There is growing evidence that tacit collusion can be autonomously achieved by machine learning technology. However, outlawing such conduct is pointless unless there are suitable remedies to address them. This article, available here, explores how fines, and structural and behavioural remedies, can serve to discourage collusive outcomes while preserving incentives to use efficiency-enhancing algorithms.
Section II provides a brief overview of the properties of deep learning methods and their applications to pricing decisions.
Different machine learning methods can be usefully deployed to make pricing decisions. Reflecting statistical analysis, machine learning has the ability to automate pricing decisions by using hundreds or thousands of variables in ways that would be otherwise unavailable to market participants. Reinforcement learning may allow firms to automate pricing strategies according to variables such as reactions by competitors and the impact this has on profits or market share. As an algorithm’s problem-solving capabilities improve, market prices will tend to become more stable and converge to a price level above the competitive benchmark.
Section III covers the theory of harm concerning algorithmic collusion that serves as a basis for the analysis of suitable remedies.
Tacit collusion is not illegal – even when achieved by AI-powered pricing software. Distinguishing tacit collusion from unlawful coordination is typically a difficult task. This difficulty has been one of the reasons, together with the equivalence of economic effects, why it has been argued that all types of oligopoly pricing should be outlawed. Economic theory has identified conditions under which oligopolistic markets can lead to supra-competitive prices without a need for overt communications between economic agents. In short, price deviations must be detectable and the threat of retaliation has to be credible. Therefore, market transparency, dispersed demand, frequent repeated interactions, stable demand and cost structure, and product homogeneity make collusion – tacit or express – more likely.
Would it make sense to prohibit tacit collusion? A broad prohibition on interdependent pricing would require firms to set a price that earns them normal profits. However, the difficulties of determining whether prices are too high, or profits abnormal, are such that competition authorities focus instead on whether a practice leads to collusion. On the other hand, establishing whether unlawful communications occurred carries its own administrative difficulties. First, there must be evidence that proves that behaviour patterns are inconsistent with competitive pricing. Second, the evidence has to show that the behaviour is the result of unlawful forms of cooperation. Despite this, it is commonly accepted that some form of communication has to occur for antitrust liability to arise.
The use of pricing algorithms changes the considerations surrounding some of these issues. Algorithms used in experimental settings have been shown to be able to learn collusive strategies without being specifically programmed to do so, while making oligopoly pricing patterns look more akin to competitive behaviour. Further, prior to the advent of algorithmic pricing, instances of pure interdependent pricing were circumscribed to a rare set of circumstances that include highly concentrated industries, homogeneous goods, symmetric cost structures across firms, and price transparency, among others. There are concerns that algorithmic pricing might make such circumstances much more common.
Section IV deals with appropriate remedies against algorithmic collusion.
An optimal remedy should effectively address and counter the particular harm caused by the pricing algorithm. Unlike the case of algorithms being used to monitor deviation from collusive arrangements, the focus here is on the use of pricing algorithms to extend and multiply the scenarios in which tacit collusion weakens the functioning of competition.
Fines might be used if they change an AI’s perception of the profit-maximising behaviour, but this may require AI to be able to identify competitive prices and the profits from the use of tacit collusion-inducing AI – and to compare them with the costs of fines. Further, given the usual source of AI instruments, the issue of whether software suppliers should be liable and fined needs to be discussed as well.
In the alternative, structural remedies can be more attractive than fines, particularly as they do not require one to determine competitive equilibrium. Instead, they can directly address market characteristics that facilitate collusive results by algorithms. Structural remedies like divestitures can be used to address concerns of tacit collusion in several ways. For example, a divestiture can aim either at creating a higher number of competitors, or at rearranging the competition dynamics between present market participants (e.g. by creating asymmetric competitors), so that tacit collusion becomes more difficult to achieve. One may also have recourse to behavioural remedies, e.g. on how algorithms are designed. Authorities could order firms to program their algorithms to play competitive instead of cooperative games. If the benign properties of algorithms can be left unharmed when software is programmed in such a way as to avoid collusive outcomes, then an injunction or an order to not price interdependently would be feasible. Simulations carried out in algorithmic incubators could, in theory, be used to police the type of game being played. One might also create guidelines on algorithmic transparency, to allow one to understand the collusive potential of an algorithm, while taking care to avoid this transparency leading to collusive outcomes.
In addition to antitrust, merger control and sector inquires could be effective instruments to address oligopolistic pricing. The use of particular pricing algorithms, and of algorithmic coordinated effects, should be considered as a factor in merger decisions. Further, structural remedies can be more easily and naturally deployed in merger control procedures. Regarding sector inquiries, where they can lead to the use of structural and behavioural remedies, they may also be useful in terms similar to those already discussed for antitrust enforcement.
This an openly speculative, yet thoughtful paper that focuses on possible reactions to tacit collusion flowing from the widespread adoption of AI. I emphasise that the solutions are possible for two reasons. First, the need to deploy such remedies has not yet been established, since there is yet no evidence of AI-driven tacit collusion being a problem. Second, were a remedy to be deployed, it would have to be appropriate to the anticompetitive practice it seeks to address, so it will necessarily have to reflect the specific situations in which algorithmic tacit collusion could amount to a competition infringement, a topic which the paper openly avoids.
This is not a criticism of the paper, which acknowledges these two elements, and focuses on the subsequent question of what remedies to apply if necessary. Further praise should be directed at the authors’ concern throughout the paper with the potential welfare-enhancing effects of AI, and the potentially detrimental effects of remedies.