AI and Economic Calculation
What is economic calculation, and can machines do it for us? Computer says "no".
Earlier this week I had the great fortune of participating in a Santa Fe Institute workshop on The Calculated Economy in the Era of Machine Learning. The workshop's broad frame was exploring how AI may change the boundaries between markets and the state, and its springboard was the socialist calculation debate. The question hinges on the substance of economic calculation and the role of human cognition in it. It has implications for how we design and use AI as well as for other institutional questions like the nature and role of regulation in complex cyber-physical-social systems like electricity.
Photo by Pietro Jeng on Unsplash
The extent to which the computational advances of AI will enable more and/or better state central planning is part of this question, which is why the socialist calculation debate is a good starting point (my summary here draws heavily on Lavoie (1985) and Boettke & Candela (2023), both of which are good sources if you want to read more, references are at the end). This debate begins in post-World War I Vienna, when the economist Otto Neurath argued that the war demonstrated the feasibility of centralized government planning of a complex economy, and that money would even become unnecessary. His colleague in Vienna, Ludwig von Mises, disagreed strongly in his “Economic Calculation in the Socialist Commonwealth” (1920). Over the next two decades, Mises and F.A. Hayek would engage in written debate with several interlocutors, most notably Oskar Lange (1936) and Abba Lerner, and the debate between Lange and Hayek continued into the 1970s.
Mises provided the starting point: socialism is defined by common ownership of the means of production, and without private property and institutions for exchange that allow (relative) prices to emerge, competition does not exist and coordination and rational resource allocation in a dynamic, complex economy is impossible. Boettke and Candela provide an excellent summary of Mises’ argument:
Mises did not claim that socialism was impossible, per se. Rather, Mises argued that rational economic calculation under socialism was impossible. Thus, Mises was not rejecting the aim of increasing economic productivity in a manner that would increase standards of living for the poorest and least advantaged in society. Rather, taking the ends of socialism as given, Mises was addressing the incoherence between the means and ends of socialism, instead of making a normative critique against the ends of socialism. ...
The critical linchpin in Mises’ argument was that the rationalization of production projects for direct use under a single central plan would be rendered senseless in the move to total socialization because without private property in the means of production, there would be no way for economics actors to engage in rational economic calculation . This is because outside the context of exchangeable private property rights, exchange ratios in the form of money prices cannot emerge to calculate the opportunity cost of capital goods in alternative consumer uses (Mises 1920 [1935]: 111). Therefore, without profit and loss signals to communicate whether or not capital goods have been directed toward value-creating consumer uses, economic actors will have no economically meaningful way to sort from the array of technologically feasible projects those which are economically viable. (Boettke & Candela 2003, p. 47)
Lange and others making the market socialist argument focused on the calculation aspect of Mises’ argument – if you could use devices like input-output tables to gather economic data, the solution to Léon Walras’ general equilibrium system of equations could be computed (Boettke & Candela 2003, p. 48). They argued that the primary constraint on central planning was computational capabilities. The growth in those capabilities from the invention of the transistor to vacuum tubes to integrated circuits to microprocessors creates more and more data and dramatically more computational power. Their fundamental claim was that the complex economy is computable.
This argument inaccurately conflates computation and calculation. Mises emphasized economic calculation, the use of relative prices for the exchange of rights embodied in the exchange of goods, to calculate opportunity costs of different goods in different uses. Human preferences determine those opportunity costs. As Stephen Phelan puts it:
As such, a machine makes a poor entrepreneur because it does not care about the significance of one economic judgment over another. ... In every AI task to date, the importance of one outcome over another is pre-specified by a human. Humans tell the AI what to care about. We care about winning a chess game, we care about making a profit, we care about not hitting a pedestrian with a vehicle. The AI, on the other hand, places no inherent value on one sequence of moves (or one combination of resources) over another. (emphasis original; Phelan 2020, p. 74)
To riff off of Boettke and Candela, economic calculation consists of the discovery of contextual knowledge, not the computation of data. Powerful computation machines are unable to discover contextual knowledge of opportunity costs without humans in the loop to distill their preferences into algorithms (which is itself a challenging epistemic question).
This is where Hayek comes in, making the well-known argument that Don Lavoie would later call the knowledge problem. Taking the liberty of quoting myself:
Knowledge is inherently imperfect, because it is dispersed, private, local, often tacit, frequently inarticulate, sometimes ephemeral, and usually contextual. Economic models based on assumptions of perfect knowledge thus do a poor job of capturing the informational and epistemological factors that are most relevant to both static and dynamic decision-making and economic calculation. Economic and social institutions designed deliberately based on those models are unlikely to perform well at generating prosperity, as research in robust political economy indicates and as Hayek suggested in his Nobel address. (Kiesling 2015, p. 62)
Knowledge is not data, and data are only an incomplete surrogate for knowledge. Knowledge is perception, interpretation, and judgement; the distillation of those elements into action in an economic system with prices (and profit and loss) creates data. Economic calculation has an irreducible cognitive dimension because it is grounded in subjective personal judgements about opportunity costs.
Clarifying these differences between knowledge and data suggests that the complex economy is in fact not computable, at least not in any meaningful sense that reflects underlying human values and can adapt to unknown and changing conditions in dynamic systems. AI can generate, process, and analyze data, but AI cannot react to data and take actions without contextual knowledge grounded in human cognition. AI cannot perform economic calculation without human input.
I don't have a strong theory for what this implies for the precise institutional delineations between “the market” and “the state”, except that this distinction between knowledge and data, between computation and economic calculation, bounds us strictly away from the Lange-Lerner vision of being able to replicate the market's outcomes in a centrally-planned “technosocialist” (Boettke & Candela 2023) system. We still have to do the hard work of comparative institutional analysis, and should do so bearing in mind the epistemic reality in which we are and will be developing AI.
References:
Boettke, Peter J., and Rosolino A. Candela. "On the feasibility of technosocialism." Journal of Economic Behavior & Organization 205 (2023): 44-54.
Kiesling, Lynne. "Knowledge problem." Oxford Handbook of Austrian Economics (2015): 45-64.
Lange, Oskar. "On the economic theory of socialism." Review of Economic Studies (1936), reprinted in Readings in Welfare Economics (1973).
Lavoie, Don. Rivalry and Central Planning (1985).
Mises, Ludwig von. Economic Calculation in the Socialist Commonwealth (1920).
Phelan, Stephen. "Can entrepreneurship be learned by intelligent machines?" (2020)
Related prior posts:
The not small item that many people ignore is that China’s rise parallels the expansionary monetary policy of Greenspan then Bernanke starting in 2003-4. Lowest interest rates in 6000 years (read “The History of Interest Rates “ by Scylla.
But AI should improve cost benefit analysis which takes relative prices and objectives as given and calculates only the optimal values of policy instruments.