Optimists talk about artificial intelligence (AI) as magnificent tools and ultimately a source of world salvation. Pessimists warn that AI can produce the implements of tyranny and ultimately soulless world domination. Many people in both camps take these views without seriously questioning the limits of AI. 

To challenge the optimists: Can AI run a human society’s economy better than humans? Governments are expanding and taking more power to run everything, and many people accept or even cheer them for doing so. To carry out that mission, such Leviathan governments invariably increase taxes and impose regulations upon economic activity: prices, wages, investments, interest rates, use of private property, transfers of ownership, and permissible or forbidden transactions.   

Leviathan’s cheerleaders don’t explain how all-powerful governments claiming to run national economies can actually succeed. What is “success” anyway, and how would you measure it? The cheerleaders imagine AI will make Leviathan’s job possible. Let’s compare the tools to the task at hand.

AlphaZero vs. Deep Blue

Consider two world famous examples of AI: Deep Blue and AlphaZero. In the 1990s, IBM computer scientists designed software and hardware to play world-class chess, naming their system Deep Blue. In 1997, Deep Blue beat the international chess champion, Garry Kasparov. The publicity sound bites heralded the age of computers smarter than humans and the end of chess as we knew it. How did Deep Blue do it?

Deep Blue used algorithms, i.e., methods strictly following pre-set analytical rules. As a logic-based system, “All it did was calculate possible outcomes of different moves and rank them based on the advantages each move would bring to the game.” Deep Blue followed the same move-by-move thought process of a human player; it won because it could evaluate from 720,000 to 200,000,000 chess moves per second. Deep Blue versus a human player analogizes to a modern diesel earth mover versus one person with a shovel. 

A different AI design arose when computer scientists built AlphaZero. As described by reporters, AlphaZero blew minds in 2017: “Given only the rules for chess, AlphaZero quickly learned the game and could defeat previous chess computers and human players alike.” AlphaZero technology beat the world champion player of the board game Go, 100 to 0. There’s no denying AlphaZero’s software marvelously displays human knowledge and design. 

In a nutshell, AlphaZero runs algorithms to set up game board positions, plays them out, then evaluates the “success” of the sequence of moves by whether the end result of the whole sequence is a “win.” AlphaZero technology can run such sequences easily in the billions. Describing the version aimed at conquering Go, George Gilder in Gaming AI (2020) explains:

In a form of “generic adversarial program,” AlphaGo Zero vied against itself repeatedly billions of times. It became its own teacher.

Gilder observes neural network style systems like AlphaZero learn by setting up positions and calculating their value:

[T]he basic steps to the solution are guess, measure the error, adjust the answer, feed it back in a recursive loop.

AI Software Cannot Forecast Human Action

Key to AI’s phenomenal success at “learning” and winning games is a set of constraints (rules) and a definable end point (game victory). As Joshua Sokol wrote in “Why Artificial Intelligence Like AlphaZero Has Trouble With the Real World” (2018): 

From an algorithm’s perspective, problems need to have an “objective function,” a goal to be sought. When AlphaZero played chess, this wasn’t so hard. A loss counted as minus one, a draw was zero, and a win was plus one. AlphaZero’s objective function was to maximize its score.

Turn now to economics – is it a game with analyzable positions and defined “success” points? Answer: No. Reason: Humans are not machines.

Preeminent twentieth century economist Ludwig von Mises, in The Ultimate Foundation of Economic Science (1962), observed that humans cannot be viewed as deterministic and predictable entities that can be predicted by “scientific” methods. Humans do have definable characteristics that allow understanding of economic behavior, said Mises, starting with their uniqueness:

The characteristic feature of man is action. Man aims at changing some of the conditions of his environment in order to substitute a state of affairs that suits him better for another state that suits him less. … The study of man, as far as it is not biology, begins and ends with the study of human action. Action is purposive conduct. It is not simply behavior, but behavior begot by judgments of value, aiming at a definite end and guided by ideas concerning the suitability or unsuitability of definite means.

Economics sees humans as beings aiming for outcomes, just as players of board games. Human action, however, arises from value judgments that are specific to each individual. Look at the menu of any restaurant: A variety of meal options reflects our recognition that people’s subjective personal preferences are neither consistent nor predictable. Tacos for lunch one day, sushi on another. 

Humans act to relieve uneasiness. Relief of uneasiness does not directly equate to having more money, or getting certain property, jobs, or toys. As Mises put it, humans act to change something so they can get a different “state of affairs” that “suits them better” as compared to some other “state of affairs.” 

Ay, there’s the rub for AI. The desired different “state of affairs” is adjudged by the mind of the individual at a specific time, in specific circumstances, and considering the individual’s unique set of thoughts and feelings at the moment. Activities in the market economy reflect all the humans acting and interacting toward relief of uneasiness – each having his or her own desired “state of affairs.” 

Even theoretically, AI cannot determine an individual human’s mental states of preferences and to-be-chosen actions toward the new state of affairs. Humans don’t even know other humans’ preferences and goals at any given time. If you can’t define the goal of a game, then you can’t figure out when the game is won or lost. You therefore cannot know what the “best next move” is.

Deep Blue step-by-step algorithm methods cannot be programmed to explore “next moves” if the human evaluation of “success” cannot be quantified in mathematical or pattern-matching terms. Similarly, AlphaZero cannot solve economic human action problems because there are no “positions” or “patterns” to be tested in the billions per second – and there is no certainty about what the winning goal would be. 

The choice between tacos and sushi is made right now by an individual human. Zillions of past human decisions and actions underlying the economic system made it possible for one individual human today to choose between tacos and sushi that are affordable and available immediately. AlphaZero cannot take the current economic system and forecast all of the future human preferences and decisions. There is no set of rules to learn that allow setting up game boards of human action in all its dimensions for one person, let alone millions of people.

Individual Human Desires Are Not Computer-Friendly

Parents cannot predict their child’s likes and dislikes, whether the child is a toddler or a teen. Spouses are moving, uncertain targets to one another. (Try buying anniversary, birthday or Christmas gifts for them!) What seems like a good idea today could be “meh” tomorrow. Algorithmic AI like Deep Blue, and neural network AI like AlphaZero, cannot obtain solutions to predict and provide for human desires and actions in the future. AI systems cannot anticipate external fact changes for each person either, such as environmental shifts or disasters, new inventions, or inter-human conflicts. In sum: AlphaZero cannot test billions of “positions” of human life on Earth, apply some internally-generated ranking of “good” or “bad,” and identify the optimal moves for everyone.

The best-intentioned AI deployed by an all-powerful government cannot predict or provide the people anywhere near how humans can for themselves via voluntary action and exchange. In the free market economy, each individual and business ascertains its own desired goals, takes action toward the goals, and does so by individual effort in voluntary cooperation with others. Each individual in the market has a new “state of affairs” in mind and acts toward it; each calculates the costs and benefits; each chooses the next best move for the circumstances, accounting for other people’s circumstances through mutually beneficial transactions. Considering the mixture of objective and subjective factors in every aspect of human action, AI systems would scarcely know where to start.