• No results found

Enforcement challenges stemming from tacit collusion and pricing algorithms in the new digital era

N/A
N/A
Protected

Academic year: 2021

Share "Enforcement challenges stemming from tacit collusion and pricing algorithms in the new digital era"

Copied!
34
0
0

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

Hele tekst

(1)

Enforcement challenges stemming

from tacit collusion and pricing

algorithms in the new digital era

Benjamin JAN Benjamin.jan@outlook.fr LL.M International and European law: EU Competition law and Regulation track Supervisor: Dr. Rein Wesseling Submission: 27th of July 2018

(2)

It has become common for firms to set prices for their products or services automatically using algorithms in digital markets. The increased transparency that comes with price monitoring and the speed with which pricing algorithms can adapt prices in these markets cause competition authorities to face new challenges. Today, business decision-making processes are supported by machines and in a near future, we can presume that machines will be able to take complex decisions themselves. At the same time, the trend in EU markets is towards concentration – oligopolies are widespread. These facts raise the question whether the use of pricing algorithms increases the likelihood of tacit collusion. If so, is the legal framework suitable to assess algorithms? This question will be answered through an analysis of Article 101 and 102 TFEU and merger control, but also with regards to other legal

(3)

I. Introduction... 4

II. Part 1: Impacts of pricing algorithms on tacit collusion...4

1. The age of pricing algorithms... 4

2. The concept of ‘tacit collusion’ and the ‘oligopoly problem’...5

3. Tacit collusion under EU competition law...6

4. Pricing algorithms, tacit collusion and its sustainability...7

5. Pricing algorithms and their impact on competition...8

a) (Potential) Anti-competitive effects...8

b) Pro-competitive effects... 10

III. Part 2: Enforcement challenges with regard to pricing algorithms...11

1. Legal approach under EU competition law...11

a) Liability... 12

i. When?... 12

ii. Who?... 16

b) Evidence... 19

c) Finding and termination of an infringement: Article 7 Regulation 1/2003...20

2. Alternative approaches to tacit algorithmic collusion...21

a) Ability to regulate and detection... 21

i. Auditing the algorithm... 21

ii. Shifting the burden to the companies...22

b) Market or sector investigations... 23

c) Unfair Commercial Practices Directive...23

d) Regulation... 24

IV. Conclusion... 26

(4)

I. Introduction

It has become common for firms to set prices for their products or services

automatically using algorithms in digital markets. The increased transparency that comes with price monitoring and the speed with which pricing algorithms can adapt prices in digital markets cause competition authorities to face new challenges. These challenges raise two issues: collusion between firms and price discrimination on consumers.

What interests us here is collusion, especially tacit collusion in a horizontal context. The latter is a complex issue under EU competition law since conscious parallelism itself is lawful. The use of algorithms makes the issue even more complex since it seems to put several business behaviours in a grey area which goes beyond conscious parallelism, but at the same time does not involve an express agreement between competitors.

Nowadays, business decision-making processes are supported by machines and in a near future, machines will be able to take complex decisions themselves. At the same time, the trend in EU markets is towards concentration – oligopolies are widespread. These facts raise the question whether the use of pricing algorithms increases the likelihood of tacit collusion. Are algorithms determinative for tacit collusion? If they are, how will competition enforcers and policy-makers face the challenges raised by them? Do they have enough relevant legal instruments in their toolbox to fight back against the potential risks raised by pricing algorithms? All these questions are related to a more central one: what are the (potential) enforcement challenges stemming from tacit collusion and pricing algorithms?

In the normative evaluation of enforcement challenges, we will proceed from official reports and literature available on this topic. Unfortunately, neither empirical evidence nor case law exist about it. Even so, through the existing ECJ’s case law on Article 101, Article 102 TFEU and Merger control, we will try to draw the potential answers which could be given regarding tacit collusion in virtual markets, and highlight the gaps in the competition law framework. Finally, outside the framework itself, we will try to find other answers to the challenges raised by algorithms through regulations and the Unfair Commercial Practices Directive.

Accordingly, this thesis focuses on two approaches: a legal approach with the core provisions of EU competition law and alternative approaches to the legal one. First, it explores the impacts of pricing algorithms on tacit collusion and their effect on competition. Second, it explores the potential answers offered by the current competition framework and other ones outside of the competition law framework.

II. Part 1: Impacts of pricing algorithms on tacit collusion

1. The age of pricing algorithms

Although there is no universal definition of the concept of ‘algorithm’, we can summarise – for the purpose of this paper only – that an algorithm has been ‘developed to automatically perform repetitive tasks involving complex calculations and data processing’1.

Algorithms are used in many fields, from detecting melanoma cancer at an early stage to matching customers willing to find love. However, what we should focus on here is pricing algorithms in digital markets. They are commonly understood as the ‘computational codes run by sellers to automatically set prices to maximise profits’2.

1 OECD, Algorithms and Collusion - Background Note by the Secretariat, 21-23 June 2017, p. 5. 2 Ibid., p. 14.

(5)

The growing use of pricing algorithms changed the digital economy and enabled new business models based on the collection and processing of data. Some industries are using them more commonly, like in the airline, hotel booking or retail industries, just to name a few. But the expansion of pricing algorithms is rising in other industry sectors, too3.

In figures, the Commission found that as many as two-thirds of EU retailers used price monitoring softwares, with some of them automatically configuring their own prices by referencing those of their rivals4. The latter – i.e. when algorithms are used to implement

continuous price changes over time – is called dynamic pricing5. Industries use it to challenge

their competitors’ prices in real time.

2. The concept of ‘tacit collusion’ and the ‘oligopoly problem’

Tacit collusion – also referred to by economists as ‘conscious parallelism’ – is commonly seen as a form of anti-competitive coordination which can be achieved without any need of an explicit agreement, but which competitors are able to maintain by recognising their mutual interdependence6. The non-competitive outcome is achieved by each participant

deciding its own profit-maximizing strategy independently from its competitors7.

This form of anti-competitive coordination typically occurs in transparent markets with few market players. The latter will be able to earn supra-competitive profits without entering into an agreement or concerted practice prohibited by competition law8.

The main objective behind this form of coordination is to raise profits to a higher level than the non-cooperative equilibrium, resulting in a deadweight loss9. In order to do so,

three conditions are needed:

- The first one is that the coordinating firms must be able to monitor to a sufficient degree whether the terms of coordination are being adhered to10.

- The second is sustainability; discipline requires that there is some form of credible deterrent mechanism that can be activated if deviation is detected11.

- The last one is the absence of competitive constraints. The reactions of outsiders, such as current and future competitors not participating in the coordination, as well as customers, should not be able to jeopardise the results expected from the coordination12.

Do the same conditions apply to tacit collusion through algorithmic pricing? The answer is affirmative, but this kind of tacit collusion will not affect all markets13. Algorithmic

collusion happens ‘in concentrated markets involving homogenous products where the 3 SK Mehra, ‘Antitrust and the Robo-Seller: Competition in the Time of Algorithms’ (2016) 100 Minn. L. Rev., p. 1338.

4 OECD, Algorithms and Collusion - Background Note by the Secretariat, 21-23 June 2017, note by the EU, p. 5.

5 OECD, supra n.1, p. 14. 6 Ibid., p. 17.

7 Idem.

8 R. Whish and D. Bailey, Competition Law (8 ed.,Oxford University Press, 2015), p. 596. 9 OECD, supra n.1, p. 17.

10 Guidelines on the assessment of horizontal mergers under the Council Regulation on the control of concentrations between undertakings, OJEU C 31, 05.02.2004, p. 5, from para. 39, on "coordinated effects", para. 41.

11 Idem. 12 Idem.

13 A. Ezrachi and M.E. Stucke, Algorithmic Collusion: Problems and Counter-Measures – Note, OECD, Roundtable on Algorithms and Collusion 21-23 June 2017, p.3.

(6)

algorithms can monitor pricing and other keys terms of sale to a sufficient degree’14. Indeed,

even though the degree of market concentration can be lower than in traditional markets, algorithmic collusion is unlikely to be a potential outcome when the products are differentiated since co-ordinating prices and other keys terms of sale is very difficult15.

Based on economic models16, tacit collusion is often referred to as the ‘oligopoly

problem’. The problem is that high interdependence and mutual self-awareness in oligopolistic markets might result in tacit collusion17. In highly concentrated and transparent

markets, firms can more easily become conscious that their respective strategic choices are interdependent and that they can set prices at a supra-competitive level without any communication18. This risk does not appear as negligible within the single market, since the

general trend in recent years has been towards an increase in industrial concentration19. It

should be noted that Ezrachi and Stuck postulated that the industry-wide use of algorithms, given the speed and enhanced transparency, could expand the range of industries susceptible to collusion beyond duopolies to perhaps markets dominated by 5 or 6 players20.

The main issue is that the oligopoly problem makes extremely difficult for antitrust authorities to distinguish between collusive and non-collusive outcomes only on the basis of what looks like parallel conduct in prices21. The question whether pricing algorithms

exacerbate the oligopoly problem will be examined below22.

3. Tacit collusion under EU competition law

Economists distinguish between two forms of collusion: explicit collusion and tacit collusion. One the one hand, under EU competition law, ‘explicit collusion’ refers to an ‘agreement’ or a ‘concerted practice’. An agreement has to be proved by the concurrence regarding the will23 of competitors. In the absence of a formal agreement, the category of

concerted practices can be applied. According to the ECJ, the concept of a concerted practice refers to ‘a form of coordination between undertakings which, without having been taken to the stage where an agreement properly so-called has been concluded, knowingly substitutes for the risks of competition practical cooperation between them’24.

On the other hand, tacit collusion occurs when oligopolists coordinate their prices without any explicit communication or contact, which makes it possible for them to behave in a parallel manner and achieve supra-competitive profits. Tacit collusion receives different treatment under the different legal instruments offered by EU competition law.

14 Idem.

15 A. Ezrachi and M.E. Stucke, supra n.13, p.3.

16 N. Jenkins and J. Kavanagh, Economics for Competition Lawyers (Oxford University Press, 2011), pp. 142-150.

17 OECD, supra n.1, p. 34. 18 Idem.

19 OECD, supra n.1, p. 595.

20 A. Ezrachi and M.E. Stucke, supra n.13, p.3.

21 P. Siciliani, Should We Act ex post Against Tacit Collusion—and How?, Journal of European Competition Law and Practice, May 2014, Vol. 5 Issue 5, pp. 294-303.

22 SK Mehra, supra n.3, p. 1346.

23 Case T-41/96 Bayer v Commission, ECLI:EU:T:2000:242, para. 69.

(7)

 European Merger Regulation: the creation or strengthening of a collective dominant position through tacit collusion that significantly impedes effective competition is prohibited25.

 Article 101 TFEU: the ECJ has been clear regarding the fact that ‘Article 101 does not deprive economic operators of the right to adapt themselves intelligently to the existing and anticipated conduct of their competitors’26. The ECJ added that ‘every

producer is free to change his prices, taking into account in so doing the present or foreseeable conduct of his competitors’27. Therefore, tacit collusion itself is legal and

will not be caught by Article 101 TFEU. However, if the parallel behaviour leads to conditions of competition that do not correspond to the market characteristics, the behaviour may amount to strong evidence of concerted practice under Article 101 TFEU28. In brief, all parallel behaviours are not automatically considered as concerted

practices.

 Article 102 TFEU: if a collective dominant position results from tacit collusion, the provision may be triggered if the undertakings abuse of their position.

4. Pricing algorithms, tacit collusion and its sustainability

Before thinking about any appropriate response from a competition enforcement perspective on pricing algorithms, we need to assess whether or not algorithms make collusive outcomes easier to sustain. If not, reflecting on enforcement challenges would prove unnecessary.

Relevant factors have been identified as increasing the likelihood of collusion in a given market and have been divided in three categories: structural characteristics, demand-side characteristics and supply-demand-side characteristics29. For the purpose of this paper, only the

structural characteristics will be examined because they are the most relevant regarding the impact of pricing algorithms on tacit collusion in digital markets. There are four structural characteristics: market transparency, frequency of interaction, barriers to entry and number of firms.

Let us start with two structural elements that are the most likely to be impacted by algorithms. Firstly, pricing algorithms increase market transparency. Ezrachi and Stucke warn us about the fact that as soon as a few sellers invest in a pricing algorithm for their business, rivals are going to develop their own to avoid being put at a competitive disadvantage30.

Therefore, if all or most sellers have their own algorithms, they will be able to observe real-time price fluctuation with the risk of it creating a transparent environment prone to collusion31.

25 Guidelines on the assessment of horizontal mergers under the Council Regulation on the control of concentrations between undertakings, OJEU C 31, 05.02.2004, p. 5, from para. 39, on "coordinated effects", para. 39-56.

26 A. Ezrachi and M.E. Stucke, supra n.13, p. 6.

27 Case 48/69 Imperial Chemical Industries (ICI) v Commission (Dyestuffs), ECLI:EU:C:1972:70, para. 118. 28 Ibid., para. 66.

29 OECD, supra n.1, p. 18.

30 A. Ezrachi and M.E. Stucke, Virtual Competition: The Promise and Perils of the Algorithm-Driven Economy (Harvard University Press, Cambridge, 2016), p. 21.

(8)

Secondly, the frequency of interaction in online markets is amazingly higher than in traditional markets thanks to pricing algorithms making price adjustments easy and costless. From a collusion perspective, those algorithms are useful because they shorten the time period between defection and detection and allow an immediate retaliation to deviations32.

Algorithms may also impact two other structural elements: the number of firms that may collude and the barriers to entry. By contrast with the above paragraphs, it is still unclear how algorithms may affect them exactly. In principle, tacit collusion may happen if there is a small number of firms and where there are enough barriers to entry to avoid new entrants coming and decrease the incentives to deviate from collusive equilibrium.

On the one hand, the impact of algorithms on the likelihood of entry is not univocal33. It is hard to assess its real impact since it will depend on the specific features of

each market: natural barriers to enter, number and size of competitive firms or network effects, just to name a few. On the other hand, pricing algorithms ‘make the number of competitors in the market a less relevant factor for collusion’34. It is likely to happen since

pricing algorithms can manage to monitor price deviations in a better – if not perfect – way than humans do with a higher number of prices.

All things considered, ‘if markets are sufficiently transparent and firm can adjust their decisions very fast, for instance by changing prices in real time, collusion is always sustainable, regardless of the potential counter-balancing effect of other factors, such as the number of firms in the industry’35. If we look at economic literature, Salcedo goes a step

further by saying that tacit collusion between firms employing pricing algorithms ‘is not only possible but rather, it is inevitable’36. However, his statement is based on strong assumptions

and still needs empirical evidence. The European Commission noted that pricing algorithms do not remove the need for some of the basic conditions for tacit collusion37. Others criticise

the fact that ‘first almost all studies explicitly focus only two players (the Duopoly situation). Second, the type of games (e.g., repeated Prisoners’ Dilemma and its variants) and the universe of possible strategies are limited, especially when compared to the real business world. Third, most studies do not consider any external uncertainty’38.

Even though scientific research on this topic still needs to be done, the risk created by the current market changes from pricing algorithms should not be underestimated. There are similarities with the classic ‘oligopoly problem’ but tacit collusion could become sustainable in a wider range of circumstances and expanding the oligopoly problem to non-oligopolistic market structures. When algorithms change certain structural characteristics of the market and increase the likelihood of collusion, it enables a new form of collusion: algorithmic collusion. The second part of this paper will therefore examine enforcement challenges against it.

32 SK Mehra, supra n.3, p. 1349. 33 OECD, supra n.1, p. 19. 34 Idem.

35 OECD, supra n.1, p. 22.

36 B. Salcedo, Pricing Algorithms and Tacit Collusion (Working paper, Cornell University, 2016).

37‘First, tacit collusion requires sufficiently homogeneous products. When sold online, even identical products can become differentiated in terms of delivery costs, delivery time, and the seller's reputation, for example. Second, tacit collusion requires effective retaliation, which in turn requires spare capacity. A capacity-constrained firm cannot initiate a price war as a means of retaliation to enforce tacit collusion. Third, before engaging in tacit collusion, a firm would need to decide that it is a better course of action than competitive pricing, especially if competitive pricing leads to drastically larger sales volumes’. OCDE, supra n.4, p. 8. 38 D. Ai, What Do We Know About Algorithmic Tacit Collusion? (May 2, 2018). Available at SSRN: https://ssrn.com/abstract=3171315, pp. 3-4.

(9)

5. Pricing algorithms and their impact on competition

a) (Potential) Anti-competitive effects

Today, there is still no empirical evidence of the effect that algorithms have on the actual level of prices and on the degree of competition in digital markets. However, the literature highlights the potential anti-competitive effect from a widespread use of algorithms as that it becomes easier from undertakings to achieve and sustain tacit collusion39. As more

firms use algorithms and more sophisticated algorithms are created, we could expect that challenges may soon increase for competition authorities.

In order to fully understand the potential effects of pricing algorithms, let us divide them in two categories that are the most relevant for tacit collusion and according to their roles as facilitators.

- Monitoring algorithms – Predictable Agent scenario

Monitoring algorithms check and adjust prices automatically. They are able to collect price fluctuation and monitor competitors’ actions. Combined with pricing algorithms, firms can automatically retaliate for any deviation from an agreed price between cartel members40.

They make explicit collusion more efficient and facilitate it. In this context, human behaviour is still needed behind the monitoring algorithms and therefore Article 101 TFEU is able to condemn it.

However, how can these algorithms affect tacit collusion? Can legal instruments within the competition framework catch all the possible unwanted effects of monitoring algorithms when they are not used to facilitate price-fixing cartels?

Ezrachi and Stucke proposed a scenario in which each firm, without communication between them, would monitor their respective prices through their own algorithms with a strategy to maximise profits. Their algorithms are programmed to monitor price changes and swiftly react to any price reduction undertaken by a competitor. At the same time, they are programmed to follow price increases when sustainable41.

Let us imagine a situation where each competitor in a digital market adopts this type of algorithm, and assess its impact on two relevant factors we have identified as increasing the likelihood of tacit collusion.

Firstly, the wide use of algorithms will increase the supply of market data (including competitors’), which will by itself increase market transparency. In traditional markets, transparency is limited because humans cannot be as efficient as algorithms42. The widespread

use of pricing algorithms will have the result of publishing online all the competitors’ current prices43. Therefore, each algorithm will rapidly inform competitors of any price fluctuation.

Secondly, the frequency of interaction is significantly higher in digital markets than in traditional ones. Each competitor will be able to shift its prices within a second44. Humans

are not in charge of monitoring price adjustments anymore, such task now falls to pricing algorithms, which ‘can assess and adjust prices – even for particular individuals at particular 39 Idem.

40 OECD, supra n.1, p. 25.

41 A. Ezrachi and M.E. Stucke, supra n.30, p. 61. 42 Idem.

43 Idem. 44 Idem.

(10)

times and for thousands of products – within milliseconds’45. There are two issues here: the

first is that a firm can instantly reduce its price in response to a rival’s discount, thus eliminating any incentive to discount in the first place; the second is that firms will follow price increases when sustainable.

Ezrachi and Stucke added that ‘in an environment dominated by similar pricing algorithms that are aware of opportunities to foster interdependence, the risk is higher prices’46. This risk is however hard to assess. Some economic studies have shown that if

market information arrives continuously and firms can react to it quickly, it doesn’t lead all the time to a price war in higher prices because ‘firms tend to act on noisy market information too fast’47 and make collusion hard to sustain48.

Are we in a science fiction scenario? It does not seem so, it is perhaps already a present issue or at least a near-future problem. In fact, two-third of retailers in the EU use software tools to monitor prices from their competitors49. The main risk of this scenario is that

it increases the risk of tacit collusion. On the one hand, markets where these algorithms are used are likely to be more transparent since an optimal use of pricing algorithms needs supply market data that all firms will be willing to share for their own use of algorithms. On the other hand, the high speed with which firms can monitor prices is likely to eliminate any incentive to discount in the first place. Moreover, if one firm departs from the appropriate behavioural standards, the other firms will be able to respond swiftly with a price war that will warn every competitor that relinquishing tacit collusion is not a good idea50.

- Self-learning algorithms – the Digital Eye scenario

Another kind of algorithm that is potentially able to sustain tacit collusion is the self-learning algorithm. It can anticipate and react to competitive threats well before any pricing change51. These algorithms achieve that through the use of machine learning and deep

learning technologies implemented to automatically set prices52. In a nutshell, these

algorithms ‘process raw data in a complex, fast and accurate way, resembling the human brain, and deliver an optimal output without revealing the relevant features that were behind the decision process’53.

Ezrachi and Stucke assume, under their ‘digital eye’ scenario, that each firm will have a technology that enables each algorithm to quickly detect any competitive manoeuvre, and thus enable it to know when and how to retaliate. This differs from the ‘predictable agent’ scenario in two ways. Firstly, tacit collusion is spread to all markets, not only the digital ones. Secondly, humans are ‘further detached from the algorithms’ tactical and strategic decisions. They don’t know whether, when, or for how long the algorithms have been tacitly

45 Ibid, p. 62. 46 Idem.

47 D. Ai, supra n.38, p.9.

48 Y. Sannikov and A. Skrzypacz, Impossibility of Collusion under Imperfect Monitoring with Flexible Production, American Economic Review 97, no. 5 (2007).

49 Commission Staff Working Document accompanying the Final Report on the E-commerce Sector Inquiry, document SWD (2017) 154 final of 10.5.2017.

50 R. Whish and D. Bailey, supra n.8, p. 598. 51 A. Ezrachi and M.E. Stucke, supra n.29, p. 61. 52 OECD, supra n.1, p. 30

(11)

colluding’54. In other words, the digital eye would tacitly collude in every market without any

human beings behind the decisions.

Self-learning algorithms entail other enforcement challenges than the previous category. The first challenge is liability. Since firms can reach a collusive outcome without being aware of it, who should be held responsible? Moreover, it raises an issue related to evidence. How could competition authorities rely on anticompetitive intent?

It should be noted that the digital eye scenario still pertains to science fiction. Nobody has proved that self-learning algorithms may reach a collusive outcome yet. When they tried, algorithms were defined by their specific environments, it is therefore currently impossible to say whether self-learning algorithms are able to collude or not55.

b) Pro-competitive effects

What if, after everything we discussed about the potential anti-competitive effects of pricing algorithms might have on competition, the anti-competitive effects would be outweighed by the pro-competitive ones? In fact, authors like Mehra postulate that ‘the significant labour cost savings and better competitive intelligence that robo-selling promises may partially or completely offset the potential for competitive harm’56. In more technical

terms, he states that ‘the possibility that producers who adopt robo-selling may see their marginal cost drop may outweight the incremental risk of deadweight loss due to increased supracompetitive pricing’57.

Pricing algorithms have tree main advantages. Firstly, dynamic algorithmic pricing helps undertakings to make their price adjustments faster. Firms can react instantaneously to changes in supply conditions – such as stock availability, capacity constraints or competitors’ prices – as well as to fluctuation in market demand58. Algorithms enable their users to adjust

their prices to the adequate price level. As a consequence, it will reduce excess supply and excess demand, thereby increasing market efficiency59.

Secondly, even though the way algorithms can monitor prices in digital markets is seen as an anti-competitive effect, it can also be seen as pro-competitive since monitoring becomes costless. The marginal cost is very low thanks to the limited human involvement. On the one hand, less staff is needed, which results in cost reduction. On the other hand, it may reduce the scope for behavioural biases60. This cost reduction could be afterwards passed on

to consumers.

Thirdly, it seems that for retailers, algorithms reduce the amount of specific market knowledge required to enter a market. Alternatively, existing retailers might find it easier to broaden their product offering and include products about which they may have less expertise61.

Without a definite answer as to whether pricing algorithms are pro- or anti-competitive, based on the current literature, we can however assert that there is a clear risk that pricing algorithms may facilitate tacit collusion.

54 A. Ezrachi and M.E. Stucke, supra n.30, p. 78.

55 Leibo, J.Z., Zambaldi, V., Lanctot, M., Marecki, J. and Graepel, T. (2017), ‘Multi-agent Reinforcement Learning in Sequential Social Dilemmas’, Proceedings of the 16th International Conference on Autonomous Agents and Multiagent Systems (AA-MAS 2017), São Paulo, Brazil.

56 SK Mehra, supra n.3, p. 1363. 57 Ibid., pp. 1363-1364.

58 OECD, supra n.1, p. 14.

59 Oxera Discussion Paper, ‘When algorithms set prices: winners and losers’, 19 June 2017, p. 15. 60 Ibid., p. 16.

(12)

III. Part 2: Enforcement challenges with regard to pricing algorithms

Since algorithms are already used in business strategies and have the potential of changing market structures, let us examine the enforcement challenges which stem from them. Monitoring algorithms already exist. Another question is: should we also consider enforcement challenges originating from self-learning algorithms? The question is legitimate since they are not representative (yet) of the current state of technology. Even though the latter could be seen as science fiction, it will no longer be the case in a few years’ time, considering the speed at which algorithms become more complex. Hence why the European Commission is already studying this matter: it is always better to be safe than sorry62.

These algorithms may do harm in the present or in a near-future competition, which raises the question: what should competition authorities do against them? Two approaches will be examined hereinafter: the legal approach under competition law and alternatives approaches to the first one.

1. Legal approach under EU competition law

We should keep in mind that explicit collusion is illegal whereas mere conscious parallelism is not. However, the use of algorithms put several business behaviours in a grey area which goes beyond conscious parallelism, but at the same time does not involve an express agreement between competitors63.

The interaction of business strategies based on pricing algorithms and competition rules in digital markets raise new challenges for competition authorities. If we combine the concern of pricing algorithms and the oligopoly problem, firms can set prices at a supra-competitive level without actually communicating in an easier way than before. In other words, algorithms can amplify the oligopoly problem and make tacit collusion a more frequent market outcome64. Moreover, a background note from the OECD Secretariat goes a

step further, stating that ‘algorithms might affect some characteristics of digital markets to such an extent that tacit collusion could become sustainable in a wider range of circumstances possibly expanding the oligopoly problem to non-oligopolistic market structures’65.

The main question is to know whether the current legal competition instruments are able to tackle the new difficulties inherent to pricing algorithms in digital markets. Even though EU competition law is in principle flexible enough to catch parallel behaviours when they do not correspond to the normal market conditions and competition authorities are able to avoid tacit collusion through merger control, the question remains open regarding the actions of non-dominant companies in using pricing algorithms whilst acting independently; these do not fall within the current competition law framework, even if such use ultimately results in higher prices for consumers. Also, which legal instrument would be best in order to tackle algorithmic tacit collusion, a quite recent phenomenon?

a) Liability

62 http://www.algoaware.eu/ 63 OECD, supra n.1, p.18. 64 Ibid, p.33.

(13)

Conscious parallelism is not unlawful in itself. Whilst dynamic pricing might have an anti-competitive effect, competition authorities will experience difficulties in finding a legal basis for intervention. Reading the statements made by German and French competition authorities in their joint report, it appears that: ‘[...] prosecuting such conducts could prove difficult: first, market transparency is generally said to benefit consumers when they have – at least in theory – the same information as the companies and second, no coordination may be necessary to achieve [...] supra-competitive results’66. If we overstep these inherent difficulties

from pricing algorithms, two questions are fundamental: when would a firm infringe competition law and who should be held liable?

i. When?

Where is the red line which businesses must not cross when using algorithms? As noted by Ezrachi and Stucke in the case of self-learning algorithms, ‘defining a benchmark for illegality can be challenging since it requires assessing whether any illegal action could have been anticipated or predetermined by the individuals who benefit from the algorithm’67. It

seems that the answer could only be given on a case-by-case basis under an assessment which includes ‘a careful consideration of the programed instructions of the algorithm, available safeguards, reward structure and the scope of its activities’68.

Which provision under competition law should be triggered? There are two ways of answering this question. Before any potential infringement, an ex-ante approach through merger control prohibits the creation or strengthening of a collective dominant position through tacit collusion which significantly impedes effective competition. After the potential infringement, two provisions are possible: Article 101 TFEU prohibiting agreements, decisions and concerted practices which are harmful to competition; and Article 102 TFEU prohibiting abuse of dominance by one or more undertakings.

The following sections are divided into ex-ante and ex-post legal approaches against algorithmic collusion.

 Ex-ante: Merger control

One way of tackling challenges raised by algorithms would be through the enforcement of merger control rules in markets with algorithmic activities. Under EU competition law, rules already exist to avoid the creation or strengthening of a collective dominant position (through tacit collusion), which significantly impedes effective competition because it is prohibited69. Nonetheless, are they able to catch all the possible anti-competitive

effects of pricing algorithms?

We already examined that pricing algorithms are increasing the likelihood of tacit collusion, especially with regard to two structural characteristics: market transparency and frequency of interaction. Therefore, there is no reason why competition authorities should not take into account pricing algorithms while assessing whether there are concerted effects in the meaning of the Guidelines of Horizontal mergers, since these two structural characteristics are

66 Autorité de la Concurrence & Bundeskartellamt, Competition Law and Data, 10 May 2016. 67 OECD, supra n.1, p. 38.

68 Idem.

69 See Guidelines on the assessment of horizontal mergers under the Council Regulation on the control of concentrations between undertakings, OJEU C 31, 05.02.2004 from para. 39, on "coordinated effects".

(14)

taken into account70. It appears that the current merger control framework can indeed tackle

tacit collusion because the conditions and concepts are broad enough to embrace pricing algorithms.

However, algorithms will probably change the way competition authorities control mergers. Firstly, algorithms may expand the tacit collusion problem to formerly non-oligopolistic market structures. Therefore, the OECD Secretariat recommend that agencies ‘consider lowering their threshold of intervention and investigate the risk of coordinated effects not only in cases of 3 to 2 mergers, but potentially also in 4 to 3 or even 5 to 4’71.

Secondly, Ezrachi and Stucke suggest that agencies may need to reconsider the approach to conglomerate mergers when tacit collusion can be facilitated by multimarket contacts. In particular, they note that ‘one aspect of machine learning is to discover correlations in large data sets. Thus, the algorithms can ascertain and respond to punishment mechanisms in distinct product markets, which to the human may appear unrelated’72.

It is doubtful that merger control will be able alone to handle challenges raised by algorithms. Moreover, tacit collusion could be tackled only ex-ante and leaves unchecked mature oligopolies where mergers do not occur. Merger control will not be helpful neither for an individual nor for the rational market strategy of a non-dominant single firm. Thus, merger control should be combined with another tool that is capable to condemn – if we decide to condemn them – anti-competitive effects from pricing algorithms outside a merger context. Should Article 101 or 102 TFEU be triggered?

 Ex-post: Article 101 and Article 102 TFEU - Article 101 TFEU

Under EU competition law, parallel behaviour – which is a rational reaction to market characteristics – does not trigger liability, but it may however amount to strong evidence of a concerted practice under Article 101 TFEU ‘if it leads to conditions of competition which do not correspond to the normal conditions of the market, having regard to the nature of the products, the size and number of the undertakings and the volume of the market’73.

Article 101 TFEU can be invoked when ‘facilitating practices’ are used by undertakings, that is to say practices which make it easier for firms to achieve the benefits of tacit coordination74. For instance, a facilitating practice is an exchange of information that

artificially increases the transparency of the market and so makes parallel behaviour easier. The structure of the market is also very important when assessing the legality of a practice under Article 101 TFEU. For example, in a context characterised by exchange of information, if the market is oligopolistic, Article 101(1) TFEU is likely to apply75. Also, the General Court

stated that decision-making independence of undertakings is more important to protect in oligopolistic markets76. The problem is that under Article 101 TFEU and the concepts that

stem from it, little guidance is given as to whether more subtle forms of communication fall under the scope of competition law. How would Article 101 TFEU and its case law be triggered in digital markets with algorithmic collusion?

70 Idem.

71 OECD, supra n.1, p. 40. 72 OECD, supra n.1, p. 40.

73 Case 48/69 Imperial Chemical Industries (ICI) v Commission (Dyestuffs), ECLI:EU:C:1972:70, para. 66. 74 R. Whish and D. Bailey, supra n.8, p. 605.

75 UK Agricultural Tractor Registration Exchange OK (1992) L 68/19, para. 16. 76 Case T-141/94 Thyssen Stahl AG v Commission (1999) ECR II-347, para 302.

(15)

One answer could be that algorithms be recognised as a facilitating factor, as they make coordination easier (e.g. by monitoring prices very quickly) and more effective (e.g. by facilitating detection of cheating and the administration of punishment of deviations)77.

Algorithms could be seen as a facilitating factors through the concept of an

information exchange. The European Commission has proposed to stretch Article 101 TFEU by expanding the interpretation of the notion of ‘communication’ in order to bring cases of algorithms-enabled price matching within the scope of Article 101 TFEU78: ‘[t]hrough

repeated interactions, two firms’ pricing algorithms could come to “decode” each other, thus allowing each one to better anticipate the other’s reactions’79. Could the ‘decoding’ between

the two algorithms be considered as an information exchange leading to a concerted practice prohibited by Article 101 TFEU? Let us follow the reasoning of the Guidelines on Article 101 TFEU in order to know if pricing algorithms can be seen as an information exchange. Pricing algorithms are artificially increasing transparency in digital market80 and are increasing the

internal stability of a collusive outcome on the market by enabling the companies involved to monitor deviations81. Moreover price is considered as a strategic data and pricing algorithms

reduce strategic uncertainty82.

Firstly, let us start by the fact that algorithms are increasing transparency in digital markets. In order to avoid being at an algorithmic competition disadvantage, firms on digital markets have a strong incentive to get their own pricing-algorithms83. ‘The result is an

industry where all market participants constantly collect and observe in real-time rivals’ actions, consumers’ choices and changes in the market environment, creating thus a transparent environment’84.

Secondly, algorithms are increasing the internal stability of a collusive outcome on the market by enabling the companies involved to monitor deviations but also reduce strategic uncertainty. Complex algorithms with powerful data mining capacity are ‘in a better place to distinguish between intentional deviations from collusion and natural reactions to changes in market conditions or even mistakes, which may prevent unnecessary retaliations’85. If

self-learning algorithms are used, ‘the combination of machine self-learning with market data may allow algorithms to accurately predict rivals actions and to anticipate any deviations before they actually take place’86. Strategic uncertainty in the market arises as there is a variety of

possible collusive outcomes available and because companies cannot perfectly observe past and current actions of their competitors and entrants. This is not the case with pricing-algorithms and by ‘decoding’ each other, it is likely that they reduce the independence of competitors’ conduct on the market.

Therefore, could algorithms be qualified as an information exchange as such? When

77 OECD, supra n.1, p. 20. 78 OECD, supra n.13, p. 8. 79 Idem.

80 Guidelines on the applicability of Article 101 of the Treaty on the Functioning of the European Union to horizontal co-operation agreements, OJ C 11, 14.1.2011,para. 65.

81 Ibid.,para. 67. 82 Ibid., para. 61. 83 OECD, supra n.1, p. 22. 84 Idem. 85 Idem. 86 Idem.

(16)

we look at their impact on transparency87, collusion stability88 and strategic uncertainty89 but

also at the fact these algorithms are related to strategic data (prices)90, the answer is likely to

be positive. However, it should be noted that Article 101 (1) TFEU is triggered depending mainly on the structure of the market and the characteristics of the data. The strategic usefulness of data will depend on its market coverage, aggregation, age and frequency of exchange. The exchange must affect a sufficiently large part of the relevant market in order for it to be capable of having a restrictive effect on competition91. As things stand currently, it

is difficult to give a definitive answer on whether pricing algorithms can be qualified as information exchange or not.

The main hurdle is that the ECJ’s case law allows economic operators to adapt themselves intelligently to the existing and anticipated conduct of their competitors92.

Therefore, it is hard to determine whether the notion of ‘communication’ could be stretched to such extent. In the context of pricing algorithms, having an answer from regulators or courts regarding whether the interaction between algorithms can be understood as ‘information exchange’ would be welcome.

As things currently stand, even though competition authorities and courts may consider algorithms as facilitating factors of concerted practice or stretch the notion of exchange of information, actions of non-dominant undertakings using pricing algorithms whilst acting independently do not fall under Article 101 TFEU. Under the ‘predictable agent scenario’ where firms, without communicating between them, monitor their respective prices through their own algorithms with a strategy to maximise profits, the provision does not fit well. It fits even less under the ‘digital eye scenario’, where each algorithm quickly detects any competitive manoeuvre, and thus knows when and how to retaliate. We can therefore conclude that at present, Article 101 TFEU is not ready to tackle pricing algorithms and tacit collusion.

This raises the following question: should we redesign Article 101 TFEU to make it fit with the new digital era? Is the case law making conscious parallelism lawful still consistent when we bring algorithms into the equation93? While not deeming illegal the use of

intelligent responses to a competitor’s behaviour seems understandable, the use of algorithms makes this affirmation more questionable. The way it impacts market transparency and the frequency of interaction is serious enough to wonder whether Article 101 TFEU should be stretched in order to catch these new market strategies and their potential anti-competitive effects.

Since pricing algorithms seem to escape the scrutiny of Article 101 TFEU (except if it leads to a concerted practice), what should competition authorities do? Should they trigger Article 102 TFEU?

- Article 102 TFEU

87 Guidelines on the applicability of Article 101 of the Treaty on the Functioning of the European Union to horizontal co-operation agreements, OJ C 11, 14.1.2011,para.58 and 65.

88 Ibid., para. 58 and 61. 89 Ibid., para. 61.

90 Ibid., para. 61 and 73-74. 91 Ibid., para. 87-88. 92 Ibid., para. 61.

93 Case 199/92, P Hüls AG v. Commission (1999) ECR I-4287; Joined Cases

C-89,104,114,116,117,125,129/85, Ahlström Osakeyhtiö and others v. Commission (Wood Pulp II), (1993) ECR I-1307.

(17)

The ECJ opened up the possibility that the conduct of oligopolists, though not within Article 101 TFEU, might be challenged under Article 102 TFEU with regard to collective dominance94.

In a nutshell, the ECJ described collective dominance as a dominant position held by two or more economic entities legally independent from each other, provided that from an economic point of view they present themselves or act together on a particular market as a collective entity95. Collective dominance may arise when the economic links result in joint

policies or activities96, even if those are not identical97. In order to establish collective

dominance, the ECJ does not require links in law, ‘other connecting factors’98 are enough to

trigger Article 102 TFEU during the economic assessment of the competition authority99.

Let us linger over the concept of ‘other connecting factors’. This will prove particularly interesting when competition authorities will have to deal with pricing algorithms. Indeed, it is conceivable that algorithms can fall under the scope of this concept in the same way that ‘facilitating factors’ do under Article 101 TFEU. These ‘other connecting factors’ will be assessed under economic criteria and, in particular, under an assessment of the structure of the market in question100. Considering what we discussed earlier regarding how

algorithms change several market characteristics, it is likely that algorithms will be assimilated as ‘other connecting factors’.

Thus, in principle, Article 102 TFEU would be a useful tool against algorithmic tacit collusion. However, things prove to be different in practice. In that regard, it is relevant to note that to date, there has been no Article 102 TFEU infringement decision in which a position of collective dominance has been found on the mere basis of tacit collusion; that is, in the absence of any other structural or contractual link101. Moreover, some have interpreted the

exclusion of collective dominance abuses from the scope of the 2009 Enforcement Priority Guidance as a sign that the Commission is reluctant to use Article 102 TFEU to tackle tacit collusion102.

Whish added that the Commission could not use Article 102 TFEU in any case for two reasons. Firstly, the case law has come to associate the economic concept of tacit coordination with the legal concept of collective dominance; and it is not an offence for a firm or group of firms to have a dominant position103. Secondly, tacit coordination occurs in certain

market conditions because the collectively dominant firms react rationally according to the conditions of the market on which they operate104. In his opinion, to condemn their parallel

behaviour as abusive in itself would amount to nonsense, and that is why the ECJ insisted in Hoffmann-La Roche that parallel behaviour should be condemned when it is attributable to an agreement or concerted practice contrary to Article 101(1) TFEU – it is not, in itself, abusive

94 R. Whish and D. Bailey, supra n.8, p. 610.

95 Joined Cases C-395/96P and C-396/96P, Compagnie Maritime Belge Transports v Commission, (2000) ECR-I I-1395, para. 36.

96 Idem.

97 Case T-228/97 Irish Sugar Plc v Commission, ECR II-2969.

98 Joined Cases C-395/96P and C-396/96P, Compagnie Maritime Belge Transports v Commission, (2000) ECR-I I-1395, para 45.

99 Idem.

100 Joined Cases C-395/96P and C-396/96P, Compagnie Maritime Belge Transports v Commission, (2000) ECR-I I-1395, para 45.

101 P. Siciliani, supra n.19, p. 1. 102 Ibid., p. 2.

103 R. Whish and D. Bailey, supra n.8, p. 616. 104 Idem.

(18)

under Article 102 TFEU105. Even though the second argument is somewhat perplexing, what it

is certain is that it would be contradictory to give firms the right to adapt themselves intelligently under Article 101 TFEU and then consider such behaviour as abusive under Article 102 TFEU.

Finally, Article 102 TFEU would not close the current gap about actions of non-dominant undertakings using pricing algorithms whilst acting independently. It therefore calls Article 102 TFEU into question when we try to tackle both monitoring algorithms and self-learning algorithms.

Is there a way to condemn algorithmic collusion without facing the hurdles stemming from price parallelism? We could assume that pricing algorithms cause unfairly high prices, explicitly condemned by Article 102(2)(a) TFEU. Here, the abuse would not lie in price parallelism, but in their level106. A national competition authority already condemned

excessive pricing in an oligopoly107. However, it raises the question of what should be

considered as an ‘unfair price’.

To sum up, Article 102 TFEU could be a useful tool, especially since the concept of ‘other connecting factors’ could encompass pricing algorithms. However, in practice, the provision is not used as a weapon against tacit collusion and undertakings still have to abuse from their collective dominant position; that is not the case in pure parallel conducts which do not always include dominant firms.

ii. Who?

Let us assume that an infringement of Article 101 or 102 TFEU has been found. Who should be held responsible? Commissioner Vestager warned that ‘companies can’t escape responsibility for collusion by hiding behind a computer program’108 and that ‘competition

enforcers need to be suspicious of everyone who uses an automated system for pricing’109.

What does this imply for algorithmic tacit collusion?

Before reflecting on who should be held liable, we need to make a distinction between the ‘predictable agent’ scenario and the ‘digital eye’ scenario. Under the predictable agent scenario, humans do not expressly collude but know that tacit collusion is the likely outcome if each adopted profit-maximizing pricing algorithms110. By contrast, under the

digital eye scenario, algorithms make decisions themselves without the approval of human beings.

In each scenario, the defendant would not be able to exonerate himself before a competition authority by stating that it was the algorithms which were colluding, and that he therefore did not have any anti-competitive intention. In the EU, an anti-competitive intention is not necessary to establish a restriction to competition in the sense of Article 101 TFEU111.

105 Idem. 106 Idem.

107 Telefonica, Vodafone and Orange, Spanish NCA decision of 19 December 2012.

108 https://ec.europa.eu/commission/commissioners/2014-2019/vestager/announcements/bundeskartellamt-18th-conference-competition-berlin-16-march-2017_en.

109 Idem.

110 A. Ezrachi and M.E. Stucke, supra n.29, p. 77.

111 Cases C-551/03 P, General Motors v Commission, ECLI:EU:C:2006:229, para. 77; C-8/08, T-Mobile Netherlands and Others, ECLI:EU:C:2009:343, para. 27; C-501/06 P et al., GlaxoSmithKline Services Unlimited v Commission, ECLI:EU:C:2009:610, para. 58; C-32/11, Allianz Hungaria Biztosito, ECLI:EU:C:2013:160, para. 37; C-67/13 P, Cartes Bancaires v Commission, ECLI:EU:C:2014:2204, para. 54 and C-286/13 P, Dole Food and Dole Fresh Fruit Europe v Commission, ECLI:EU:C:2015:184, para.118.

(19)

Liability is an important question and it is particularly problematic when pricing decisions are made by a machine using an algorithm, rather than by human beings using algorithms. Under the predictable agent, human beings could be held liable because they know that tacit collusion would be the outcome of their behaviour. As for self-learning algorithms, S.K. Mehra proposed three solutions in attributing responsibility: ‘to the robo-seller itself, to the humans who deploy it, or to no one’112.

Let us first rule out the hypothesis according to which there is no liability because it would provide immunity to anticompetitive conduct through the intermediary of an algorithm. It leaves us with two possibilities for the attribution of liability: the algorithm itself or the human being deploying it.

The main problem in attributing liability to the algorithm itself is that it does not fit well with the existing concepts within the Treaties. Of course, this issue does not arise when it is only an accessory to an overall agreement between competitors. However, things are different when price decisions are delegated to an algorithm and when humans have no ability to influence the way in which such decisions are made. EU competition law has not anticipated self-learning algorithms and their ability to act and establish prices autonomously. It is hard to assess to which extent the concept of ‘undertaking’ will cohabit with self-learning algorithms, the latter not being considered as a legal person. S. Chopra and F. White argue that self-learning algorithms such as robo-sellers increasingly deserve recognition as true actors, beyond their function as mere tools113. However, recognizing algorithms as something

beyond a mere tool equals to a dog chasing its own tail. Indeed, even if algorithms had a legal personality, a natural or legal person should be held responsible (and pay) for what the algorithms eventually infringed. It is hard to conceive, today, that an algorithm could possess assets which could be used to pay fines.

Therefore, in the present context, the easiest way to attribute responsibility seems to be holding human beings liable in both scenario. EU Commissioner Vestager stated that ‘what businesses need to know is that when they decide to use an automated system, they will be held responsible for what it does. So they had better know how that system works’114.

Nonetheless, we still need to determine who should be held responsible. To answer the question, competition authorities have different tools at their disposal.

If liability is indeed to be attributed to human beings, who should take the blame? Since undertakings have to comply with antitrust rules, both the developer and the user of the algorithm would be held liable. On the one hand, users have to ensure that their conduct do not restrict competition. On the other hand, the ECJ’s case law held liable intermediaries which facilitate collusion by competitors115.

The ECJ has been clear on the fact that an undertaking cannot adopt a passive attitude and that liability for an anticompetitive infringement can arise from its inaction. Moreover, undertakings which collect commercially sensitive information from their competitors must take the necessary steps to ensure compliance with Article 101 TFEU116. In

other words, we can presume that undertakings have an obligation to own algorithms that comply with EU competition law: ‘if a designer or user of pricing bots fails to take the 112 SK Mehra, supra n.3, p. 1366.

113 S. Chopra and L. White, A legal theory for autonomous artificial agents (University of Michigan Press, 2011), ch. 4 and 5.

114 https://ec.europa.eu/commission/commissioners/2014-2019/vestager/announcements/bundeskartellamt-18th-conference-competition-berlin-16-march-2017_en

115 Cases T-99/04, AC-Treuhand v Commission, ECLI:EU:T:2008:256; T-27/10, AC-Treuhand v Commission, ECLI:EU:T:2014:59 and C-194/14 P, AC-Treuhand v Commission, ECLI:EU:C:2015:717. 116 Jan Blockx, "Antitrust in digital markets in the EU: policing price bots", paper for the Roundbond Economic Law Conference of 9 June 2017, pp. 4-5.

(20)

necessary steps to stop those bots from engaging in collusion, they can be liable for that collusion, regardless of whether the pricing bot is self-learning or not’117. This is however

only true under Article 101 TFEU, if undertakings have a parallel behaviour that leads to conditions of competition which do not correspond to the market characteristics118. Thus, only

these undertakings could be held liable, leaving the actions of non-dominant companies using pricing algorithms whilst acting independently free from restraints in terms of competition rules.

Another question is: when self-learning algorithms decide to tacitly collude without any influence from human beings, who should be held liable for them? A company will be held liable for any anti-competitive practices of their employees, even if the company tried to prevent such behaviour119. Self-learning algorithms are not much different from sales

employees learning and adapting the prices to market dynamics; undertakings should therefore be held liable for the actions of the pricing algorithms.

Under 102 TFEU, all the undertakings which hold the collective dominant position are held liable when they abuse it. There is an interesting difference with Article 101 TFEU in that the Irish Sugar case could have been the perfect jurisprudence for closing the gap when a firm using pricing algorithms acts independently and tackles the two aforementioned scenarios. Indeed, there is no need for a ‘collective’ abuse of collective dominance. An individual abuse by a firm in a collective dominant position seems to be enough120.

Nevertheless, since the firm needs to have a dominant position, the gap is still open in the current competition framework.

Overall, it should be noted that holding human beings liable instead of algorithms themselves would have the advantage of fitting with the current concept offered by EU competition law, thus making economy legislative.

Although the current framework appears have the tools for holding human beings liable, the European Commission is currently taking in consideration self-learning algorithms and made some specific recommendations about it. It highlighted that autonomous decision-making ‘may conflict with the current regulatory framework which was designed in the context of a more predictable, more manageable and controllable technology’ and that ‘the increasing degree of autonomy thus poses a challenge to the current regulatory framework as a natural or legal person needs to ultimately be held responsible for such an impact’121. The

Commission proposed to further discuss different legal approaches with stakeholders: - a strict liability regime;

- a liability regime based on a risk-generating approach (whereby ‘liability would be assigned to the actors generating a major risk for others and benefitting from the relevant device, product or service’), and

- a risk-management approach (whereby ‘liability is assigned to the market actor which is best placed to minimise or avoid the realisation of the risk or to amortise the costs in relation to those risks’)122.

117 Idem.

118 Case 48/69 Imperial Chemical Industries (ICI) v Commission (Dyestuffs), ECLI:EU:C:1972:70, para. 66.

119 See joined cases C-100-103/08, Musique Diffusion française v Commission, ECLI:EU:C:1983:158, para. 97, and, more recently, case T-588/08, Dole Food and Dole Germany v Commission, ECLI:EU:T:2013:130, para. 581.

120 Case T-228/97 (1999) ECR II-2969, para. 66.

121 European Commission, ‘Commission Staff Working Document on the free flow of data and emerging issues of the European data economy Brussels’, 10.1.2017 SWD (2017) 2 final, at 43.

(21)

As Ezrachi and Stucke point out, ‘one significant obstacle with a risk-based approach for algorithmic tacit collusion is our ability to understand the magnitude and likelihood of risk and the actuality of harm’123. Tacit collusion is hard to detect, and it is even harder when

algorithms play a role in it because it is not always possible to know exactly how they work. We should therefore be careful about who should be held liable and gather a substantive body of studies about pricing algorithms before creating a specific liability regime.

b) Evidence

Let us now assume an undertaking is under investigation for infringing EU competition law through its pricing algorithms. Another challenge falls to competition authorities and courts: the legal burden of proving algorithmic collusion.

The traditional difficulty with tacit collusion lies in proving that parallel behaviour is attributable to tacit collusion between firms and not to the oligopolistic structure of the market124.

Even though there has not been any case law on algorithmic tacit collusion yet, we can presume that it will be even harder than before for competition authorities to prove tacit collusion. Algorithms will pose new challenges in that traditional approaches for finding collusive activity by incentivising whistle-blowers are no longer likely to work. Moreover, ‘in an environment where algorithms are making autonomous decisions based on information in the public domain, and there is no record of pricing decisions, what would constitute evidence of collusive activity is unclear’125.

If there is evidence that an algorithm has been programmed in a particular way in order to soften competition, this will clearly be useful in assessing whether it can and does restrict competition. However, due to the difficulty of reaching the requisite legal standard for competition authorities and the fact that algorithms are sometimes impossible to audit, should new tools be added to the current framework?

Pricing algorithms are a potentially powerful device to circumvent antitrust rules. The Commission stated that in respect to self-learning algorithms, their autonomous decision-making may ‘conflict with the current regulatory framework which was designed in the context of a more predictable, more manageable and controllable technology’126. Therefore, it

is conceivable to give competition authorities tools as powerful as pricing algorithms to oppose any misuse of the latter. Another danger will be that competition enforcers will too readily reach the conclusion that parallel conduct means that there is collusion and thus agree on a false positive condemnation.

123 Idem.

124 Under Article 101 TFEU, the General court held that when the Commission establishes the existence of a concerted practice solely on the parallel behaviour of the firms in question as proof of a concerted practice, it must address any alternative explanations put forward by the parties regarding that behaviour and demonstrate why they are not convincing. In other words, parallel behaviour can be evidence of a concerted practice where there is no plausible alternative explanation. Case T-442/08 CISAC v Commission (2013) EU:T:2013:188, para. 134-81. Under Article 102 TFEU, when markets characteristics give rise to tacit collusion, collective dominance may be established without the existence of an agreement or links in law. The competition authority has however the duty to prove that the conditions to establish tacit collusion are fulfilled. Three-part test under which the undertakings involved are capable of (1) reaching a tacit agreement, (2) detecting breaches and (3) punishing deviations from the agreement. See Case T-342/99 Airtours plc v Commission (2002) ECR II-2585.

125 Oxera, supra n.59, p.4.

126 European Commission, ‘Commission Staff Working Document on the free flow of data and emerging issues of the European data economy Brussels’, 10.1.2017 SWD (2017) 2 final, at 43.

(22)

Trying to answer ‘when’ and ‘who’ should be held responsible for pricing algorithms highlights how complex it is for Article 101 and 102 TFEU to catch the potential anti-competitive effects of algorithms within the current case law. It also raises the fundamental question of whether we should condemn a firm for behaving rationally and developing, unilaterally, an algorithm which processes publicly available information while operating interdependently on the market. To sum up, reaching a legal standard for proving an infringement will be even more arduous than before for competition authorities.

Such hurdles raise the question of whether the core provisions of competition are the right answer to pricing algorithms and tacit collusion. Should we adjust the legal framework in order to catch algorithmic collusion through competition law or should the digital market be regulated?

c) Finding and termination of an infringement: Article 7 Regulation 1/2003

Before looking at the alternatives to the legal approach to algorithmic collusion, we should keep in mind that Article 7 Regulation 1/2003 may help the European Commission in tackling pricing algorithms challenges beyond merger control and imposing fines under Article 101 or 102 TFEU trough remedies.

 Behavioural remedy

One way of dealing with tacit collusion in digital markets is to prevent the use of pricing algorithms insofar as they facilitate tacit collusion. However, determining what we should indeed forestall is a thorny issue. The ‘extreme’ answer under this behavioural approach would be to make any parallelism in price between undertakings through pricing algorithms illegal. Such measure – in addition to not making sense economically speaking – would not fit with the case law, since firms can behave in a parallel manner provided that they do so as a rational response to the structure of the market127.

A behavioural remedy may either be a negative or positive obligation. An appropriate answer would be to prohibit pricing algorithms that encourage parallel behaviour in a way that goes beyond mere conscious parallelism. This idea is, however, utopic. If prohibiting certain types of algorithms in oligopolistic markets is conceivable, in order to be efficient we first need empirical evidences on a specific kind of algorithm.

 Structural remedy

Where an infringement of Article 101 or 102 TFEU is the consequence of the structural conditions of the market and there is no effective behavioural remedy, Article 7 of Regulation 1/2003 provides the possibility of structural remedies to bring that infringement to an end128.

The issue is that Article 7 Regulation 1/2003 can be triggered only if there is an infringement of Article 101 or 102 TFEU. In the current competition framework, it is still uncertain whether these provisions bite on pricing algorithms.

127 R. Whish and D. Bailey, supra n.8, p.601. 128 Idem.

Referenties

GERELATEERDE DOCUMENTEN

Alongside this, new modes of economic governance are evolving on the European and national levels, triggering harmonisation and market integration through the application

Tabel 3. Plaaginsecten 'Top Twintig' van 50 jaar. Totaal aantal soorten gemelde plaaginsecten op loof- en naaldhout in Nederland.. hout zoals Aesculus, Robinia en Platanus,

Gebleken is dat de teelt- en oogstkosten voor voederbieten circa 400 gul- den per ha hoger zijn dan voor mais.. Dit wordt met name veroorzaakt door de hoge kosten van het zaaizaad en

De ammoniakemissie in de afdeling voor vleesvarkens met een bolle dichte vloer voor in de hokken en alle roosters van metalen driekant, was gemiddeld over vijf ronden 1,9 kg

Toepassing van het product ter bestrijding van de peren- bladvlo moet aan het eind van de winter plaats- vinden, voor de bloeitijd, en een of twee keer worden herhaald.. De boom

struction. The DR of the MS-CLEAN image is slightly lower than that of CoSch-CLEAN. This can be explained by the fact that the data are dominated by the source, making it easier

Using the Politie Output Monitor, it is possible to gain an impression of the 'order of magnitude' of police efforts in various areas of activity.78 Roughly 135 million guil- ders

The most important results of this study were that the PURE participants required education on foods associated with weight gain, what foods and drinks to purchase, how