Automated decisions, Coase’s theory of the firm, and how it all applies to companies like Uber.

Adam Gurri is the founder and Editor‐​In‐​Chief at Liberal Currents.

Modeled Behavior’s Adam Ozimek, no enemy of markets or decentralization, thinks that algorithms might give the price system a run for its money. He emphasizes the centralized, automated mechanism employed by a company like Uber, as against the decentralized knowledge system expounded by Hayek. In fact, this sort of argument has been made for decades, with Oskar Lange’s “The Computer and the Market” in 1967 going so far as to claim that computers could directly solve all the simultaneous equations required for general equilibrium. But there is no evidence that computation can replace the price system. Though it was not Ozimek’s intention to make such a bold claim, his argument could be extended to that logical conclusion. But the examples he picks are no different from the make‐​or‐​buy decisions that have always been present in markets, decisions about whether to purchase a good or produce it oneself.

Lange’s argument is unlikely to be taken seriously by anyone who isn’t an economist, and even then only by those who take general equilibrium models extremely seriously. The cornerstone of such models is that the economy can be represented as a series of equations which must all be solved simultaneously in order to bring the entire market into equilibrium; that is, so that all markets clear. But the idea that an economy can be represented by a series of equations is very different from the idea that it actually is a series of equations, which Lange’s piece seems to insist.

Lange’s earlier and more famous argument is that central planners can approximate market conditions by trial and error. They start with an arbitrary price, and if there is a surplus, they lower prices. If there is a shortage, they raise prices. But because they are thinking beings rather than mindless market orders, they can adjust for social ills such as inequality and class exploitation. This latter moral element aside, the trial and error approach better captures how algorithms are used today.

Consider Uber, Ozimek’s prime example of an algorithm displacing an auction. What Uber’s algorithm ostensibly does is set prices in such a way as to strike a balance between the supply of drivers and the demands of riders. They anticipate or respond to a shortage of drivers by charging higher prices, which also reduces the quantity of drivers demanded by riders. The algorithms set prices in the equal and opposite way when there is a driver surplus of the same magnitude. Thus, within its own sphere, Uber seems to behave like Lange’s trial‐​and‐​error central planner.

Tempting as it is to come to that conclusion, it is utterly false. The appropriate comparison here is not with hypothetical socialist systems, but with the make‐​or‐​buy decision faced by individuals and firms every day for as long as commerce has existed. It was this decision point that fascinated a young Ronald Coase, a fascination which drove him to write the foundational paper “The Nature of the Firm.” It was in this paper that he coined the term “transaction costs,” which boil down to the relative cost of buying rather than making. Like any good price theorist would, Coase argued that it was relative costs (and benefits) which drove each decision.

Rather than operating a Langean technocratic system of price controls, Uber has in reality helped to create a whole new market. Whatever California courts may say on the matter, the relationship of Uber the company to the drivers who pick up their users is not one of traditional employment inside of a firm. The drivers are independent contractors; effectively Uber is their client much more than the riders are. Just as a newspaper would have standard rates set for freelance writers, Uber sets the rates on their drivers’ productivity—a productivity measured in customers given rides, rather than articles. Customers can give feedback on specific driver performance, of course—but newspapers care about how a freelancer’s article is received, too. Uber may be the drivers’ client, but as such they have reason to care about how Uber’s customers perceive them.

In short, rather than subsuming everything into the make side of things through vertical and horizontal integration, Uber has actually increased the activity out on the open market. And as such, the price system is brought to bear, no less than if Uber set its prices through an individual human decision each time rather than an algorithm. Uber drivers often are also Lyft drivers, and many taxi drivers drive for both services. In Austin, Texas recently, I rode in an Uber car with a sign on the back of the driver’s seat addressed to both Uber and Lyft riders. The sign was advertising his bakery.

You see this dynamic everywhere that algorithms are brought to bear. Search engines came into existence entirely because of an abundance of cheap material to “buy” (though they did not literally have to pay for it). As more and more data has become proprietary, companies like Google have sought to increase vertical integration by buying services like YouTube or offering products like Android. At the same time, they have tried and failed to strike deals with companies like Twitter and Facebook to access their data—the transactions costs still appear to be too high for any such deals to last long.

Amazon both sets its own prices algorithmically, and offers a marketplace in which merchants can set their own prices. And its suggestion and search algorithms draw from both sorts of products. In the advertising world, Google offers an ad network that weights advertisers’ bids by an algorithmically determined probability that users will click on their ad. Meanwhile, supply side platforms determine winners simply by who bids the highest for a given pageview, but the demand side platforms that place the bids determine who to bid on and what price to offer largely on the basis of algorithms.

Where algorithms are concerned, we see many mixes of make‐​or‐​buy decisions; what we do not see is the eradication of that decision. So long as that decision remains foundational, so too does the price system. Nor is there any evidence that this is likely to change in the future. Harry Collins has done compelling work on the limitations of automation in general and computers in particular. He points to the myriad of ways that specifically human knowledge must always be brought to bear in order to get machines and computers to perform their functions successfully.

Computers, according to Collins, are a “social prosthesis.” “It only works because the surrounding social organism makes up and ‘repairs’ its deficiencies. These repairs usually take place without anyone noticing, which is probably why the standard analysis of these things goes so wrong so often.”

Consider Uber. If Uber’s algorithms set wildly inappropriate prices in a given situation—raising prices during a surplus or lowering them during a shortage—how would this situation remedy itself? Well, according to Time , “Demand surges have also been monitored by Uber’s human staffers, who have on rare occasions used their discretion to lower prices.” Human judgment is ultimately needed to overcome deficiencies in the algorithm. Google search would be a worthless product now if the very human engineers and product team hadn’t continued to make improvements, search for weak spots, and guard against spammers and malware.

Even if you get an algorithm that works correctly 99% of the time, it takes an error that shows up much less than 1% of the time to cause a lot of havoc. In the case of Knight Capital, it bankrupted the company is just 30 minutes. Can you imagine what an error like that would look like in an algorithm running a Langean computer economy?

I can: It would look like the bizarre and extreme resource misallocations that occurred in the USSR on a regular basis. Nothing would fundamentally have changed—we would be right back to the same old problems.

The Soviet Union constitutes possibly the largest experiment in seeing how far the choice to make rather than buy can be taken. And the results are instructive. They were unable to do away with the price system entirely. For one thing, they had to use their own centrally controlled imitation of it. For another, as the economist P. Wiles explained as far back as 1957, the Soviets were highly reliant on world prices to determine their artificial ones. If it wasn’t for the information generated by the global market, they wouldn’t have been able to operate their supposed alternative to markets at all. As it was, global prices could tell them next to nothing about local conditions, to which they were infamously terrible at responding to.

The dream of a perfectly logical order without prices or commerce should have died with the Iron Curtain, and all we have learned of what went on behind it. I am sure that with every new generation of computers and every new breakthrough in automation and data science, Lange’s arguments will be trotted out all over again, as they were in the early 90s. But like Ozimek’s piece, such arguments ignore the fundamental fact that no computer or algorithm has yet challenged the logic of make‐​or‐​buy decisions one iota. And such decisions only exist in the context of the price system, and of commerce as usual.