• No results found

Algorithm-Facilitated Tacit Collusion under Article 101 TFEU

N/A
N/A
Protected

Academic year: 2021

Share "Algorithm-Facilitated Tacit Collusion under Article 101 TFEU"

Copied!
44
0
0

Bezig met laden.... (Bekijk nu de volledige tekst)

Hele tekst

(1)

Algorithm-Facilitated Tacit Collusion under Article 101 TFEU

Anastasiia Kachalova anastasiia.kachalova@student.uva.nl 12778826 LLM International and European Law: European Competition Law and Regulation

Thesis Supervisor: dr. Jan Broulík 24.07.2020

(2)

2

Abstract

Nowadays algorithms are performing an important role in the daily operation of undertakings, taking Uber and Airbnb as a well-known example. The main focus of this thesis lies on the competition concerns arising from algorithm-facilitated tacit collusion. Tacit collusion or, from the legal perspective, parallel behaviour – the intelligent response of an undertaking to the changes on the market, does not trigger the application of art. 101 TFEU. The main question is to what extent does the legality of parallel behaviour preclude competition law to fully address the anti-competitive effects arising from the algorithmic operation, taking into account the existence of alternative tools to tackle such issue.

This thesis examines the concept of tacit collusion and market-specific conditions which are necessary for the rise of tacit collusion on the market. After such conditions are identified, algorithms and their specific characteristics such as machine learning are closely examined. In this regard, two types of algorithmic collusion established by Ezrachi and Stucke – Predictable Agent and Digital Eye are of central importance to this thesis. Furthermore, it is discussed how algorithms can facilitate tacit collusion by contributing to the market-specific conditions of coordination. It is argued that the relevance of those coordination conditions, rooted in human behaviour, do not properly represent the contemporary reality of digitalisation due to specific characteristics inherent to computer-based machines.

After the possible impact of algorithms on the market has been identified, this thesis will address available options of the current system to adequately address the emerging adverse effects. The main findings of this research show that the current competition law framework is not able to adequately and fully tackle the issue of algorithm-facilitated tacit collusion, mainly, due to the adopted formalistic approach which rests on the human collusive intent. As we shall see the abilities of the independent computer programmes based on AI cannot be equated to human. Ultimately, this research does not offer solutions to the emerged problem but instead identifies regulatory challenges for competition enforcers.

(3)

Table of Contents:

INTRODUCTION ...4

1. TACIT COLLUSION OR PARALLEL BEHAVIOUR ...8

1.1. THE CONCEPT OF TACIT COLLUSION ...8

1.2. CONDITIONS FOR TACIT COLLUSION ...9

2. ALGORITHMS AND COMPETITION ...12

2.1. ALGORITHMS AND THEIR CHARACTERISTICS ...12

2.1.1. ALGORITHMS ...12

2.1.2. MACHINE LEARNING CAPACITY ...13

2.1.3. TYPES OF ALGORITHMIC COLLUSION ...13

2.2. EFFECTS OF ALGORITHMS ON COMPETITION ...14

2.2.1. GENERAL POSITIVE EFFECTS ON COMPETITION ...14

2.2.2. NEGATIVE EFFECTS ON COMPETITION ...15

B. ALGORITHM AS A FACILITATOR OF TACIT COLLUSION ...16

2.3. ALGORITHMIC TACIT COLLUSION AND CONDITIONS FOR COORDINATION ...17

3. AVAILABLE COUNTER-MEASURES TO ADDRESS ALGORITHMIC TACIT COLLUSION ...22

3.1. REVISION OF A CONCEPT “AGREEMENT” AND/OR “CONCERTED PRACTICE” UNDER ART. 101TFEU ...23

3.1.1. THE EXTENSION OF THE SCOPE OF AGREEMENT REQUIREMENT ...23

3.1.2. THE CONCEPT OF CONCERTED PRACTICE AND TACIT COLLUSION ...26

3.2. ALGORITHMS AS A FACILITATING PRACTICE ...27

3.3. ALGORITHMIC PRICE SIGNALLING AS A CONCERTED PRACTICE ...29

3.4. AVAILABLE COUNTER-MEASURES: ADEQUATE TOOLS TO TACKLE ALGORITHMIC TACIT COLLUSION? ...30

3.5. REGULATION CHALLENGES AND REQUIRED CHANGES ...31

4. CONCLUSION ...36

(4)

4

Introduction

This thesis focuses on the influence of digitalization on competition law, in particular, the ability of algorithmic tacit collusion to give rise to anti-competitive behaviour. The enhancement of digitalization and growth of artificial intelligence (AI) play an important role in the competition law of today. In general, artificial intelligence is changing how businesses operate and consumers consume. The use of technology is capable of reducing search costs and entry barriers by increasing information transparency. Despite such benefits, technology allows undertakings to implement anti-competitive strategies more easily, increase concentration, strengthen their dominant position and prevent potential competitors from entering the market.1 Even though most consumers do not anticipate the role algorithms play in their trading and decision-making processes, when buying a specific product, algorithms have gained a prominent position in the current commercial model.2

Pricing algorithms have taken an important place in heated discussions among various economists, legal scholars and competition enforcement authorities. Even though consensus has been reached on the certain aspects of algorithmic pricing to a certain extent, the issue of algorithm-facilitated tacit collusion remains open.

The algorithmic collusion infringes competition law when the criteria of art. 101 TFEU are fulfilled. First and foremost, according to that provision, there has to be an agreement, decision or concerted practice among undertakings or associations of undertakings. Interestingly enough, this first condition already precludes the successful application of art. 101 TFEU to algorithmic tacit collusion. This happens when such conduct is neither based on agreement nor concerted practice, but instead is a result of an autonomous decision of algorithm which adapts its behaviour to the changes on the market. Furthermore, according to the current competition law regime, undertakings are free to intelligently adapt their conduct to other actors on the market.3 As a consequence, tacit collusion, which is also named as a parallel behaviour, is considered to be lawful under art. 101 TFEU.4

1 European Commission, ‘Competition Policy for the Digital Era. Final Report’ 4 April 2019, p. 11-13. 2 Alžbeta Krausová, ‘EU Competition Law and Artificial Intelligence: Reflections on Antitrust and Consumer

Protection Issues’(2019) International Journal for Legal Research 1, p. 79-80.

3 Case 48/69 Imperial Chemical Industries Ltd v Commission of the European Communities (‘Dyestuffs’) [1972]

EU:C:1972:70, para. 118.

(5)

However, according to Ariel Ezrachi and Maurice E. Stucke, algorithmic tacit collusion raises competition concerns and poses one of the main challenges to competition law in the area of digitalization.5 Nowadays there is an increasing tendency to move from collusion in smoke-filled rooms to collusion online with the use of pricing algorithms.6 This means that in today’s world the relevance of prerequisite requirement of anticompetitive intent or concurrence of wills between the competitors to collude is significantly mitigated.7 This may have inevitable consequences on the current competition law model, which is predominantly based on the formalistic approach, requiring the existence of an agreement or concerted practice.

Up until now, there is no or little empirical evidence proving that tacit collusion facilitated by algorithms has been liable for collusive outcomes and distortion of competition.8 Nonetheless, one cannot immediately conclude that algorithmic tacit collusion is a matter of legal sci-fi. The actual or foreseeable possibility of collusive outcomes or at least its facilitation should be enough to raise competition concerns.9 The seriousness of the challenge has been recognised by the EU Commissioner for Competition - Margrethe Vestager in 2017 during the 18th Conference on Competition in Berlin where she stated that:

“The challenges that automated systems create are very real. If they help companies to fix prices, they really could make our economy work less well for everyone else. So as competition

enforcers, I think we need to make it very clear that companies can’t escape responsibility for collusion by hiding behind a computer program. We do have to keep a close eye on how algorithms are developing. So when science fiction becomes a reality, we’re ready to deal with

it.” 10

Pursuant to the above-mentioned citation, the competition law and its enforcement authorities have to be ready to adequately tackle the problem before it hits. As one saying states: once forewarned

5 Ariel Ezrachi and Maurie E Stucke, ‘Emerging Antitrust Threats and Enforcement Actions in the Online World’

(2017) 13 Competition Law International 2.

6 Ariel Ezrachi and Maurice E Stucke, ‘Artificial Intelligence & Collusion: When Computers Inhibit Competition’

[2017] University of Illinois Law Review 5, p. 1782.

7 Ibid.

8 Ashwin Ittoo and Nicolas Petit, ‘Algorithmic Pricing Agents and Tacit Collusion: A Technological Perspective’

(2017) IO: Regulation.

9 OECD, Algorithmic Collusion: Problems and Counter-Measures, Roundtable on Algorithms and Collusion - Note by A. Ezrachi and M. E. Stucke, (21 June 2017), DAF/COMP/WD(2017)25, p. 61.

10 Margrethe Vestager, ‘Speech. Algorithms and Competition’ Bundeskartellamt 18th Conference on Competition,

Berlin (16 March 2017), <

https://wayback.archive-

it.org/12090/20191129221651/https://ec.europa.eu/commission/commissioners/2014-2019/vestager/announcements/bundeskartellamt-18th-conference-competition-berlin-16-march-2017_en> accessed on 24 July 2020.

(6)

6 is forearmed. Therefore, the evaluation of the progress that competition law has made in the field of algorithmic collusion, especially three years after the conference, is of vital importance to this research.

This thesis attempts to demonstrate a direct link between the legality of parallel behaviour under art. 101 TFEU and the (in)ability of competition law to adequately deal with algorithmic tacit collusion. As a consequence, the research question reads as follows: to what extent does the

legality of parallel behaviour under art. 101 TFEU preclude EU competition law to adequately address the possible anti-competitive effects on competition caused by algorithmic tacit collusion?

Subsequently, this question can be divided into the following sub-questions: What is meant by algorithmic tacit collusion? What are the effects of such algorithmic tacit collusion on the competition? What role does the legality of parallel behaviour under 101 TFEU play in this matter? What tools does the current EU competition law have at its disposal to tackle issues arising from algorithmic tacit collusion? Are those available means adequate in relation to the magnitude of the problem? If the available tools are not adequate, and the regulatory change is required, what are the challenges for regulation in this area?

The main purpose of this thesis is twofold. Firstly, it aims to further contribute to the debate on algorithmic tacit collusion and its capacity to produce anti-competitive effects. Secondly, it evaluates the ability of European competition law to adequately tackle possible adverse effects of algorithm-facilitated tacit collusion. In order to provide answers to the main question, this thesis relies on descriptive as well as normative methodology of legal research. The descriptive part depicts the current competition framework concerning tacit collusion and represents the existing opinions regarding the ability of algorithms to adversely affect competition. The normative part evaluates the capacity of the competition law to fully address algorithmic tacit collusion and expresses own viewpoint and arguments on the main issue.

The above-mentioned sub-questions form and guide the structure of the thesis which is as follows: Chapter 1 introduces the concept of tacit collusion and its legality under the current competition law framework. Furthermore, Stigler’s market-specific conditions for coordination are introduced.

Chapter 2 provides with the short introduction to the concept of algorithm, its specific characteristics and its relation to competition law. It also presents four types of algorithmic collusion, according to Ezrachi and Stucke, and identifies which types are relevant to this research.

(7)

Moreover, this chapter answers the question of how algorithms can facilitate tacit collusion and what are the effects of such algorithmic-facilitated collusion on competition law.

Chapter 3 refers to the legal framework of competition law and its available means to adequately tackle the anti-competitive effects raised by algorithmic tacit collusion. Essentially, it will address the last part of Margrethe Vestager’s speech on whether competition law is ready to deal with the adverse effects of algorithmic tacit collusion. If the answer is negative and the regulatory action is required, the main challenges of regulation in this field are going to be discussed.

Chapter 4 concludes the discussion and briefly summarises the main competition law issues arising from algorithmic tacit collusion and readiness of competition law to fully address those issues.

(8)

1. Tacit Collusion or Parallel Behaviour

1.1.The Concept of tacit collusion

The concept of tacit collusion presupposes collusion that is contrary to express. In the European competition law terms, tacit collusion refers to the individual alignment of interests between competitors which is neither based on agreement nor concerted practice within the meaning of art. 101 TFEU. Furthermore, tacit collusion is an economic concept. Economists do not distinguish between agreements, concerted practices and tacit collusion, on the contrary, their focus predominantly lies on the effects of the particular conduct on the market. This distinction, however, is of central importance to lawyers. From the legal point of view, only the existence of an agreement or concerted practice triggers the application of art. 101 TFEU. In the lawyers’ perception, the use of ‘collusion’ in the concept of ‘tacit collusion’ sounds too rigid as collusion assumes infringement of art. 101 TFEU, provided that the existence of an agreement or concerted practice is established.11 Hence, instead of ‘tacit collusion’ the concepts of ‘conscious parallelism’ and/or ‘parallel behaviour’ are more commonly used in the legal literature. For the purpose of this research, all above-mentioned notions are implemented.

Under the current competition law framework, tacit collusion or parallel behaviour is lawful as it does not fall under the notion of agreement or concerted practice which triggers the application of art. 101 TFEU. The concept of “agreement” was clarified in the Bayer judgement as it encompasses “the existence of concurrence of wills between at least two parties, the form in which it is manifesting being unimportant so long as it constitutes the faithful expression of the parties’ intention”.12 Whereas the notion of “concerted practice”, according to the judgement in Imperial

Chemical Industries, relates to “a form of coordination between undertakings which, without

having reached the stage where an agreement properly so-called has been conducted, knowingly substitutes practical cooperation between them for the risks of competition”.13 According to the

Court, undertakings are free to adapt prices and intelligently respond to present or foreseeable conduct of their competitors.14 Therefore, following these definitions, it is clear that a pure form

of tacit collusion falls under neither of concepts.

11 Whish and Bailey (n 4), p. 573.

12 Case T-41/96 Bayer v Commission [2000] ECR II-3383, para. 69. 13 Dyestuffs (n 3), para. 64.

(9)

Nevertheless, the Court recognised that: “although parallel behaviour may not in itself be identified with a concerted practice, it may however amount to strong evidence of such a practice if it leads to the conditions of competition which do not correspond to the normal conditions of the market”.15

From this passage, the consideration can be drawn that concerted practice can be inferred from tacit collusion when the latter is the reason for abnormal modifications of the market conditions, leading to coordination. As a consequence, the following paragraph focuses on the conditions which are necessary for coordination to occur.

1.2.Conditions for Tacit Collusion

After the clarification on what falls under the concept of tacit collusion or parallel behaviour has been identified, the conditions under which tacit collusion gives rise to coordination have to be closely considered. It is true that tacit collusion does not occur on every market.16 Nonetheless, specific market conditions such as concentrated markets involving homogenous products are more prone to give rise to tacit collusion.17

For coordination to take place three market-specific cumulative conditions, as was identified by George Stigler, have to be met.18 Firstly, competitors have to reach an understanding or agreement

on the terms of their coordination. Secondly, the parties to coordination can detect deviations from the terms and, thirdly, they are capable of exercising a credible threat of retaliation to prevent deviations. The same cumulative market-specific characteristics necessary for coordination were adopted by the European competition law. 19 According to most legal scholars, algorithmic tacit collusion can occur in the markets where coordination is likely and sustainable.20 This is the case when previously mentioned cumulative conditions for coordination are fulfilled.

The competitors are likely to opt for coordination and reach a common understanding when coordination is more beneficial for competitors than competitive conduct. This also happens when the economic environment is less complex, more stable with a smaller amount of players and

15 Ibid, para. 66. 16 OECD (n 9), p. 3.

17 European Commission, Guidelines on the assessment of horizontal mergers under the Council Regulation on the

control of concentration between undertakings, (2004/C31/03), (“Horizontal Mergers Guidelines”) para. 41.

18 George J Stigler, ‘Theory of Oligopoly’ [1964] J Political Economist 72, p. 44-46. 19 Horizontal Mergers Guidelines (n 17), paras. 39-57.

20 Ezrachi and Stucke (n 6); Michal Gal, ‘Algorithms and Illegal Agreements’ [2019] Berkeley Technology Law

(10)

10 involves homogenous products available on that market.21 The high level of market transparency

and availability of information provides firms with an ability to timely and sufficiently detect deviations.22 Thirdly, the consequences of deviations have to be drastic for the coordination to work.23 The use of a credible threat of retaliation keeps coordination intact.24 Last but not least, the fourth condition for successful coordination was added relating to high entry barriers for potential competitors and low countervailing buying power of consumers.25 Therefore, all market-specific characteristics are of vital importance to establish the likelihood of coordination, meaning that some markets are more prone to coordination than the others.26

According to the economic view, both explicit and tacit collusion can give rise to coordination, leading to prices above competitive levels. The main difference however is that price increase in case of explicit collusion is based on established agreement, whereas, in case of tacit collusion, it can be explained by the firm’s rational unilateral decision which takes into account competitors’ responses.27 In the determination of the likelihood of coordination arising from tacit collusion, the “amount of players” criterium plays a central role. It is assumed that sustainable tacit collusion may arise with two players in the market, rarely with three, and almost never in markets with four players.28 That is why tacit collusion is often called oligopolistic coordination.29 The reason behind this statement is that the more actors are involved in coordination, the harder it is for them to sustain it. In the oligopoly markets, where the amount of players is small, the deviation from coordinated behaviour is easily detectible, stimulating undertakings not to compete and to base their independent decisions on the foreseeable reactions of competitors. In the oligopoly theory, competitors are interdependent because they are aware of each other’s presence and share the common interest of maximising profits. In this case, the specific structure of the market affects the behaviour of the competitors, who tend to increase their prices, which subsequently raises competition concerns.30

21 Ibid, paras. 44-48. 22 Ibid, paras. 49-51.

23 Horizontal Mergers Guidelines (n 17), paras. 52-55. 24 Ibid, para. 52.

25 Ibid, paras. 56-57. 26 Gal (n 20), p. 74.

27 Guan Zheng and Hong Wu, ‘Collusive Algorithms as Mere Tools, Super-Tools or Legal Persons’ (2019) 15

Journal of Competition Law and Economics 2-3, p. 132.

28 Jan Potters and Sigrid Suetens, ‘Oligopoly Experiments in the Current Millennium’ (2013) 27 Journal of

Economic Surveys 3, p. 448. Niklas Horstmann, Jan Krämer and Daniel Schnurr, ‘Number Effects and Tacit

Collusion in Experimental Oligopolies’ [2016] 66 Journal of Industrial 3, p. 9. 29 Gal (n 20), p. 74.

(11)

Moreover, tacit collusion is not easy to achieve due to the lack of communications between competitors. Communication is an important factor in fulfilling the first condition for coordination, namely, to reach agreement on terms of coordination. As a consequence, the legal approach predominantly focuses on the means of collusion and human intentions. However, according to the economic model, the method of reaching coordination is not confined within a specific form and mode of communication.31 Even though communication is helpful in reaching coordination between competitors, it is not a prerequisite requirement for coordination, provided that competitors understood each other and coordinated their behaviour.32

When all the above-mentioned conditions are fulfilled, the coordination as a result of tacit collusion can be expected.33 There is no clear distinction between markets that are and not concentrated enough to give rise to either express or tacit collusion.34 The implementation of algorithms by competitors may affect market characteristics in such a way as to make markets more conducive to coordination. Therefore, the following chapter focuses on algorithms and their ability to facilitate tacit collusion on the market.

31 Louis Kaplow, ‘On the Meaning of Horizontal Agreements in Competition Law’ (2011) Harvard Law and

Economics Discussion Paper No 691, p. 98.

32 Ibid, p. 101. 33 OECD (n 9), p. 5.

34 Ariel Ezrachi and Maurice E Stucke, ‘Sustainable and Unchallenged Algorithmic Tacit Collusion’ (2020) 17

(12)

2. Algorithms and Competition

2.1.

Algorithms and their characteristics

2.1.1. Algorithms

In the Cambridge English Dictionary, the term ‘algorithm’ is defined as ‘a set of mathematical instructions or rules that, especially if given to a computer, will help to calculate an answer to a problem’.35 The concept of algorithm has been also defined in the EU Commission’s report as ‘any

set of rules about turning digital inputs into digital outputs. An algorithm is decision-making software’.36 Two significant elements could be derived from the above-mentioned definitions:

firstly, algorithms provide an answer or a decision; secondly, they are capable of reaching conclusions based on digital inputs.

Algorithms differ in their computer abilities and procedures. For example, expert algorithms are algorithms that work based on pre-determined features and set of rules implemented by a developer which limit their adaptability to changes. Contrary, learning algorithms are algorithms that use data inputs and decision-making processes to deliver a decision.37 The machine learning ability of algorithms, which is specific to the latter type, is of vital importance. Learning algorithms are not limited in the capacity to learn beyond the scope of already implemented instructions. They can adjust their responses according to newly emerged information.38 Human cognitive capacity cannot be compared to the algorithmic one in the context of the decision-making process. The collection, analysation and comparison of available sizable data are swift and mostly accurate with algorithms. They can also use any parameters and the unlimited number of variables to reach conclusions which are seen as an additional complex burden to human brains. The accuracy of the decision depends on the quality of available data and computer-specific characteristics.

35 The Cambridge English Dictionary, <https://dictionary.cambridge.org/dictionary/english/algorithm> accessed on

14 June 2020.

36 OECD, Algorithms and Collusion - Note from the European Union, 21-23 June 2017, DAF/COMP/WD(2017)12,

p. 2.

37 Gal (n 20), p. 9-10.

38 Madeleine de Cock Buning, ‘Artificial Intelligence and the Creative Industry: New Challenges for the EU

Paradigm for Art and Technology by Autonomous Creation’ in Woodrow Barfield and Ugo Pagallo (eds), Research

(13)

2.1.2. Machine learning capacity

Due to the fact that this thesis focuses on algorithmic tacit collusion, the machine learning capacity of AI is of particular importance. The example of self-learning algorithms can be easily understood with the reference to games. In general, human when playing a game against computer encounters an increasing level of difficulties to win with every new round because the computer can learn from every loss and adopt new strategies.39 Such enhanced machine learning capacity has been recently proven in various researches, for example, in Google research project DeepMind the computer has learned how to play different games: Chess, Shogi and Go based on trials and errors.40

Amongst the available literature on AI, generally, three main types of machine learning are identified: supervised, unsupervised and reinforced learning.41 Firstly, as regards to supervised learning, the algorithm uses the set of examples and goal variables to be able to learn how to reach the desired data output when confronted with new examples. When the achieved result is incorrect, the algorithm is modified to be able to reach the desired level of accuracy.42 Secondly, in

unsupervised learning, algorithms are categorising and organizing available data inputs in such a way that a new pattern is created. This type of machine learning is used to identify consumer preferences and to personalise available data accordingly.43 Thirdly, algorithms with reinforced

learning focus on reaching the best strategy through trials and errors which are neither based on previously implemented data inputs/outputs nor the modifications of the incorrect goals. The unsuccessful strategy is rarely achieved because it is punished, while successful strategy is rewarded.44

There are different ways in which algorithms may facilitate collusion. The existing types of algorithmic collusion are discussed in the following section.

2.1.3. Types of Algorithmic Collusion

39 Stephen Marsland, Machine Learning: An Algorithmic Perspective (2nd edn, CRC Press 2015), p. 4.

40 DeepMind research, ‘A General Reinforcement Learning Algorithm that Masters Chess, Shogi and Go Through

Self-Play’ (2018).

41 Ulrich Schwalbe, ‘Algorithms, Machine Learning and Collusion’ [2018], p. 7; Marsland (n 39), p. 6; OECD, Algorithms and Collusion: Competition Policy in the Digital Age (2017), p. 9.

42 Schwalbe (n 41), p. 6-9. 43 Ibid, p. 8.

(14)

14 Ezrachi and Stucke introduced four non-exclusive types of algorithmic collusion in their book “Virtual Competition” which are going to be briefly discussed.45

The first type is a Messenger – algorithm which is used by operators to assist in the implementation of an already existing collusive agreement. In the case of a Hub and Spoke algorithm, the hub (developer) creates world-wide collusion through vertical agreements concluded with numerous users who in turn use the hub’s algorithm to identify market prices. The third type - Predictable

Agent is an algorithm created by each market operator unilaterally to predict outcomes on the

market based on available or collected data. Last but not least, Digital Eye is an algorithm unilaterally implemented by the operator which achieves its strategic goals independently through the use of machine learning.46 This distinction plays an important role in the discussion on the level of required regulatory change when dealing with possible anti-competitive effects of algorithmic collusion.

These anti-competitive effects will be discussed in the following paragraph. It has to be borne in mind that the scope of this thesis is limited to algorithm-facilitated tacit collusion, meaning that some types of previously mentioned algorithmic collusion are going to be discussed in a greater extent than the others.

2.2.

Effects of algorithms on competition

2.2.1. General positive effects on competition

In the era of digitalization the use of algorithms becomes more and more customary, some businesses, if not most of them, have closely incorporated AI in their business model so that the absence thereof can prevent them from achieving essential efficiencies.

Algorithms optimise business processes through faster and more correct monitoring and analysation of a substantial amount of information in a shorter period of time.47 Algorithmic pricing allows both businesses and customers to enjoy extensive data information, accurate price adjustments and lead to the best market outcomes. Moreover, sophisticated algorithms with

45 Ariel Ezrachi and Maurice E Stucke, Virtual Competition: The Promise and Perils of the Algorithm-Driven Economy (Harvard University Press 2016).

46 Ibid, p. 36-37. 47 OECD (n 36), p. 11.

(15)

machine learning are capable of using data available on the market including the one from competitors’ and adapt prices according to modifications on the market.48 The real-life example

of such a business model is Airbnb – an online platform which offers housing for rent world-wide based on the listings provided by local hosts. Each host has a chance to set his/hers price per night for its property or use the smart pricing option offered by Airbnb. Such smart pricing mechanism is based on an algorithm with machine learning which adjusts prices according to several factors such as location, size, demand for property in that period of time, duration of booking, season, availability.

Furthermore, algorithms can promote competition due to a sufficient number of undertakings using technologically advanced business models and engaging in a more rigid competition online. Consequently, new undertakings with more advanced algorithms can easily enter the market. Undertakings and their competitors can timely adjust their prices while a variety of consumers’ choice has been extended with the ability to swiftly change their preferences.

2.2.2. Negative effects on competition

In general, there are two main competition concerns that algorithms call into question.49 The

first issue relates to Big Data and privacy concerns.50 This competition concern falls outside the

scope of this research. The second concern, which is of central importance to this thesis, is the ability of algorithms to facilitate anti-competitive behaviour on the market.51 Algorithms can either be used as facilitators of pre-existing agreement or they can affect market characteristics in such a way, making markets prone to coordination.

a. Algorithm as a facilitator of a pre-existing agreement

Due to the fact that the scope of this thesis is limited to algorithm-facilitated tacit collusion, algorithms as facilitators of the existing agreement fall outside its scope. In short, such algorithms bring little to no challenges to the current model of competition law as they, contrary to algorithmic tacit collusion, fulfil the first condition for the application of art. 101 TFEU. The existence of an

48 Oxera, ‘Algorithmic Competition. Prepared for European Commission’ (2018),

<https://ec.europa.eu/competition/information/digitisation_2018/contributions/oxera/oxera_algorithmic_competition .pdf> accessed on 24 July 2020.

49 Schwalbe (n 41), p. 3.

50 OECD, Big Data: Bringing Competition Policy to the Digital Era - Background Note by the Secretariat (2016),

29-30 November 2016, DAF/COM/(2016)14.

(16)

16 agreement or concerted practice can be proved due to the possibility to establish collusive intent.52

As has been stated by Richard Whish and David Bailey in ‘Competition Law’: “this is a simple case: the exchange of information to support a cartel is unlawful, and the use of price-tracking software to achieve the same end is no different”.53

In this regard, both scenarios’ of Messenger and Hub and Spoke are seen as mere tools which facilitate the pre-existing agreement. The Messenger is used as a tool to execute or foster a collusive agreement. In case of Hub and Spoke, as it has been ruled in the UK’s Posters case,54 the existence of the intent or the knowledge of possible collusion between algorithms implemented through a common hub can be established, triggering the application of art. 101 TFEU. For example, in the Eturas case,55 the degree of knowledge about concerted practice played a significant role in finding the infringement of art. 101 TFEU with the use of Hub and Spoke algorithm.

b. Algorithm as a facilitator of tacit collusion

In contrast to algorithm as a facilitator of pre-existing agreement, algorithms are also able to adversely affect competition in the absence of collusive agreement. The tacit collusion may arise when unilaterally implemented algorithms, based on the available information and modifications in the market behaviour, decide that collusion is the best available strategy. Moreover, algorithms may alter market conditions in such a way that markets, including those where collusion was less probable before the implementation of algorithms become more inclined to coordination.56

“Firstly, algorithms are fundamentally affecting market conditions, resulting in high price transparency and high-frequency trading that allows companies to react vast and aggressively.

Secondly, by providing companies with powerful automated mechanisms to monitor prices, implement common policies, send market signals or optimise joint profits with deep learning techniques, algorithms may enable firms to achieve the same outcomes of traditional hardcore

cartels through tacit collusion”.57

52 Schwalbe (n 41), p. 5.

53 Whish and Bailey (n 4), p. 536.

54 Online sale of posters and frames, Case 50223, decision of the CMA, 12 August 2016; OECD, Algorithms and Collusion – Note form the United Kingdom, 21-23 June 2017, DAF/COMP/WD(2017)19, p. 6.

55 Case C-74/14 Eturas and other v Lietuvos Respublikos konkurencijos taryba [2016] ECLI:EU:C:2016:42, para.

45.

56 OECD (n 36), p. 33. 57 OECD, (n 50), paras. 49-50.

(17)

Regarding the rise of tacit collision, both scenarios of Predictable Agent and Digital Eye are of central importance. In both cases, the establishment of the collusive intent of undertakings, which triggers the application of art. 101 TFEU, is complicated due to the ability of such algorithms to independent decision-making. Therefore, the focus of the following sections lies on the question to what extent can algorithms, in particular Predictable Agent and Digital Eye scenarios, affect market characteristics so that tacit collusion emerges?

2.3.

Algorithmic tacit collusion and conditions for coordination

In this section, the algorithmic ability to tacitly collude and to affect above-mentioned Stigler’s conditions for coordination are discussed. To illustrate the possible impact of algorithms on market conditions the following hypothetical example is provided.

Imagine that three undertakings A, B and C are active on the market of provision of online meetings and communication services similar to Zoom, Webex and Google Hangout. Each online platform has incorporated a reasonable usage fee which can be viewed online. Also, each service provider offers the possibility to businesses to buy yearly licenses to use and share such rights of usage of online communication services with people employed by the firms. Due to the newly emerging tendency to shift from the office-based working activity to online workplaces, the demand in such communication services as provided by undertakings A, B, C has enormously increased among businesses and consumers. Firms make use of such license possibility to a significant extent. All three undertakings are striving to expand the number of users to increase their profit. For this purpose, these undertakings are making use of different yet similar algorithms to collect available data and monitor market changes to improve their technologies and services according to changing consumer preferences. In addition, the goal of profit-maximization was set for all algorithms. After a certain period, the demand for online communication services went up to such extent that algorithm A raises the price for services based on available data such as demand, amount of users and number of sold licenses. Subsequently, both algorithms of undertakings B and C detect such behavioural change and simultaneously increase their prices based on analysation of data available to them and a specific set of inputs. Those specific set of inputs could entail, for example, that algorithm B is programmed in such a way as to be able to compare prices set by algorithms A and C and to evaluate other available parameters such as demand and coverage to rationally adapt prices accordingly to any modifications in those parameters. As a consequence of such comparison and evaluation exercise, which is set to be conducted every 5 seconds, coordinated pricing arises. In this case, the price-fixing is not based on the collusive agreement

(18)

18 between undertakings A, B and C but instead, it is a result of the individual rational decision of the implemented algorithms. There is a lack of collusive intent on the part of developers whose primary aim was to employ efficient data collection and pricing decisions through the use of algorithms. In short, the increase in pricing decisions of the algorithms is the implied result of tacit collusion.58

Following this example, the substantial effects on the market conditions can be identified so that Stigler’s conditions for coordination are fulfilled. Firstly, in the economic sense, three undertakings have reached an agreement on their coordination based on the transparency of available accurate and actual data regarding prices, consumer demand and consumer preferences. Machine learning algorithm can analyse and implement the best strategy. When it appears that coordination is more rational than competitive behaviour, according to the algorithmic decision, the collusion will follow.59 Even when not all relevant information is available to reach a perfect profit-maximizing price, algorithms may observe and indirectly infer such information from the already existing benchmarks and modifications on the market.60 Furthermore, algorithms are free from human biases and thus they are not conducive to deception which increases the sustainability of coordination.61 Secondly, due to the increase in market transparency, algorithms are swift in detecting behavioural deviations and quick in reacting to the alterations in pricing. Thirdly, as a consequence of fast algorithmic responses to changes, the deviating undertaking would not be able to confer any benefit for price-cutting. In the absence of algorithms, as the criticism of the oligopolistic coordination theory assumes, the delay in discovering deviations would provide a diverging undertaking within a reasonable period to make sufficient profit.62 This is however different when algorithms come into operation. Moreover, in general, algorithms diminish the incentive to deviate from coordination which mitigates the need for the credible and strong threat of retaliation.63

Some economic studies have shown that to satisfy previously mentioned conditions for tacit collusion, some sort of pre-communication is required.64 The undertakings have to at least agree on basic terms and conditions for their coordination and somehow threaten each other to prevent deviations from coordination.65 According to Ezrachi and Stucke, such requirement of

58 Ezrachi and Stucke (n 6), p. 1790. 59 Gal (n 20), p. 83.

60 Ibid, p. 83-84. 61 Ibid, p. 84.

62 Whish and Bailey (n 4), p. 574. 63 Gal (n 20), p. 88-89.

64 Horstmann, Kramer and Schnurr (n 28).

(19)

communication is contrary to the current approach of competition law which defines tacit collusion as a unilateral alignment of interests. The prerequisite requirement of communication would make it possible for competition authorities to establish anti-competitive intent which is necessary for the application of competition law. This implies that tacit collusion would also have a possibility to be considered unlawful, which is not the case under current legal perspective.66

In my opinion, if we accept the argument of economists that communication is one of the prerequisite requirements, algorithmic tacit collusion would nevertheless satisfy such requirement. Communication is not limited to the particular type or mode of communication, meaning that non-verbal signals should also be sufficient.67 “While algorithms generally do not sign agreements, wink to each other, or nod their consent, they communicate through decisional parameters coded into them. Other firms can rely on such communications in order to shape their actions”.68 This highlights the fundamental difference between human and algorithmic collusion where, in case of the latter, transparency of information in combination with certain decision mechanisms enables algorithms to read each other’s mind and predict each other’s conduct.69

The distinction between Predictable Agent and Digital Eye lies in the degree of sophistication and ability to learn. The extent and limit of the reinforced learning capacity of algorithms, which is directly dependent on developments in AI, remain unclear.70 In case of the Predictable Agent

scenario, where undertakings unilaterally implement algorithms which adjust prices based on the available data on the market, is now considered to “be a real challenge for competition law”.71

The lack of clear empirical evidence and real case studies is often used as an argument to demonstrate that algorithmic tacit collusion is a matter of legal sci-fi. Also, some writers argued that scholars find it simply interesting to write on topics that can easily catch the attention of the public such as robots.72

This statement was questioned by Ezrachi and Stucke during Roundtable on Algorithms and Collusion in 2017.73 They referred to the number of case studies in the field of gasoline retail as

66 Ezrachi and Stucke (n 34), p. 230-232, 238. 67 Kaplow (n 31), p. 98.

68 Gal (n 20), p. 109. 69 ibid, p. 85.

70 Ezrachi and Stucke (n 34), p. 250-252.

71 Václav Šmejkal, ‘Cartels by Robots – Current Antitrust Law in Search of an Answer’ (2017) 4 InterEULawEast

2, p. 10.

72 Thibault Schrepel, ‘Here’s why algorithms are NOT( really) a thing’ (Concurrentialiste, May 2017). 73 OECD (n 9), para. 60.

(20)

20 real examples of algorithmic tacit collusion.74 For example, Germany imposed the obligation on

oil companies to publish changes in prices on the government’s electronic system which transferred these data to consumers. This system aimed to increase price transparency and provide greater freedom of choice to the customers. In general, transparency is considered to be a positive characteristic of the market. Transparency enables undertakings to freely trade online and discourage them from price discrimination, while it allows consumers to freely choose among available products and compare prices available online. However, the increase of transparency in the concentrated markets of homogenous products increases the risk of tacit collusion. Also, such transparency leads to price increase which harms consumers.75

Now imagine that the government introduced a petrol price transparency programme which obliged petrol companies to publish their tomorrow’s petrol prices. Also, these undertakings have unilaterally implemented algorithms entrusted with the task to adjust prices according to alterations on the market. These algorithms were using available information directly collected from the price transparency programme to adapt prices immediately after the modification. As a result, price increases became more probable since algorithms mimic price modifications. This example is taken from the real case which happened in Perth, Australia.76

Consequently, based on the examples provided above, one cannot argue that algorithm-facilitated tacit collusion purely amounts to the matter of legal sci-fi. “It is neither based on futuristic prediction nor vague assumptions. It simply describes the current state of technology”.77

Nevertheless, due to the existing uncertainty regarding the actual capacity of algorithms to autonomous decision-making in the absence of human interference, the following two outcomes, according to Ezrachi and Stucke, can be anticipated.78 Firstly, if we accept the statement that algorithms are not able to establish parallel behaviour independently, then either the undertakings will be able to develop and train algorithms to do so or they would refrain from using algorithms at all. Most probably, bearing in mind the efficiencies algorithms bring about for undertakings,

74 Ibid, p. 8.

75 ibid, p. 7; Ralf Dewenter, Ulrich Heimeshoff and Hendrik Lüth, ‘The Impact of the Market Transparency Unit for

Fuels on Gasoline Prices in Germany’ Discussion Paper no 220 ( May 2016) Dusseldorf Institute for Competition Economics,

<https://www.dice.hhu.de/fileadmin/redaktion/Fakultaeten/Wirtschaftswissenschaftliche_Fakultaet/DICE/Discussio n_Paper/220_Dewenter_Heimeshoff_Lueth.pdf> accessed on 24 July 2020.

76 OECD (n 9), p. 9. 77 Ibid, p. 16.

(21)

humans will continue adopting and implementing interdependent algorithms.79 Secondly, contrary

to the first outcome, if algorithms can reach tacit collusion independently through machine learning, then the legality of parallel behaviour would prevent the application of competition law to deal with algorithmic tacit collusion.80 The competition law enforcers cannot completely disregard the possibility of intelligent machines to be commonly implemented in the commercial model of tomorrow.81

Therefore, based on above-mentioned considerations, the conclusion can be drawn that sophisticated algorithms with machine and deep learning capacities facilitate tacit collusion and foster the fulfilment of coordination conditions. Furthermore, algorithms set aside the need for Stigler’s conditions in the first place, particularly, regarding the necessity of explicit agreements, credible mechanisms of punishment and decrease in the incentive to deviate.82 The main reason behind this is that Stigler’s coordination conditions and oligopolistic coordination theory are heavily based on human, rather than algorithmic, behaviour. Therefore, these theories do not reflect current realities, especially, taking into account the characteristics and capacities of smart machines. This being said, the next chapter focuses on the available tools of competition law to adequately tackle the rise of algorithmic tacit collusion.

79 Ezrachi and Stucke (n 34), p. 252. 80 Ibid, p. 253-255.

81 Šmejkal (n 71), p. 9-10. 82 Gal (n 20), p. 89.

(22)

3. Available counter-measures to address algorithmic tacit collusion

Under the considerations in the previous chapters, one of the inevitable consequences of the use of algorithms by undertakings is that they are capable of fostering the fulfilment of coordination conditions mentioned in section 1.2. Moreover, algorithms are capable of facilitating tacit collusion on the markets where it would not otherwise be possible due to their speedy detection and reaction abilities and minimization of human biases. Due to the computer-specific characteristics inherent in pricing algorithms, they can effectuate tacit collusion in oligopolistic markets and make it probable in non-oligopolistic markets.83

As has been previously emphasised, the current competition law policy rests on the formalistic approach which requires the existence of an agreement or concerted practice to infringe art. 101 TFEU. As a consequence, the price-fixing, that would otherwise be condemned if it was classified as a concerted practice, remains unenforced when it falls under tacit collusion leading to enforcement gap. For this purpose, the distinction should be drawn between decisions which are purely unilateral and decisions that derive from some sort of agreement or concerted practice. For example, various practices such as signalling, information exchange and other manipulations through the use of technologies, can be potentially appraised by the competition authorities when the link with an agreement or concerted practice is established. However, when we refer to an exclusively independent decision of the undertaking to alter its behaviour to the changing market conditions and no trace of the concerted practice can be found, the alleged anti-competitive behaviour escapes the enforcement leading to so-called enforcement gap.84 This calls into question the competition law ability to effectively tackle the anti-competitive effects arising from tacit collusion, specifically the one facilitated by machine learning algorithms. The arisen enforcement gap can encourage undertakings to further develop and enhance its pricing algorithms and facilitate tacit collusion.85

The possible effects of algorithms on the market, including the significant increase in the probability of tacit collusion and, as we have seen in section 2.3, mitigation of the actuality of the Stigler’s coordination conditions, may justify the reconsideration of the current competition law

83 Catalina G Verdugo, ‘Horizontal Restraint Regulations in the EU and the US in the Era of Algorithmic Tacit

Collusion’ [2018] UCL Journal of Law and Jurisprudence 7, p. 130.

84 Ezrachi and Stucke (n 34), p. 255. 85 ibid, p. 256.

(23)

framework. “The fact that tacit collusion is rational conduct cannot, and should not, be a cause for excuse under the competition rules”.86

Therefore, in this chapter, the available options to enforce algorithmic tacit collusion under the current competition law framework will be thoroughly assessed with simultaneous discussion on whether such options have to be further revised to adequately address the issue at stake.

3.1.

Revision of a concept “agreement” and/or “concerted practice” under

art. 101 TFEU

Following the definitions in the Bayer87and ICT88 cases previously discussed in section 1.1, it is clear that, from a legal perspective, a pure form of tacit collusion falls under neither of concepts.

3.1.1. The extension of the scope of agreement requirement

One of the possibilities to capture (algorithmic) tacit collusion is to extend the scope of agreement requirement.89 The proponents of such view are Posner90 and Kaplow.91 In Posner’s view, the economic approach to collusion has to be taken where the collusive pricing itself triggers the application of competition law, regardless of whether the collusion is tacit or explicit.92 In case of the former, a tacit meeting of minds should suffice for the application of competition law.93

In this case, the central question is what is the limit of such an extension? It is clear that if the notion of the agreement is defined too broadly capturing tacit collusion, then the distinction between agreement, concerted practice and other types of parallel or unilateral conduct would be blurred, triggering the application of art. 101 TFEU. This would lead to overregulation of what is supposed to be a competitive market. The boundary should be established between the deterrence of infringement and “chilling of desirable economic activity”.94

86 Nicolas Petit, ‘The “Oligopoly Problem” in EU Competition Law’ in Ioannis Lianos and Damien Geradin (eds), Handbook on European Competition Law Substantive Aspects (Edward Elgar 2013), p. 19.

87 Bayer (n 12). 88 Dyestuffs (n 3). 89 Verdugo (n 83), p. 130.

90 Richard A Posner, Antitrust Law (2nd edn, University of Chicago Press 2001), p. 57. 91 Kaplow (n 31).

92 Posner (n 90), p. 69. 93 Ibid, p. 94-95.

(24)

24 In order to avoid such overregulation, the clear classification of conduct is desirable. For example, Kaplow draws the line between independent behaviour, interdependent behaviour and express agreement.95 Interestingly enough, according to Kaplow, parallelism is considered to be independent yet similar behaviour of two or more undertakings. Furthermore, such parallelism becomes conscious but remains independent when undertakings are aware of each other’s behaviour. However, oligopolistic price coordination renders parallel behaviour as a highly interdependent action.96 Independent behaviour does not trigger the application of art. 101 TFEU, while interdependent behaviour may and express agreement does.97 The clear distinction is blurred between what can be classified as an express agreement and pure interdependent behaviour. The intention of the parties and their “concurrence of wills” is predominantly used as a norm to set the boundary between these two concepts. However, this approach does not lead to the desired outcome when algorithms are completely autonomous in their decision-making. The collusive intent of undertakings cannot be derived from the independent conduct of algorithms in the absence of a link.98

It can be argued that, in case of Predictable Agent, the developer incorporates certain parameters and hand-designed specific features, enabling the developer to predict the outcome of algorithmic decision-making or at least prove his awareness of the probability of collusion. According to the

Anic case, the collusive intent at least requires the possibility of the undertaking to reasonably

foresee the other parties’ conduct and its readiness to accept risks arising from it.99 Consciously developers are aware of the probable existence of other algorithms and their likelihood to use each other’s data inputs. This awareness is based on the common knowledge between undertakings that the adoption of pricing algorithms would be ineffective when unilaterally adopted. As a result, there is a “mutual dependence” between rivals to use pricing algorithms which can be used as a proof to establish intent within the meaning of art. 101 TFEU.100To the contrary, Zheng and Wu argue that: “regardless of whether it is traditional machine learning or deep learning, there are elements that humans cannot accurately understand and predict – the so-called ‘black boxes’”.101 This means that the prediction of algorithmic decisions and full control over algorithmic operations

95 Ibid, p.16. 96 Ibid, p. 22. 97 Ibid, p.11-14.

98 Zheng and Wu (n 27), p. 153.

99 Case C-49/92 P Anic Partecipazioni [1999] ECLI:EU:C:1999:356, para. 203.

100 Paolo Siciliani, ‘Tackling Algorithmic-Facilitated Tacit Collusion in a Proportionate Way’ (2019) 10 Journal of

European Competition Law & Practice 1, p. 33-34.

(25)

are not always possible by undertakings, preventing competition law from establishing collusive human intent.102

Furthermore, the distinction between express agreement and pure interdependence can become clearer with the use of specific factors such as the intentions of the competitors, way of communication, type of shared information.103 The alternative approach for distinction is to focus on the particular conduct of the undertaking and whether it leads to successful oligopolistic coordination, resulting in prices above the competition level.104 This differentiation method, when applied to pricing algorithms and their effects on market conditions, would qualify an interdependent action as an agreement and thus trigger the application of art. 101 TFEU. However, this method is already used to differentiate between interdependent and independent conduct where unsuccessful oligopolistic collusion falls under the latter.105

Finally, Kaplow argues that a narrow interpretation of agreement is not required by competition law because modern oligopoly theory does not distinguish between express agreement and pure interdependence. This is the case, especially regarding price-fixing arrangements which, in economists’ perspective, have such a negative effect on social welfare which has to be appraised even when price-fixing is a result of pure interdependence.106

In my opinion, both considerations of Posner and Kaplow do not merely advocate for the broader interpretation of agreement requirement. They shift the focus from the formalistic approach, which concentrates on the existence of an agreement, to outcome approach where the anti-competitive effect of the conduct is of central importance.

Ultimately, the extension of the scope could lead to undesired outcomes of overregulation by including behaviour that has pro-competitive effects. Therefore, the impact of under-regulation has to be weighed against (over)regulation in order to adequately evaluate the degree of the required enforcement change.107 In my opinion, the regulation in this area is necessary due to the algorithmic ability to tacitly collude both in oligopolistic and non-oligopolistic markets, while the original economic theory of tacit collusion exclusively relies on oligopoly. This contradiction in

102 Zheng and Wu (n 27), p. 129-130. 103 Kaplow (n 31), p.15.

104 Ibid. 105 Ibid. 106 ibid, p. 114.

(26)

26 theory and in practice proves that algorithmic tacit collusion is a peculiar case, different from human tacit collusion, which can justify the refined approach to the current legal system.108

3.1.2. The concept of concerted practice and tacit collusion

Besides the extension of agreement requirement, the alternative solution would be widening of the scope of concerted practice to capture tacit collusion. However, under specific circumstances, tacit collusion can contribute to the finding of a concerted practice. The relation between concerted practice and parallel behaviour does not relate to a clear-cut case and it will be further elaborated in this section.

Even though the Court in Dyestuffs ruled that parallel behaviour itself does not fall under the notion of concerted practice, it can nevertheless be considered as strong evidence of such a practice when it leads to abnormal conditions of the market.109 The breakthrough case which paved the way in the direction of the possible illegality of tacit collusion under art. 101 TFEU is Wood Pulp.110 In this case, the Commission argued that parallel behaviour amounts to a concerted practice because price announcements lead to artificial market transparency, while the coordination arose in less tight oligopoly market where it would not be expected. The same approach has been later affirmed in the CISAC case111 where the Commission found that collusion was a result of parallel conduct

between competitors.112 Even though the Commission’s finding was rejected by the Court in both cases due to the lack of sufficient proof of concerted practice, the Court affirmed that tacit collusion could be used as evidence of a concerted practice when no other plausible alternative explanation exists.113

Therefore, the relationship between parallel behaviour and concerted practice is interrelated due to the fact that the existence of the latter could be inferred from the former, provided that parallel behaviour is the only plausible explanation of the collusive outcome. It is sufficient for an undertaking to provide proof of circumstances which show another explanation of

108 Zheng and Wu (n 27), p. 154. 109 Dyestuffs (n 3), paras. 65-66.

110Joined cases 89/85, 104/85, 114/85, 116/85, 117/85, 125/85, 126/85, 127/85, 128/85,

C-129/85 A. Ahlström Osakeyhtiö and others v Commission of the European Communities (‘Wood Pulp II’) [1994]

EU:C:1993:120.

111 Case T-442/08 CISAC v Commission [2013] EU:C:2013:188.

112 CISAC (Case COMP/C2/38.698) Commission Decision of 16 July 2008 [2008] C(2008) 3435 final, paras.

156-223.

(27)

concentration.114 The burden of proof rests on the Commission which has to establish that any

alternative explanation is implausible.115

3.2.

Algorithms as a facilitating practice

The preceding paragraph has shown that, in general, the Court is reluctant to accept parallel behaviour as evidence of concerted practice, imposing a high burden of proof on the Commission which precludes the effective enforcement of (algorithmic) tacit collusion. However, the concept of facilitating practice provides an alternative solution when dealing with behaviour contrary to express under art. 101 TFEU.116 Facilitating practice is used to tackle undertakings’ conduct

which exceeds the limits of the pure parallel behaviour and does not reach the threshold of the express agreement. The conduct which falls within the so-called “grey area”.117 Verdugo explicitly argues that facilitating practices could be used to address algorithmic tacit collusion.118 Moreover,

from the Court’s consideration in Wood Pulp follows that parallel behaviour can nevertheless be appraised under art. 101 TFEU when it amounts to practice that artificially facilitates collusion.119 In case of algorithms, their use and design enables them to facilitate and sustain coordination beyond what would normally happen on the market.120

According to OECD: “the concept of “facilitating practices” refers to the conduct by firms, typically in an oligopolistic market, which does not constitute an explicit, “hardcore” cartel agreement, and helps competitors to eliminate strategic uncertainty and coordinate their conduct more effectively”.121 The facilitating practice can also be unilateral.122 For example, the undertaking’s unilateral decision to implement pricing algorithms which tacitly collude can fall under the unilateral facilitating practice.123

114 CISAC (n 111), para. 99; Joined Cases T-305/94 to T-307/94, T-313/94 to T-316/94, T-318/94, T-325/94,

T-328/94, T-329/94 and T-335/94 Limburgse Vinyl Maatschappij and Others v Commission (‘PVC II’) [1999] ECR II-931, paragraph 727.

115 CISAC (n 111), paras. 140-145; Whish and Bailey (n 5), p. 580-581. 116 OECD (n 41), p. 20.

117 Ibid.

118 Verdugo (n 83), p. 136. 119 Wood Pulp (n 110), para. 126.

120 Stephan Thomas, ‘Harmful Signals: Cartel Prohibition and Oligopoly Theory in the Age of Machine Learning’

(2019) 15 Journal of Competition Law & Economics 2-3, p. 173-174.

121 OECD Policy Roundtables, ‘Facilitating Practices in Oligopolies’ DAF/COMP(2008)24, p. 9. 122 Verdugo (n 83), p. 136.

(28)

28 One of the clear examples of facilitating practice is an exchange of information which through the increase of market transparency leads to the rise of tacit collusion.124 The Commission’s assessment of the compatibility of information exchange with art. 101 TFEU is published in the

Guidelines on Horizontal Agreements (‘Guidelines’) where pro-competitive and anti-competitive

effects of information exchange are evaluated. In addition to market foreclosure, the significant increase of probability of coordination between competitors is considered to be one of the main competition concerns of information exchange.

The information exchange infringes art. 101 TFEU when it is used in support of an agreement or concerted practice125 which, in principle, means that pure tacit collusion as a result of information exchange is once again left unenforced.126 Nevertheless, in case of algorithmic tacit collusion, collusion arises as a result of how algorithms function which heavily depends on the information that algorithms receive.127 The application of information exchange as facilitating practice to algorithmic tacit collusion depends on the specific characteristics of algorithms, its ability to analyse available information, the type of consumed information and the degree to which that information is implemented in the algorithmic decision-making process.128

The likelihood of anti-competitive effects as a result of information exchange substantially depends on the market characteristics. According to the Guidelines, the information exchange is likely to have adverse effects on competition not only in tight oligopolies but also in fragmented markets provided that information exchange increases transparency or alters the market conditions in another way that is prone to coordination.129 Due to the fact that pricing algorithms, as was discussed in section 2.3, react swiftly to behavioural changes in the market and make it easier for undertakings to monitor and deter deviations, the coordination becomes probable even in non-oligopolistic markets.

Nonetheless, in practice, the Court is reluctant to accept the unlawfulness of information exchange under art. 101 TFEU, especially when such exchange is neither based on pre-existing agreement nor concerted practice.130 Moreover, each case of information exchange has to be individually assessed as its compatibility depends on the characteristics of the algorithm, its effect on the

124 Verdugo (n 83), p. 137.

125 European Commission, Guidelines on the applicability of Article 101 of the Treaty in the Functioning of the

European Union to horizontal co-operation agreements (2011/C 11/01), (‘Horizontal Guidelines’), paras. 55-59.

126 Petit (n 86), p.323. 127 Verdugo (n 83), p. 138. 128 Ibid.

129 Horizontal Guidelines (n 125), para. 79.

(29)

conditions of the relevant market, the type and quality of exchanged information.131 Such

individual evaluation of risks and benefits of information exchange is problematic in case of algorithms due to the existing uncertainty regarding the characteristics of algorithms, the manner of its operation and inadequate ability to detect collusion produced by algorithms.

3.3.

Algorithmic price signalling as a concerted practice

Similar to considerations in the previous paragraph, price signalling can be classified as a facilitating practice resulting from algorithmic tacit collusion. Thomas argues that: “signalling or any other type of information exchange is an emanation of tacit collusion in itself”.132

The debate exists as regards to the lawfulness of price signalling under art. 101 TFEU.133 Price

signalling encompasses situations when undertakings, regardless of the absence of a price-fixing agreement or concerted practice, unilaterally send signals about their commercial intentions to third parties leading to coordination.134 It is not implausible that pricing algorithms are devices that signal prices which significantly increase the likelihood of parallel behaviour. Undertakings that choose to signal their prices from time to time can be held liable for participation in a concerted practice, based on its definition in the Dyestuffs case, as such conduct knowingly substitutes practical cooperation between competitors for the risks of competition.135 According to the Guidelines, the price signalling can potentially be classified as a concerted practice under specific circumstances, for example, when public price announcement was followed by the competitor’s announcement which could constitute a ground for finding a mutual understanding in reaching coordination.136

However, due to the lack of case-law on price signalling, its application under art. 101 TFEU remains unclear. For instance, the Commission found in the Wood Pulp case, that quarterly price announcement could itself constitute a concerted practice or at least an evidence of a concerted practice.137 In this case, pursuant to the Commission, the price announcements lead not only to the

131 Case C-238/05 Asnef-Equifax v Asociacion De Usuarios De Servcicios Bancarios [2006] EU:C:2006:734, para.

54.

132 Thomas (n 120), p. 159-160. 133 Whish and Bailey (n 4), p. 582. 134 Whish and Bailey (n 4), p. 535. 135 Dyestuffs (n 3), para. 64.

136 Horizontal Guidelines (n 125), para. 63.

137 Case C-89/85 A Ahlström Osakeyhtiö and others v Commission of the European Communities (‘Wood Pulp’)

Referenties

GERELATEERDE DOCUMENTEN

Omringd door professionals, die beroepsmatig veel te  maken hebben met borstvoeding, heeft Julia er alle vertrouwen in dat zij kan rekenen  op een goede begeleiding. .. Onze

This is confirmed by Compte and Jehiel (2001), for a more general case of coordination in groups. There is an additional influence of the degree of concentration on the

Het toevoegen van paralinguïstische symbolen aan webcareberichten heeft volgens het huidige onderzoek geen effect op de attitude ten opzichte van de organisatie (H1), terwijl uit

We prove that any n-vertex triangle-free graph has at most 3 n/3 ≈ 1.4423 n maximal induced matchings, and this bound is attained by any disjoint union of copies of the

Zowel de Wet financieel toezicht als de Wet marktordening gezondheidszorg beoogt de positie van de consument te versterken, maar de toezichthouders op deze wetten gaan hier in

Archive for Contemporary Affairs University of the Free State

to-day who speak Afrikaans and think that they are speaking High Dutch), these. men saw that the only possible development for them was to gain recognition for

The relative contributions of each of the individual processes to coastal flooding are also assessed (i.e. surge, tide, wave setup and relative sea level rise), as are the