before computational power: Machine-learning models can estimate demand curves and specify supply chains for the most complex plants in the world, from the most sophisticated semiconductor lithography machines down to the janitor’s mop.

　　What AI cannot possibly know is the price of a product that doesn’t yet exist. As Reuven Brenner, one of our great theorists of innovation, wrote in 1987:

　　How are prices set in a world where competition is done through innovations? Since people introduce new products and bet on numerous new competitive strategies in either marketing, the internal organization of the firm, or advertising practices, there are no standards to which one can just adapt. The answer is that pricing just becomes on particular competitive strategy, an art.

　　A related question is: Why is Nvidia worth $3.2 trillion at the June 17 close, selling for 77 times earnings? Firms offering new products bet on the correct price, which by definition can’t be observed, let alone modeled using machine-learning algorithms trained on past data. As Brenner observes, it is an art, not a science. The same issue arises when we try to assign prices to firms that offer new products.

　　The standard theory of asset pricing, the Capital Asset Pricing Model (CAPM), states that the price of an individual stock relative to the overall market depends on the expected return to the market as a whole, and the individual stocks sensitivity (covariance) to the overall market, or beta. Price is a function of risk and reward. CAPM is an indispensable intellectual construct and a brilliant simplification. The only problem is that it doesn’t work. Observed prices are unrelated to what CAPM predicts, and generations of finance theorists have tried to turn the lead of financial statements into the gold of excess returns by searching for other factors.

　　Why CAPM doesn’t work is obvious when we consider that not one of the original 30 members of the Dow Jones Industrial Average of 1896 still exists. The original 12 stocks in the DJIA included American Cotton Oil, American Sugar, American Tobacco, Chicago Gas, Distilling Cattle Feeding, Laclede Gas, National Lead, North American, Tennessee Coal and Iron, and other long-dead enterprises.

　　In 1973, Nobel Laureate Robert Merton pointed out that CAPM is a one-period model that assumes the investable universe (the “investment opportunity set”) is constant. A multi-year (“intertemporal”) model, though, has to take into account changes in the investment opportunity set. Youre supposed to buy the market portfolio plus hedges against changes in the investment opportunity set. That’s the stock-picking equivalent of belling the cat. How do you know the correct hedge for technological changes you can’t possibly predict? If you bought Intel—one of the investors of chip design—five years ago at the then price of $50 a share, you would be disappointed with today’s price of just $21 a share. Of course, you could have bought Nvidia, a startup selling at $4 a share, and held it until today’s close of $130 a share. But who knew?

　　Merton’s formula doesn’t help us price tech stocks, because we can’t possibly know in advance which innovations will succeed. It does help explain why the market sometimes assigns arbitrary prices to tech stocks, for example, the great Internet bubble of the late 1990s. I am not suggesting that Nvidia’s stock price is a bubble; I simply don’t know. I discuss these and related issues in a paper published earlier this year by the Heritage Foundation. 

　　The work of Merton, Brenner, and others points to an important conclusion: The exercise of freedom in the marketplace depends on the creativity of entrepreneurs who bet on innovations. The outcome of prospective innovations cannot be known in advance. Mises complexity argument isnt enough. Freedom is the counterpart of uncertainty.

　　The worst thing that governments can do is to try to pick winners among commercialized technology. But it is a proper and necessary function of government to support fundamental scientific research. Private corporations won’t fund RD that has no evident commercial purpose. Their job is to make money for their shareholders. There is a good reason to direct government research subsidies through private corporations, as opposed to universities. Private companies with experienced engineers and manufacturing facilities are better suited to identify the practical utility of scientific discoveries than academic laboratories. Public-private partnerships where the government supports basic RD, along the lines of NASA in the 1960s and the Defense Advanced Research Projects Agency at the peak of the Cold War, are a kind of Hamiltonian public improvement. Private entrepreneurs should bear the whole risk of commercialization.

　　AI can’t replace the innovative creativity of entrepreneurs. On the contrary: AI itself is an innovation whose outcome is uncertain. Some applications (replacing human beings on corporate help desks, for example) may turn out to be trivial; others, for example devising new pharmaceuticals, may be revolutionary. Only in an imaginary world in which no innovation occurs could we envision an AI-driven marketplace.

　　Artificial Intelligence isn’t intelligence in the first place. It can replicate the lower-order functions of the human mind, the sorting and categorizing faculty, and perform such operations much faster than humans. But it cannot reproduce the higher-order functions of the mind—what Immanuel Kant called Vernunft (roughly, critical reason) as opposed to Verstand (usually translated as “understanding”). It can mine data from past experience, but it can’t stand at a distance from experience and ask, “What if we did things differently?” Freedom is the freedom to create, and that is what free societies must preserve.