Algorithmic Collusion

One of the most interesting aspects in the relationship between AI and antitrust law is the risk that algorithms – and, in particular, pricing algorithms, increasingly used by undertakings to determine the best pricing strategy in real time – may facilitate collusive behaviors between competing undertakings and give rise to new forms of anticompetitive coordination.

 

Explicit collusion and tacit collusion

The term “collusion” refers to any form of coordination among competing undertakings with the objective of raising or maintaining profits above the level they would reach in a competitive scenario, i.e. in the absence of the collusive behavior.

It is common to distinguish in economic science between “explicit collusion“ and “tacit collusion“. The former term refers to collusive behaviors that are the result of agreements or concerted practices between undertakings, ie the conscious and voluntary coordination. The second term refers to forms of coordination based on parallelism of behaviors by competing undertakings that, although conscious (conscious parallelism), is the result not of agreements but rather of autonomous choices.

In principle, antitrust law does not prohibit parallelism of behaviors between competing undertakings (tacit collusion), but prohibits anticompetitive agreements (explicit collusion), which are prohibited at EU level by Art. 101 TFEU and at national level by Art. 2 of Law No. 287/1990.

 

The risk of collusion generated or facilitated by the use of algorithms

The main risk from an antitrust point of view connected to the use of algorithms is that they are capable of facilitating collusive behaviors between competing undertakings or making new forms of coordination possible; in some cases, even in the absence of the prior programming of the algorithm to achieve the collusive outcome.

It is therefore usually made reference to “algorithmic collusion” and the economic and legal literature has identified four scenarios: (i) monitoring algorithms, (ii) parallel algorithms (iii) signalling algorithms and (iv) selflearning algorithms.

 

Four possible scenarios of algorithmic collusion

Monitoring algorithms

In this scenario, the case is that each undertaking participating to a price cartel can verify, through the use of an algorithm duly programmed to monitor the prices applied on the market, that the cartel has been effectively implemented by the other participants (ie verify that other undertakings participating to the cartel actually apply the agreed price) and detect possible deviations. In this case the algorithm would act as an instrument facilitating the stability of an anticompetitive cartel.

Parallel algorithms

It is noted in economics that one of the difficulties in implementing an anticompetitive cartel in highly dynamic markets is the fact that continuous and sudden changes of the market conditions require similar frequent adjustments of the price agreed within a cartel and thus continuous communications between the cartel participants to agree on new prices. Such communications increase the risk that the cartel may be detected (and thus investigated) by the competent authorities. In this scenario it is deemed that undertakings may use pricing algorithms with the aim of determining simultaneous and parallel reactions to changes of the market conditions, reproducing a scenario of conscious parallelism of behaviors. In this context, competitive concerns may arise if undertakings agree to programme price algorithms not with the purpose of competing with each other, but to set their prices in a coordinated manner at a supracompetitive level.

Signalling algorithms

Competing undertakings can reach a common understanding to coordinate their conducts without explicit communications, by means of unilateral signals and announcements concerning the commercial terms they intend to apply and by aligning their conducts on the basis of such announcements. But undertakings might not observe the signals sent by competitors or voluntarily decide not to align their behavior with the competitors’ announcements. In this scenario, it is deemed that appropriately programmed algorithms could favor an automatic alignment of the behavior of the undertakings with the announcements of competitors. Each undertaking could transmit in this way – through signalling algorithms – continuous signals concerning, for example, the price it intends to apply. When all undertakings concerned send the signal announcing the same price, each firm’s algorithms process the competitors’ signals and ensure that the firms apply the price of the message concerned. This corresponds to an agreement between the undertakings to apply that price.

Self-learning algorithms

Finally, it may be the case that using machine learning and deep learning technologies, a highly sophisticated algorithm can itself make autonomous business decisions, based on the analysis and processing of market data. This is the case of self-learning algorithms, programmed to maximize companies’ profits, learning autonomously, without any human intervention, that, at least in certain market contexts, the most effective way to achieve this goal is to coordinate business conducts with those of competitors. The use of this kind of self-learning algorithm could lead to a collusive outcome that occurs without the algorithm having been programmed to implement a restrictive cartel or facilitate its implementation.

 

Self-learning algorithms and antitrust law

The Fact-finding Survey on Big Data jointly carried out by the Italian Competition Authority, the Italian Communications Authority and the Personal Data Protection Authority noted that “the spread of procollusive pricing algorithms can facilitate the stability of cartels and the creation of market contexts leading to collusive equilibrium”.

The Fact-finding Survey points out that, in presence of sophisticated algorithms, characterized by machine learning mechanisms, it‘s very difficult to identify the “decisive ingredient for an infringement of Article 101 TFEU” – ie the “exchange of wills” between competitors aimed at agreeing and coordinating a given commercial practice, although it highlights that the investigation of anticompetitive agreements facilitated by the development of "sophisticated algorithms" is one of the priorities of the Italian Competition Authority’s activity.

The first three algorithmic collusion scenarios above can be more easily brought within the perimeter of application of the anticompetitive agreements’ prohibition, since in such cases the algorithms help implement or facilitate an anticompetitive cartel. Greater doubts arise with reference to the collusive outcome achieved through the self-learning algorithms, since it does not constitute the result of an “exchange of wills”.

It is the latter case that gives rise to the most relevant competition concerns.

A self-learning algorithm programmed to define the best pricing strategies to maximize the firm’s profit – which is per se a legitimate goal – may learn that one of the ways, if not the most effective way, to achieve this goal is to align its price with that of its competitors on supra-competitive values.

The widespread use of self-learning algorithms could lead to collusive outcomes even in markets not particularly concentrated, taking into consideration the high capacity of AI systems to quickly process market information (such as possible price deviations of competitors from the supra-competitive price) and the ability of the algorithms to determine in real time the best pricing strategy (ie one that maximizes the undertaking’s profits) on the basis of this information.

In the absence (for the time being) of explicit provisions expressly regulating the use of algorithms and sanctioning algorithmic collusion, the main issue at stake is: (i) whether the competition rules suffice to detect and prevent algorithmic collusion conducts detrimental to competition, eg by interpreting extensively the concepts of “agreement” and “concerted practice” or applying the provisions on the collective dominance and its abuse; (ii) whether it is necessary to adopt new provisions and, if so, what type of provisions (eg provisions which expressly identify algorithmic collusion capable of constituting a competition infringement or provisions which set forth certain requirements and conditions for the use of the algorithms and appropriate controls to prevent collusion).