I. Introduction
The use of algorithms is associated with a risk of collusion. This bears on the construal of the cartel prohibition, on which the present chapter focuses. The hypothesis is that algorithms may achieve a collusive equilibrium without any involvement of natural persons. Against this backdrop, it is questionable whether and to what extent such an outcome can be qualified as a concerted practice in terms of the law.
The analysis will be structured as follows: first, it will be assessed in what way algorithms can influence competition on markets (Section II). Subsequently, the article will deal with the traditional criteria of distinction between explicit and tacit collusion, which might reveal a potential gap in the existing legal framework with respect to algorithmic collusion (Section III). Finally, it must be analyzed whether the cartel prohibition can be construed in a way that captures the phenomenon appropriately (Section IV). The chapter will close with a summary (Section V).
II. Algorithmic Collusion As a Phenomenon on Markets
It is widely accepted that the use of algorithms can precipitate collusive outcomes, at least in theory. There is no lack of attempts to systematize the different ways algorithms can be involved here. Since the first groundbreaking publications by Ariel Ezrachi, Maurice Stucke, and Salil Mehra, as well as other authors in the following, the matter has come into the focus of antitrust scholarship and practice.Footnote 1 Agencies have started to look into it.Footnote 2 First cases have emerged about restraints implemented on platforms involving computer technology. With respect to the United States (US) TopkinsFootnote 3 must be mentioned, which involved alleged horizontal price fixing on a digital platform based on algorithms. For the European Union (EU), the EturasFootnote 4 case comes to mind. The operator of a travel booking platform had informed the travel agencies using that platform that it intended to cap rebates granted to end-consumers. The European Court of Justice (ECJ) held that this amounted to a horizontally concerted practice among the travel agencies to the extent that these had not objected to this proposal. The Luxembourg competition authority in 2018 found the taxi booking app Webtaxi to be exempt from the cartel prohibition although the system involved an algorithmic horizontal alignment of prices. The agency found offsetting efficiencies to the benefit of consumers.Footnote 5 These cases have in common that the natural persons representing the companies involved were aware of the restrictions, or that they at least ought to have known about them. Algorithms were an element of implementing the restriction, yet the ultimate decision about the competitive restraint was taken by human beings. Under these conditions the cases did not pose severe difficulties in establishing explicit collusion. There is not a fundamental difference to cases in which parties communicate, for example, by way of hub-and-spoke cartelization through traditional means and forms of communication.Footnote 6
A greater legal challenge is caused by the risk of autonomous algorithmic collusion. Computers with machine learning capabilities can possibly achieve or sustain a collusive equilibrium without any involvement of human knowledge or intent. The underlying scholarly discussion usually orbits around q-learning mechanisms.Footnote 7 The hypothesis is that algorithms with machine learning capabilities can act as computer agents exploring the success of their own actions, from which a collusive strategy can emerge as the optimum. In this event, it is conceivable that even the programmer of the algorithm was not aware of this potential outcome.Footnote 8 The market effect, therefore, can be a collusive equilibrium, albeit absent any human involvement.
Lawyers, economists, and computer scientists are still at odds over the likeliness and actual occurrence of autonomous algorithmic collusion. Some consider it a realistic scenario that tends to be underestimated.Footnote 9 The German Bundeskartellamt and the French Autorité de la Concurrence have refrained from a definitive conclusion so far.Footnote 10 Ulrich Schwalbe, in his seminal article, points out that the game theoretical dilemma that has to be solved by autonomous computer agents to achieve a stable collusive equilibrium is huge and cannot be easily overcome in practice.Footnote 11 A more recent study by Emilio Calvano and othersFootnote 12, however, concludes that q-learning algorithms, in fact, can autonomously collude. The EU Commission, in its proposal for a ‘New Competition Tool’, mentions the risk that digital platforms can create ecosystems in which collusion arises, which can be read as a recognition of the phenomenon as a matter of concern.Footnote 13 Against this backdrop, it is expedient to elaborate further on the application of the cartel prohibition in such cases.
III. On the Scope of the Cartel Prohibition and Its Traditional Construal
The conceptual problem behind the traditional construal of the cartel provision is the difference in the structure of the law on the one hand and the economic determinants for collusive equilibria on the other.Footnote 14 Anticompetitive collusive equilibria are characterized by the fact that the participants consider it individually rational to pursue such a strategy. That is the case for types of collusion that are usually referred to as explicit and which are illegal in the same way as it holds true for so-called tacit collusion, which is seen as not to fall within the scope of the prohibition. Agreements in breach of the cartel prohibition are null and void so that they cannot be enforced. Therefore, any cartel is only stable so long as the firms participating in it consider it rational to remain involved. All types of collusive equilibria can, therefore, be considered as non-cooperative games in terms of game theory.Footnote 15 Yet still, the law does not prohibit the achievement or sustaining of a collusive equilibrium as such. Instead, the provision is confined to certain types of measures, which are described as agreement, concerted practice, or decision. Mere tacit collusion is supposed to be distinct from these aforementioned types of anticompetitive conduct. While explicit collusion (i.e. achieved by agreement), concerted practice, or decision, is prohibited, tacit collusion is found to be legitimate.
As becomes obvious, the traditional construal of the law rests on a description of the means and forms by which firms interact when defining collusion. In the context at hand, the category of concerted practices has the greatest relevance. It is conceived of as something whereby a ‘practical cooperation’ is substituted for the risks of competition. Such practical cooperation, in turn, is supposed to be different from merely observing a rival’s conduct and reacting to it.Footnote 16 Similar approaches of distinction apply under section 1 of the US Sherman Act.Footnote 17 There, so-called conscious parallelism is deemed not to fall within the scope of the prohibition.Footnote 18 For the finding of a cartel, so-called plus factorsFootnote 19, or ‘facilitating practices’/‘facilitating devices’ need to be established.Footnote 20
It is argued that, whereas firms when tacitly colluding merely observe each other and react independently, the mechanism allegedly differs if they opt for a practical cooperation. A private exchange of pricing information can serve as an example for the latter. The jurisprudence of the courts requires that such practical cooperation must be substituted ‘knowingly’ for the risks of competition for it to amount to a concerted practice.Footnote 21 The concept, therefore, hinges on the inner sphere of the firms involved.
Against the afore, it is questionable whether autonomous algorithmic collusion is prohibited under the traditional enforcement paradigms. If the firms lack of knowledge or intent with respect to the fact that their computer agents pursue a collusive strategy, it cannot be said that these firms ‘knowingly’ substitute a practical cooperation for competition. Several authors, therefore, point to the risk that the cartel prohibition might stop short of preventing such outcomes. Ezrachi and Stucke argue that collusion achieved by machine learning systems can fall outside the scope of the cartel prohibition for ‘lack of evidence of an anticompetitive agreement or intent.’Footnote 22 In a similar vein, Calvano and others conclude from their studyFootnote 23:
From the standpoint of competition policy, these findings should probably ring an alarm bell. Today, the prevalent approach to tacit collusion is relatively lenient, in part because tacit collusion among human decision-makers is regarded as extremely difficult to achieve. While we have no direct comparative evidence for algorithms relative to humans, our results suggest that algorithmic collusion might not be that improbable. If this is so, then the advent of algorithmic pricing could well heighten the risk that tolerant antitrust policy will produce too many false negatives.
Some authors, therefore, highlight that the enforcement paradigms might warrant amendments to close such regulatory gaps.Footnote 24
IV. Approaches for Closing Legal Gaps
1. On the Idea of Personifying Algorithms
It is questionable whether the potential enforcement gap in antitrustFootnote 25 can be overcome by defining algorithms as ‘undertakings’. The notion of an undertaking in EU antitrust, indeed, is a very broad concept which has many facets and functions. As a working definition that applies to the most common types, an undertaking can be described as a combination of assets and people that act on a market governed by a management body or further representatives irrespective of the legal personality or corporate structure. For an undertaking to act, or to have knowledge or intent, it is the action, the knowledge, or the intent of the human beings representing it which is attributed to it. If a company manager knowingly enters into an exchange of sensitive pricing information with the manager of a rival, it can, therefore, be said that these ‘undertakings’ substituted a practical cooperation for the risks of competition.
Yet what substantive meaning would terminology such as a ‘practical cooperation’ or ‘knowingly’ have, if they were applied to an algorithm? The problem is that the legal terminology is coined on human interaction and cognition, so that it will be vastly deprived of its meaning if transferred to a computer system. How shall an action of an algorithm be identified as ‘knowingly’, as opposed to another action that is supposed to happen un-knowingly? In what way is it meaningful to consider the actions of an algorithm as a ‘practical cooperation’, as opposed to a mere intelligent adaption to information obtained by this algorithm on the market? To rely on such human concepts of cognition with respect to the regulation of algorithms will likely end up in semantic exercises with limited substance.
2. On the Idea of a Prohibition of Tacit Collusion
Another way of dealing with the problem could be in equating tacit collusion and explicit collusion.Footnote 26 This would mean that under antitrust law any collusive strategy would qualify as an illicit cooperation so that any further distinctions based on the inner sphere of the persons involved would become obsolete. Such view has been suggested in the past independently of the issue of algorithmic collusion. Notable proponents were Richard Posner in his earlier writingsFootnote 27 (he has since changed his viewFootnote 28), Richard MarkovitsFootnote 29, and more recently Louis Kaplow.Footnote 30
One might feel inclined to hold that such a view does not reconcile with the structure of the law. Yet it is questionable whether this counterargument would be very strong. The notion of concerted practices can possibly be construed in a way as to extend to cases in which the collusive outcome is based on a mechanism of observation and retaliation, which is characteristic for tacit collusion as it is for explicit collusion. As outlined earlier, both categories share the feature that they can be described as a non-cooperative game in terms of game theory. It is, therefore, rather a semantic issue whether some ways of engaging in such a non-cooperative strategy can be tagged as a ‘practical cooperation’ or not, while the underlying economic principles remain the same. Firms observe and react in ways that are deemed individually optimal irrespective of which words are used to describe this phenomenon. Especially in the grey area between typical cases of explicit cooperation on the one hand and tacit oligopoly conduct on the other, it becomes apparent how brittle the traditional concept of distinction is.Footnote 31 Consider that even in cases that would usually be qualified as tacit collusion, firms cooperate in that they observe each other and react to the information they have obtained from observing each other. One might ask the question: In what way is this not a ‘practical cooperation’? Also, a collusive equilibrium can bear negatively on consumer welfareFootnote 32 irrespective of the means and forms used by firms to sustain it. This seems to add to the argument that the distinction between tacit and explicit collusion is of limited expedience.
Yet still the idea of equating tacit collusion with explicit collusion faces severe objections, which might explain why it has not gained more recognition among scholars and enforcers.Footnote 33
Any law must provide the addressee the opportunity to abide by it through choosing a compliant course of action. Otherwise, the law would be perplexing. Recall, now, that collusion in the realm of antitrust is a non-cooperative game.Footnote 34 This means that the strategy is individually rational for each participant. Sanctioning collusion without more, therefore, would amount to prohibiting the pursuit of an individually rational strategy. Kaplow reacts to this objection by pointing out that it is a common feature of the legal order to prohibit types of conduct that are individually rationalFootnote 35, such as the stealing of an apple.Footnote 36 By imposing a sanction, the law ensures that it becomes rational to refrain from that course of action, he advances.
This argument, however, is not able to overcome the conceptual problems that would arise if the law prohibited the collusive outcome as such. While it is perfectly clear what a person must do to ‘not steal an apple’, it is much less obvious what course of action a firm would have to take in order to ‘not pursue a collusive strategy’. The negatory of stealing an apple is unambiguous. The course of action in order to avoid the sanction is to not steal the apple. With a prohibition of a collusive market outcome, however, it would be much more difficult to describe, in an unambiguous way, what the compliant course of action would be. One of the reasons for this is that, from an ex ante perspective, it is unclear what outcome a collusive strategy among firms might produce. If that is not known, however, it is equally unclear what price a firm must set in order to not charge at a collusive level. If the firms do not know whether a collusive strategy would yield a market price of € 5 or rather of merely € 4, they cannot know ex ante whether individually charging a price of € 4 would be ‘non-collusive’ or not.
On a more philosophical level, another objection buoys. Effectively, the law would require the addressee to pursue any market strategy so long as it is not rational, assumed that the rational strategy would be collusion. It is questionable, however, whether an addressee of the law can be required to intentionally act irrationally. Any legal prohibition must conceive of the addressee as an intelligent entity, for if the addressee were not intelligent, it could not abide by the law in the first place. From a philosophical perspective, it is unclear, however, whether a person can ‘rationally act irrationally’. Can a firm be obliged to randomize prices in order to protect itself from being accused of acting rationally-collusive?
Yet even ignoring this fundamental problem and moreover assuming it were possible to define a hypothetical collusive price level ex ante, it would remain unclear whether a non-collusive strategy could be defined with a sufficient degree of precision so that firms could abide by the law. Would it suffice to undercut the hypothetical collusive price by, for example, 2%? Or would the law require every firm to price at marginal cost, for under perfect competition this would be the hypothetical competitive price, even though perfect competition usually does not exist? Merely imposing a sanction on the pursuit of a collusive strategy does not solve any of these conceptual problems.Footnote 37
3. Harmful Informational Signals As Point of Reference for Cartel Conduct
a. Conceptualization
Another conceptual venture to solve the issue could be to maintain a distinction between illicit cartel conduct and legitimate coordination, yet to substitute a new criterion for the traditional paradigm, viz. for the practical cooperation adage. Such alternative criterion of distinction could be checking the informational signals released by colluding firms for their propensity to create a net consumer harm. Accordingly, an illicit concerted practice would be found if firms, or the algorithms relied on by firms, released informational signals which reduced consumer rent if compared to a counterfactual in which such informational signals were absent. The counterfactual, therefore, would not be a hypothetical market without collusion. It would be the same market absent a particular informational signal.Footnote 38 To clarify the specifics of this approach, and to contrast it with the view expressed by Kaplow and others, the following explanations shall be made.
In contrast to a situation where any collusive equilibrium is prohibited, so that it is unclear which course of action to take in order to abide by the law, confining the prohibition to harmful signals leaves the addressee a binary choice that is conceptually clear: an informational signal that is ultimately harmful to consumers can either be released or not. There is no element of imposed irrationality in such a prohibition. To refrain from the release of a harmful informational signal can also mean that the addressee must choose to release a different signal in order to avoid a harmful effect. To exemplify the idea, reference can be made to the EU Commission’s remedy decision in Container Shipping, where public price announcements were not abandoned completely but limited in scope in order to avoid the creation of a consumer harm.Footnote 39
A relevant signal in this sense can be any release of market-related information independent of whether it takes place publicly or in private, whether consumers are involved or not. Even price lists, price announcements, or other types of public display of prices or other parameters can qualify as a signal that ultimately leads to a collusive equilibrium. In contrast to the aforementioned opinion of Kaplow, however, this would not suffice for the conclusion that a restriction of competition in terms of the cartel prohibition is established. Rather, the release of such informational signals would fall outside the scope of Article 101(1) TFEU to the extent that it produces offsetting consumer benefits.
This reflects the fact that there is a plethora of cases where public market information leads to collusive equilibria, although the benefit to consumers being derived from this information outweighs the gross harm of the collusive outcome if compared to a situation in which such informational signals were not made. The reason why, in most cases, public price lists and similar forms of information are not prohibited despite their risk to precipitate collusive equilibria lies in the fact that they usually bring about benefits to consumers that offset the potential harm.Footnote 40 If firms refrain from the release of such pricing signals, the distribution of goods might be significantly impeded. Consumers can face difficulties in planning their purchases ahead or in pursuing multisource strategies. It can be said, therefore, that sometimes ‘vertical transparency’ creates a greater benefit to consumers than what is taken away by the horizontal collusion that concomitantly results from it. It is necessary, therefore, to balance both effects in order to get a full picture of the economic impact of an informational signal. If such an analysis of the economic effects is integrated into the assessment of a concerted practice, this creates a system in which harm and benefit of a measure can be distinguished without the need to make recourse to notions such as the ‘practical cooperation’ adage or the intention of the firms involved.
This concept would allow an entity to deal with the phenomenon of autonomous algorithmic collusion under the cartel prohibition. Achieving or sustaining such a market outcome would be prohibited if and to the extent that the harm to consumers were greater than the benefits associated with the release of the underlying informational signals.Footnote 41 If algorithms achieved a collusive equilibrium by communicating with each other and without this interaction providing any useful information to consumers, this could, therefore, constitute a concerted practice in terms of the law. If, on the other hand, a digital platform aggregated and provided information to consumers in a way that benefits them, and if the efficiency potential of this platform could not materialize without this release of information, Article 101(1) TFEU would not be triggered even though the conduct might, ultimately, give rise to a collusive equilibrium, if and to the extent that the benefit of the former offsets the harm of the latter. The counterfactual, therefore, would not be a digital platform without collusion. It would be a situation in which the informational signals were not released. If the platform, in such a counterfactual scenario, were not operated or operated with a lesser efficiency, consumers could be deprived of the benefits resulting from it. This could mean a smaller range of suppliers being visible to them, less information being available to help consumers plan their purchases, etc.
This demonstrates that a collusive outcome can be an inevitable consequence of an algorithm-based system that produces benefits, although horizontal restraints, even on prices, are involved. The decision of the Luxembourg competition agency has made this clear with respect to the taxi app Webtaxi.Footnote 42 The agency found consumer benefits in the fact that the platform improved on the supplies with transportation services, even though a horizontal collusion on prices was an inevitable side effect. While the decision accounted for these efficiencies within the scope of an exemption from the cartel prohibition, the concept presented here would integrate this analysis already into the assessment of a concerted practice. As previously outlined, this is a consequence resulting from the substitution of an effects analysis for the less useful ‘practical cooperation adage’ in order to discriminate between legitimate and illicit collusion in algorithm cases.
b. Possible Objections
Such a concept of harmful informational signals, of course, provokes objections. They shall be dealt with in the remainder of this chapter. One might want to invoke that this construal of the law does not reconcile with the structure of the provision as shaped by the jurisprudence. Admittedly, the courts have not yet recognized an interpretation as suggested here. Rather, the Court of Justice, currently, relies on the notion of ‘practical cooperation’ in order to distinguish between concerted practices and tacit collusion. The established set of criteria does not contain a place for an economic effects assessment. On the other hand, the Court of Justice has not yet had an opportunity to hone the law with respect to the phenomenon of autonomous algorithmic collusion. It is, therefore, conceivable that the courts, upon preparatory enforcement steps done by the agencies, ultimately consider the option to readjust some of the enforcement paradigms in order to close potential regulatory gaps.
Also, it should be noted that even under the current decisional practice an effects analysis can be part of the assessment of Article 101(1) TFEU. As to the distinction between restrictions by effect and those by object, the potential of the measure to produce restrictive effects bears significance.Footnote 43 In that regard, the Court of Justice has made clear that, among other things, it must be analyzed whether and in what way consumers might suffer from the measure at stake.Footnote 44 This line of reasoning already mirrors an effects analysis based on the consumer welfare paradigm, which also is the backbone of the present proposal. In a similar way, the Commission practice demonstrates that the notion of a competitive restraint involves an analysis of the effects on consumer rent. Even though the Court of Justice has ruled that, for a restriction of competition to arise, it is unnecessary to demonstrate concrete consumer harmFootnote 45, the Commission takes this into account if it helps to discriminate between legitimate and illicit conduct. The Commission, for example, argues that even horizontal price fixing, depending on the structure and function of the cooperation, might warrant a close examination within the ‘by effect category’, which demonstrates that it is possible to conceive of cases where such conduct falls outside of the scope of Article 101(1) TFEU without any further examination of the legal exemption rule in Article 101(3) TFEU.Footnote 46
On a practical level, one might want to invoke that the assessment of whether an informational signal precipitates a net consumer harm or not is too complicated to rely on as an enforcement paradigm. Yet it must be noted that even the currently existing decisional practice is not void of elements of an effects analysis, as demonstrated previously. Beyond that, firms and enforcers face exactly these difficulties already within the realm of Article 101(3) TFEU, so that it could not be said that such difficulties are idiosyncratic to the proposal made here.
Beyond that, one might raise the question in what way firms can become responsible for the conduct of an algorithm, if the strategy pursued by the latter is unknown to the former. In legal terms, however, it is possible to hold someone responsible for the organization of an enterprise. Firms, therefore, could be considered obliged to terminate the use of an algorithm or to alter its paradigms, if and to the extent that it produces more harm than benefit to consumers. Firms could, therefore, be ordered, by way of an administrative decision, to make such adjustments to their business strategy either by tweaking the algorithm or by stopping its use altogether. Antitrust economists already expound ways to design platforms in a way that counters collusive risks emerging on it from machine learning systems. Justin Johnson, Andrew Rhodes, and Matthijs Wildenbeest published a study in 2020 on how a choice algorithm on a sales platform can impact the likeliness of collusion among independently acting q-learning algorithms.Footnote 47 It goes without saying that the imposition of a fine, or a liability for damages would, in any event, require negligence or intent on the firm’s part. That, in turn, would require the agency to establish that some degree of knowledge or intent with respect to the achievement of a collusive equilibrium can be established among the actual firms that relied on the algorithm. Such would be absent if the equilibrium were achieved or sustained independently by a machine learning process not guided or anticipated by the firms that rely on the outcome.Footnote 48 Yet still, in such an event a mere administrative order, absent any sanction or damages award, could be issued.
Finally, one might want to invoke that it would be disproportionate to make such far-reaching amendments to the construal of the cartel prohibition for the sole purpose of closing a regulatory lacuna with respect to algorithms. It is not the intention behind this proposal, however, to render the established enforcement paradigms obsolete in their entirety. Rather, the suggestion made here should be conceived of as a mere addition to the established principles that would still bear relevance in the majority of non-algorithmic collusion cases. A private exchange of pricing information between the managers of two rivals, for example, could still be considered a practical cooperation without further effects analysis required, for its potential to precipitate a consumer harm, as reflected in the Commission’s Horizontal Guidelines.Footnote 49 The informational signal approach suggested here, on the other hand, could become relevant if, in an algorithm case, the traditional criteria did not allow the application of the law in a meaningful way. The present suggestion, therefore, is meant as a humble contribution to existing paradigms, not as a postulate of a total substitution for them.
V. Conclusion
This article is a conceptual sketch of a way to deal with the intricacies coming along with autonomous algorithmic collusion. Such risk is being discussed, especially, with respect to q-learning algorithms. Even though practical cases have not yet emerged, there is sufficient reason to address potential issues as a precautionary measure at this point in the scholarly debate. It was the intention to demonstrate that the traditional construal of the law, which relies on a description of human behavior, appears inapt for effectively tackling machine-induced equilibria. Applying the established criteria in such cases would very likely lead to a harping on words without substance. There is simply no point in venturing to assess whether algorithms ‘practically cooperated’ as opposed to ‘merely observed each other’. The present proposal, therefore, is intended to serve as a conceptualization of the notion of concerted practices for cases that would otherwise elude the cartel prohibition. To conclude, the entire problem of autonomous algorithmic collusion is an example of the necessity for interdisciplinary research between lawyers, economists, and computer scientists. At the same time, the problem highlights how enforcement paradigms, that hinge on descriptions of the inner sphere and conduct of human beings, may collapse when applied to the effects precipitated by independent computer agents. The subject matter of this chapter is, therefore, an example for the greater challenges that the entire legal order faces in light of the progress of machine learning.