Michal Gal discusses the regulatory hurdles to deal with the impacts of algorithmic price collusion. In the meantime, she says, market fixes include algorithmic consumers and platform nudges to mitigate price coordination.
In the past few years our knowledge regarding algorithmic harms in the marketplace has greatly evolved. Today there is wide-spread agreement that algorithms can assist firms in putting in place a cartel, but, more importantly, that algorithms can autonomously learn to coordinate prices and set them at supra-competitive levels (algorithmic coordination). This awareness is based on theoretical, experimental, and empirical studies.
For example, one study of the German gasoline market found that when two firms in a duopoly market switched from manual to algorithmic pricing, the price increased substantially (9–28%). Together, these studies lead to an undeniable and credible conclusion: under some market circumstances, pricing algorithms can achieve coordination at supra-competitive prices without any human intervention or prior agreement, in their quest to achieve a general goal (e.g., “maximize the seller’s profits”).
The next frontier, therefore, involves finding ways to effectively deal with the negative welfare effects of algorithmic coordination, while not unnecessarily limiting the ability of suppliers to enjoy the benefits that the use of algorithms can bring. The exponentially growing use of pricing algorithms underscores the importance of this quest.
Existing legal tools are quite limited in their ability to deal with algorithmic coordination. This is because competition law prohibitions of coordinated conduct require the existence of an “agreement in restraint of trade”, interpreted as the offer and acceptance of an agreement not to compete. Accordingly, pure oligopolistic coordination—coordination which results from unilateral autonomous decisions of suppliers that take into account the expected reaction of their rivals to their actions when forming their own decisions—is not captured under the law, even though its effects on consumers are similar to those of an illegal cartel. The focus on the mode of communication may be partly explained by the difficulties in remedying pure oligopolistic coordination. This implies that as the use of sophisticated learning algorithms becomes more commonplace, more markets might engage in legal—yet harmful—algorithmic coordination.
Our toolbox is, however, not completely empty. In fact, it may be the right time to revive the use of a legal tool that is mostly dormant, which might provide a partial solution in some circumstances. The doctrine of “facilitating practices”/“plus factors” prohibits intended and avoidable acts that facilitate coordination by creating conscious commitments to a common scheme, and are not justified on procompetitive grounds. As elaborated in my research, such doctrines limit the ability of algorithms to coordinate, at least in some instances. Acts that raise red flags may include, inter alia, making it easier for competitors to observe one’s algorithms and/or databases in order to better or more quickly set an optimal coordinated price; or technologically “locking” one’s algorithm so that it is difficult to change, thereby increasing the commitment to the pricing scheme embedded in it.
Such practices could amount to plus factors, in that they may facilitate coordinated conduct; they are potentially avoidable; and they are unlikely to be necessary in order to achieve procompetitive results. They should thus trigger a deeper investigation into procompetitive justifications. The remedy in such cases is clear and easy to apply: prohibiting the facilitating practice. Furthermore, it is time to delve more deeply into the old doctrines, and determine when exactly price signaling crosses the legal line. Yet the doctrine does not capture the hardcore case of autonomous algorithmic coordination.
Merger review is another tool that can assist on the margins. In a recent article, co-authored with Dan Rubinfeld, we have suggested that merger review’s wide scope for inquiry, the fact that it is outcome-based rather than process-based, and the flexibility of its potential remedies, all increase its potential effectiveness in limiting mergers that increase the possibility of algorithmic coordination without offsetting benefits. To do this, some presumptions might need to be changed (such as with regard to the importance of asymmetry in the market to reduce coordination; the use of financial turnovers as a stand-alone basis for capturing mergers that lead to algorithmic coordination; or that coordination concerns only take place in markets with high levels of concentration). In addition, competition authorities must be aware of the potential for algorithmic coordination, where algorithmic coordination is already prevalent or is potentially profitable.
Calvano et al. have suggested dealing with algorithmic coordination by changing the law, to focus on the process that leads to coordination or its outcome, instead of on the mode of communication. This implies prohibiting the use of those lines in the code that produce a predictable coordinated outcome, while ensuring that the efficiency gains from using such algorithms are not lost (competition-by-design). The advantages of this solution are obvious: it goes to the root of the problem and the algorithm can be tested before it is put to use. Unfortunately, it also raises significant difficulties. One difficulty involves identifying the parts of the code that lead to coordination and distinguishing them from other parts of the code. The researchers suggest that to create certainty, competition authorities should test the algorithms in the lab before approving their use.
Yet such experiments often depend on the environment in which the algorithm is tested. Significant challenges thus arise regarding the market conditions authorities should take into account when testing the algorithm. In addition, it might be difficult to identify and separate the coordinating parts of the code in deep learning algorithms. But the more difficult problem involves the fashioning of a suitable remedy. The regulator would have to determine what weight, if any, firms should be allowed to give different factors, such as the prices set by rivals, in their decision-making. Economic theory, however, does not supply good answers as to how much weight should be given to rivals’ prices in order to set a price that is optimal for long-term consumer welfare. Regulators would thus need to tread carefully. For the same reason, simply prohibiting the use of all pricing algorithms or applying a “no fault” regulation, is problematic, although we might still identify some “red cases” in which the use of a certain type of pricing algorithm leads only to anti-competitive results.
Other suggestions have ventured even more outside the box, such as the introduction of a disruptive algorithmic seller. This idea is based on insights from economic theory that “noise”—(perceived) changes in market conditions which may change the optimal equilibrium—makes coordination more difficult. Accordingly, deployment of a disruptive algorithm, which is given the task of introducing noise, can potentially limit the ability of other algorithms to engage in coordinated conduct. The interference is external, and mimics the entry into the market of a maverick supplier that does not adhere to the coordinated equilibrium. Yet this solution has many limitations, such as requiring a high degree of intervention in the market, and an operation of an algorithm by the government or by a government-subsidized supplier.
As can be seen, there are no easy fixes for most instances of algorithmic coordination without changing the law and applying more interventionary measures. And even then, regulators should tread with caution, to enable suppliers to enjoy the pro-competitive benefits that algorithms bring. Many researchers and competition authorities are thus putting their heads together to fashion of better solutions, including a better use of existing ones.
In the meantime, the market can also devise its own solutions. While none provide a panacea, they might provide some respite. Let me offer two suggestions. The first involves creating a counterforce that would change market dynamics, in the form of algorithmic consumers. These are algorithms, operated by consumers, consumer groups, or third parties, that make purchase decisions on behalf of consumers and act as agents for buyers. Beyond the reductions they offer in search and transaction costs, algorithmic consumers can help limit algorithmic coordination in several ways. By aggregating consumers into buying groups, they can introduce another element into each supplier’s decision-making: the ability to supply a large quantity at lower price. This can potentially weaken the stability of the coordinated conduct. Alternatively, should algorithmic consumers represent a sufficiently large number of consumers, they could negotiate a deal outside the digital sphere. Also, by using AI, algorithmic consumers can test, devise, and apply other strategies to motivate suppliers to reduce prices. For example, they could decide not to buy beyond a certain price, thereby reducing consumers’ collective action problem.
Finally, by engaging in multi-homing, algorithmic consumers may reduce the extent of network effects, thereby potentially reducing the efficient size of market participants and creating more fragmented and contestable markets, which might be less prone to coordination. Indeed, we might already see this happening with the introduction of add-ons to ChatGPT that are designed to find the best offer on the web. Yet to function, algorithmic consumers need access to relevant data and access to potential users, which firms might try to limit, for example via limitations on viewing content by algorithms (e.g., “I am not a robot” requirements). Regulators should be aware of such imitations and attempt to deal with them effectively.
The second market tool involves nudges used by online marketplace platforms to mitigate price coordination between third-party merchants, by making deviation from a coordinated price more attractive. Take, for example, Amazon’s Buy-Box—the top-right box on the results page which allows for customer’s direct purchase and avoids a visual comparison with other offers available for that product. As most consumers buy the product offered in the Buy-Box, without checking other options, and only one offer is included in it, suppliers have a strong incentive to have their products included in it. This creates incentives for firms to compete over it. Furthermore, should the offer chosen be mainly based on price, and the platform may also compete over the Buy-Box, it can set the price for its own products relatively low, thereby nudging others to outbid it. This may be profitable to the platform, depending on its business arrangement with sellers operating on its platform. Governments can, in turn, nudge platforms to engage in such conduct.
Articles represent the opinions of their writers, not necessarily those of the University of Chicago, the Booth School of Business, or its faculty.