Can Computers Engage in Anticompetitive Behavior? Antitrust Questions Raised by Algorithmic Pricing
Businesses competing in the digital marketplace increasingly rely on sophisticated algorithms to monitor their rivals’ price movements and adjust their own pricing strategies accordingly. As the use of pricing algorithms has become more common, so has debate over whether allowing algorithms to set prices without human intervention can result in collusive or anticompetitive behavior.
To shed some light on how regulators and litigators might approach this question, Vice President David Smith spoke with Steve Tadelis, Professor of Economics and Sarin Chair in Leadership and Strategy at the University of California, Berkeley’s Haas School of Business. Professor Tadelis is a widely published expert on e-commerce and the economics of the internet, and has extensively researched online auctions and online bargaining, digital advertising, seller reputation and the determinants of trust, price salience, and algorithmic pricing.
Is there something new or different about algorithmic pricing that raises red flags with regulators and watchdog groups over potentially collusive behavior?
Steve Tadelis: Professor of Economics, Business and Public Policy, and Sarin Chair in Leadership and Strategy, Haas School of Business at the University of California, Berkeley
Although this is not a settled debate, I don’t think algorithmic pricing raises new and different concerns for regulators. A useful way to think about it is that the tool may be new, but the behavior isn’t. Companies have long relied on predetermined pricing rules or strategies to guide their pricing decisions across business cycles, and in reaction to changes in input prices, market conditions, or their competitors’ prices.
A pricing strategy could be as simple as an independent gas station’s owner looking across the street to see what prices a competitor was posting that day, and then meeting or beating that price by a predetermined amount. Or it could be as complex as combining cost input data and active monitoring of hundreds of competitor prices with consumers’ online browsing and purchasing data.
A key difference now is the enormous growth in the number of businesses using algorithms, instead of people, to monitor a variety of signals and automatically adjust pricing in reaction. The exponential increase in online activity, even before accounting for the effects of the pandemic, has made it easier and cheaper, and in some cases simply necessary, to use algorithms instead of people for monitoring prices of suppliers, competitors, and distributors.
And while we are good at tracking the growth in machine activity in place of human activity, it is not quite as clear where we can draw the boundaries between machine and human responsibility, particularly when the machine may have been built by someone else.
Does that raise questions about whether such activity is, in and of itself, anticompetitive?
Not necessarily – or perhaps a better answer is, “It depends.” On the one hand, as with the introduction of computer processing into countless other business processes, algorithmic pricing may, in fact, promote competition and enhance consumer welfare. The appropriate use of algorithmic pricing makes businesses better able to offer efficient and competitive prices to attract more customers, some of whom may not have participated in the market otherwise.
The other side of the coin is that algorithmic pricing potentially could also enable illegal collusion by unethical actors. If a collusive or previously coordinated agreement already exists, then, in theory, machines could be used to monitor and enforce (through price manipulation) the agreement, without requiring any human intervention, communication, or coordination after the fact.
This, in turn, has the potential for modifying the evidence available to prove the existence of collusion in the first place. The types of “smoking guns” may be different, thereby modifying the types of empirical analyses and reviews necessary to identify such collusion.
“[A]lgorithmic pricing is like any other tool – evidence of its use alone is not evidence of anticompetitive behavior. It all depends on how it’s used, which circles back to the question of human behavior, not machine behavior.”
Most of the focus from regulators and litigators has been on horizontal price-fixing between competitors, although interesting questions are starting to be raised about the potential for vertical price-fixing and for price discrimination. My own research was primarily focused on horizontal price-fixing, where two or more competing firms explicitly agree on a particular pricing strategy that is advantageous to the businesses involved but is divorced from market forces.
Successful collusion of this sort typically requires complex coordination between companies to agree on a pricing strategy, as well as to enforce compliance post-agreement. Here is where algorithmic pricing could be a tool used by colluders to automatically coordinate price changes as expected demand changes. The companies in the agreement may also be able to detect and punish deviations from the collusive pricing strategy without explicit communication between the parties.
What should regulators or litigators be looking for?
To date, traditional tools of competition policy, such as the Sherman Act, have been capable of dealing with cases involving algorithmic pricing. However, the empirical models that are needed to investigate anticompetitive concerns may need to be revised. For example, to examine whether a given pricing algorithm is capable of tacitly colluding, it may be necessary to understand the code underlying the algorithm or to test the algorithm in a simulated environment.
Importantly, regulators should not ignore that algorithmic pricing provides many benefits, such as reducing costs of pricing decisions and incorporating information on variations in supply and demand.
We know that questions have also begun to be raised as to whether the use of algorithmic pricing might allow or support tacit collusion – that is, whether algorithms can be written by different parties so that they autonomously interact or coordinate with each other to achieve an anticompetitive effect, but without the parties making an explicit agreement to do so.
Whether such a scenario is at all practical is still very much open to debate. Much of the research to date suggests that implicit collusion of this sort may be beyond the capabilities of algorithmic communication, given the massively complex nature of coordinating a collusive agreement.
On the one hand, very recent research has shown that pricing algorithms driven by artificial intelligence may be able to arrive at stable, collusive outcomes in certain circumstances. On the other hand, the setup and simulations of this new research are significantly simpler than the real world, and attempts to complicate the setting with more sellers either are intractable, or suggest that collusive outcomes become significantly harder to support. The question of how, or even whether, to regulate different uses of artificial intelligence has been receiving a great deal more attention recently, and is likely to receive a great deal more in the coming years.
You mentioned that these types of concerns are beginning to be raised about vertical price-fixing as well.
In resale price maintenance, or RPM, which is the most commonly used mechanism for vertical price-fixing, prices are held above thresholds dictated by the supplier. These arrangements are not stable unless all distributors comply, and so algorithmic price monitoring can be a cheaper and more effective way for both suppliers and distributors to identify and report pricing that deviates from the arrangement.
This is even more of a grey area than horizontal price-fixing, because RPM may very well have procompetitive justifications, such as solving the free-riding problem. For example, an online retailer may be benefiting from in-person services provided by its brick-and-mortar competitors while it offers lower prices. To date, we haven’t seen evidence to suggest that algorithms are more likely to be used to enforce anticompetitive RPM agreements than procompetitive ones. The body of evidence is likely to grow, however, as the use of algorithmic pricing increases across retailers.
Regulators have raised concerns about algorithmic pricing with respect to vertical price-fixing, but in the absence of horizontal collusive efforts, the effects of vertical agreements may be more likely to be weakened than enhanced by algorithmic pricing, which can quickly exploit departures from a competitive outcome.
From an antitrust perspective, then, should we be concerned?
There is no categorical answer to that question. It simply is too early to be able to say definitively. What I can say, however, is that algorithmic pricing is like any other tool – evidence of its use alone is not evidence of anticompetitive behavior. It all depends on how it’s used, which circles back to the question of human behavior, not machine behavior.
Whether the use of any particular pricing algorithm facilitates collusion or reduces consumer welfare is a question that should continue to be examined empirically for now. Ultimately, to answer this empirical question you have to look at the actions of the competitors, at the algorithms themselves, and at the net effect on prices, output, and innovation. However, we should continue to adjust our empirical methods to review pricing algorithms and to distinguish procompetitive uses of pricing algorithms from approaches that raise anticompetitive concerns.
Given the increasing pace of progress in artificial intelligence, pricing algorithms are sure to be a subject of continued study and debate. ■