Original version With this story appeared in Quanta Magazine.
Imagine a city with two gadget sellers. Customers prefer cheaper widgets, so sellers must compete to set the lowest price. Dissatisfied with their meager profits, they meet in a smoke-filled tavern one night to discuss a secret plan: if they raise prices together instead of competing, they can both make more money. However, this type of intentional price fixing, called collusion, has long been illegal. Widget sellers choose not to take the risk and everyone else can enjoy economical widgets.
For more than a century, American law has followed this basic template: back-office trading is prohibited and prices must be kept fair. Nowadays it’s not that straightforward. Across broad swaths of the economy, sellers increasingly rely on computer programs called learning algorithms that repeatedly adjust prices in response to novel data about market conditions. They are often much simpler than the “deep learning” algorithms that form the basis of current artificial intelligence, but they can still be susceptible to unexpected behavior.
So how can regulators ensure that algorithms set fair prices? Their time-honored approach will not work because it relies on finding clear collusion. “Algorithms definitely don’t drink together,” he said Aaron Rothcomputer scientist at the University of Pennsylvania.
Still a widely cited publication from 2019 showed that algorithms can learn tacit collusion even if they were not programmed to do so. The research team pitted two copies of a straightforward learning algorithm against each other in a simulated market and then allowed them to explore different strategies to enhance profits. Over time, each algorithm has learned through trial and error to retaliate when the other lowers prices by lowering its own price by a huge, disproportionate amount. The end result was high prices, supported by the mutual threat of a price war.
These types of hidden threats also underlie many cases of interpersonal collusion. So if you want to guarantee fair prices, why not simply require sellers to utilize algorithms that are inherently incapable of expressing risks?
IN recent articleRoth and four other computer scientists showed why that might not be enough. They have proven that even seemingly benign algorithms that optimize for their own profit can sometimes produce bad results for buyers. “You can still get high prices in a way that looks reasonable from the outside,” he said Natalia Collinagraduate student working with Roth who co-authored the novel study.
Not all researchers agree on the implications of this finding – much depends on how you define “reasonable.” But it shows how subtle questions about algorithmic pricing can be and how hard it can be to regulate.