Ina Fey
Supervisor: Joe Goté
22 July 2018
The application of current antitrust law
to explicit collusion
by autonomously acting pricing algorithms
Thesis
European Competition Law and Regulation
2017/2018
Abstract
This thesis aims to answer the question whether the current antitrust law is well-equipped to deal with explicit collusion by pricing algorithms that act independently from its using com-panies. The paper is divided into three main parts: first, it outlines the exact circumstances of the later discussed scenario and the differentiation to other ways in which algorithms can be related to infringements of antitrust law. Then, it investigates whether antitrust law can be ap-plied in general to the specified algorithms, that is whether one of the terms ‘agreement’ or ‘concerted practices’ according to Art. 101 (1) TFEU is applicable to independent algorithms. Finally, it assesses if the company that uses the algorithm can be held liable for any breach of antitrust law that the algorithm is engaged in. It does so by applying the concept of attribution of liability for acts of employees to the employing firm.
Table of content
A. Introduction...5
B. Algorithms and modern E-Commerce...8
I. How do pricing algorithms work?...8
1. Definition and working method of algorithms...8
2. Deep learning machines...9
3. Dynamic pricing by means of algorithms...10
a) The concept of dynamic pricing...11
b) The particularities of dynamic pricing with algorithms...11
c) Examples for dynamic pricing in modern e-commerce...12
d) Dynamic pricing by taking competitors’ prices into account...12
II. Categories of possible collusive conduct by pricing algorithms...13
1. Algorithms as “messenger”...13
2. “Hub and spoke” conspiracy...13
3. Predictable agent...14
4. Assessment of the first three scenarios...15
5. Autonomous machine...16
6. Underlying scenario of this paper...17
C. Is an adoption of the current antitrust law necessary?...18
I. Applicability of antitrust law to deep learning algorithms...18
a) Definition of the term ‘agreement’...18
aa) European law definition...18
bb) U.S. definition...19
b) Issue with autonomously acting deep learning algorithms...19
aa) Critical voices against the application to algorithmic pricing...20
bb) The particular communication of deep learning algorithms...20
cc) A ‘mind’ or ‘will’ of deep learning algorithms?...21
c) The concept of concerted practices as a possible solution...21
aa) The European law concept of concerted practices...22
(1) The term ‘concerted practice’...22
(2) Distinctive factors between tacit collusion and concerted practices...23
bb) Applicability of the concept of concerted practices to deep learning algorithms ...24
d) Interim conclusion...25
II. Liability for employees and third parties...26
1. How is liability for others designed in current antitrust law?...27
a) Design in legislation and similar sources...27
b) Design in European case-law...27
aa) Employees – Suiker Unie, Jean Claude Becu and VM Remonts...28
bb) Third persons – FNV Kunsten Informatie and VM Remonts...28
c) The theory behind this design: the single economic entity doctrine...29
d) Interim conclusion...30
2. Can these concepts be applied to deep learning machines?...31
a) Against the applicability to algorithms...31
b) For the applicability to algorithms...32
aa) Possible counter-arguments...32
bb) Application of the single economic entity doctrine...33
c) Interim conclusion...34
D. Conclusion...36
A. Introduction
This paper investigates whether the current antitrust law is well-equipped to deal with poten-tial collusion by deep-learning algorithms in e-commerce. More specifically, it will focus on 1) whether the concept of “concerted practices” as interpreted in reference to Article 101 TFEU can be applied to determine if such algorithms have improperly colluded, and 2) whether the concept of attribution of liability for employees to their firms can be applied for such algorithms in order to determine who could be held liable for collusion by algorithms.
Algorithms as part of the digital environment and development have become an integral part of this world and almost every aspect of our lives. Modern E-commerce, too, would not be possible without them. Since they can operate more and more independently of their users, it becomes obvious that they do not only provide huge benefits, as for instance enabling us to efficiently allocate resources, to regulate complex tasks like traffic and to take the most ratio-nal decisions, but they also raise immense ethical, philosophical and legal challenges. A well-known example of these ethical and legal problems is Google’s driverless car. Behind most of these problems is the question who would be held responsible if the algorithm threatens or damages something that is considered as an important collective or individual good, as for in-stance consumer and general welfare in competition law.
This thesis will investigate whether the existing antitrust law is prepared to deal with the most sophisticated form of algorithms – that is self-learning algorithms that are able to act au-tonomously from their human users. More specifically, it will assess whether current cartel law provides tools and concepts to handle collusion by self-learning algorithms without any human intervention or whether we need an adaption of the antitrust law in order to face the upcoming challenges of sophisticated software.
So far, the technique for completely autonomously acting algorithms is still in its infancy; but given the speed by which these technologies have been developed in the recent years and decades, it seems possible that at some point in the near future, the artificial pricing intelli-gence of one company will be fully capable to put in place a joint profit maximization strat-egy with the artificial pricing intelligence of a competing company, without any human inter-vention, instruction or control. This would be a collusion in every meaningful sense as we
have collusion cases today, except that the roles now played by humans are then played by machines with artificial intelligence.
However, this thesis (as a legal paper) will not investigate in what exact technical ways algo-rithms could collude but take it as a postulate that they will be able to do so at some future time.1 Since this scenario is still a cutting-edge topic, this paper aims to open the legal
discus-sion on this topic and to offer some first possible approaches to the problem.
Methodology
In order to assess whether current antitrust law is prepared to deal with explicit collusion by such deep learning machines, I first tried to investigate whether cartel law can be applied in general to this scenario, meaning whether there could be an agreement or a similar form of collusion between the algorithms of competing firms. I decided for an evaluative approach: After taking a look at the definition and interpretation of the term ‘agreement’ I tried to sub-sume the conduct of algorithms under this definition. However, since the current definition refers to the existence of a will or mind and the (potential and future) existence of such things in artificial intelligence is one of the most discussed and controversial topic in computer sci-ence, it would have gone beyond the size of the thesis to investigate whether deep learning al-gorithms can fulfill the criteria of the current definition of the term ‘agreement’.
Instead, I investigated whether there is another option to apply antitrust law to algorithms without the necessary of using the term and definition of ‘agreement’. Indeed, the antitrust law provides some tools that do not rely on direct evidence of the manifestation of a will and nevertheless apply antitrust law. These tools are concerted practices.
I tried to find examples for the application of concerted practices in case-law and investigated whether the finding in the cases allowed for an application to deep learning algorithms.
As a second step, I moved on to the question that almost all law deals with in the end: the question who is held liable for what behavior and why. As the question “for what behavior” was answered with my focus on explicit collusion, I could concentrate at the two other ques-tions; I thought of existing concepts of attribution of liability in cartel law and came to the at-tribution for employees and other comparable third parties. My idea was that algorithms could be considered as “employees” (or rather tools) of their users so that the liability could be
at-1 Mathematical indication that they are able to do so: see for example: studies presented in OECD, ‘Algorithms and Collu-sion: Competition Policy in the Digital Age’ (2017) <http://www.oecd.org/competition/algorithms-collusion-competition-policy-in-the-digital-age.htm> accessed 17 July 2018 31; also agreeing on that this is “the current state of technology”: Ariel Ezrachi and Maurice E Stucke, ‘Algorithmic Collusion: Problems and Counter-Measures’ (2017) <http://www.oecd.org/offi-cialdocuments/publicdisplaydocumentpdf/?cote=DAF/COMP/WD%282017%2925&docLanguage=En> accessed 27 May 2018 para 58ff.
tributed to them in the same way that employers are held responsible for the conduct of their employees. For this part as well, I chose for an evaluative approach. First, I took a look into EU legislation and examined some case-law related to the liability for employees. Then I tried to define the theory behind the legislation and the judgments. In a third step, I considered ar-guments pro and contra the application of the findings to algorithms, that is the question of whether users of algorithms can be held liable for the algorithms in the same way as they are responsible for their employees.
B. Algorithms and modern E-Commerce I. How do pricing algorithms work?
This chapter will provide a short introduction in the terms and concepts of computer-based learning, in order to help the reader to understand the later-on developed basic scenario of this paper, and will then define on what exact type of algorithm the discussed scenario will be based on.
1. Definition and working method of algorithms
Even though a concrete universal definition of the term ‘algorithm’ is still missing to date,2
there are three commonly accepted aspects that are considered necessary to define ‘algo-rithm’: an algorithm needs 1) a given form of input over which it can 2) perform a pre-de-signed exact sequence of commands to generate 3) an output in a clearly defined format.3
The term “artificial intelligence” “refers to the broad branch of computer science that studies and designs intelligent agents.”4 A subfield of artificial intelligence is the so-called “machine
learning” which refers to intelligent machines that iteratively learn from data and experience through the use of algorithms.5
Depending on the learning pattern, there are three different types of machine learning algo-rithms6:
1) With supervised learning,7 the algorithm is given a set of “correct” example pairs of
input and output data (called “the training set”); it then figures out the general mathe-matical rule that connects input and output. The aim of this learning is to enable the al-gorithm to perform the mathematical operation on a new input in order to create cor-rect new output.
2) Instead of being provided with a correct answer as an example of “success”, the algo-rithm in case of unsupervised learning attempts to identify hidden structures and pat-terns from a data set that is not pre-classified or pre-categorized in any form, and then resorts to pre-coded measures as indicators of “success”.8
2 OECD (n 1) 8.
3 See for example: Avigdor Gal, ‘It's a Feature, not a Bug: On Learning Algorithms and what they teach us’ (2017) <https://one.oecd.org/document/DAF/COMP/WD(2017)50/en/pdf> accessed 7 June 2018 para 1,5; OECD (n 1) 8.
4 OECD (n 1) 9. 5 ibid.
6 Hereafter, this thesis will refer to machine learning algorithms only as ’algorithms’. 7 Gal (n 3) para 7; OECD (n 1) 9.
3) In a dynamic environment, the algorithm performs tasks and learns through a trial-and-error approach, the so-called reinforcement learning.9 In this learning type, data is
not given to the algorithm, but generated by its interaction with the environment. The algorithm receives a kind of feedback of its last actions in order to improve future ac-tions.
Reinforcement learning has to be distinguished from supervised as well as from unsupervised learning in two major regards: The difference between reinforcement learning and unsuper-vised learning is in the respective goals: unsuperunsuper-vised learning is usually used to find similari-ties and differences between data-points whereas the goal of reinforcement learning is to achieve a certain behavior determined by the user. Reinforcement learning is about to find a certain strategy for a game with specified rules.
The difference to supervised learning is that there is no strict pre-given set of labeled data (the so-called “training set” by which the algorithm can explore what answer to a certain question the user considers to be ‘correct’), but a constantly changing set of data, influenced by the al-gorithm’s last action.
2. Deep learning machines
As conventional machine learning systems have limitations in their ability to process large raw data, it is necessary to extract from such raw data the relevant features before applying the algorithm (a process called “feature engineering”). This feature engineering can either be performed by humans, which is costly and time-consuming, or through so-called deep learn-ing by which a computer system learns by imitatlearn-ing the activity of human neurons and creat-ing an artificial neural network.10
The concept is described by many terms such as artificial neural networks, machine learning algorithms or deep learning algorithm. These systems attempt to simulate the structure and the way of operating of the human brain at a basic level.11 The brain functions are simulated in a
way that allows information to be distributed throughout all the network components.12
9 OECD (n 1) 9; see also: Ashwin Ittoo and Nicolas Petit, ‘Algorithmic Pricing Agents and Tacit Collusion: A Technological Perspective’ [2017] SSRN Journal 5f.
10 OECD (n 1) 9.
11 Diego Rasskin-Gutman, Chess metaphors: Artificial intelligence and the human mind (MIT Press 2009) 79; Daniel Kle-fors, ‘Artificial Neural Networks’ (1998)
For this thesis, it is not necessary to further develop the characteristics and way of working of neural networks. However, the following brief overview aims to give a glance of the deep complexity a neural network is designed and operating with.
The smallest basic element of neural networks are simple processing elements called neurons which can receive input and produce output. The neurons are arranged in different layers of which at least three are necessary in every neural network13 (however, there are usually
multi-ple layers within a single neural network): input neurons interface the external environment to receive inputs; output neurons communicate the output to the user; in between, there are inter-mediary neurons. Each neuron is connected to other neurons within the same layer and to neu-rons in different layers. The connection between the neuneu-rons carries the output of one neuron as input to another neuron. A single neuron can receive input from many other neurons but produces only one output. Neural networks learn by adjusting the strength of the connections between the neurons to allow the overall network to output appropriate results.14
As only the input and the output layer are “visible” from outside and all the rest of the neu-rons on all layers between are hidden from view,15 the exact way in which a neural network
finds its outcome is intransparent to the user (therefore, their functioning is often referred to as a “black box”). This hiding/intransparency can create some legal problems, as explained below.16
Neural networks do not have to have neurons in a physical way; a certain type of composition of algorithms can also be considered as artificial neural network as long as it provides the de-scribed structure in its operation-process.
Even though deep learning algorithms can in principle learn by supervised learning as well, they are mostly used with unsupervised learning (so-called deep learning). The application of the reinforcement learning method is possible as well.
3. Dynamic pricing by means of algorithms
One option that deep learning machines could be used for is the algorithmic-driven setting of prices in e-commerce, the so-called dynamic pricing.
13 ibid.
14 Klefors (n 11). 15 ibid.
a) The concept of dynamic pricing
In principle, the term ‘dynamic pricing’ describes a method of pricing that responds flexibly to market fluctuations in regard to aspects such as demand, supply, competition price or sub-sidiary product prices.17 The aim of dynamic pricing is to optimize prices for customers by
charging different prices for the same goods to different customers based on their purchase habits.18
b) The particularities of dynamic pricing with algorithms
Pricing algorithms self-learn by reinforcement learning in a highly dynamic environment where various factors such as demand, supply and other factors are constantly changing. They learn how to best achieve their pre-coded target by monitoring these factors, setting a certain price at a time, monitor the outcome of this set prices and adapt future price-setting according to the monitored outcome. Their specific task is ‘prediction’ which means that they provide an assessment of the likelihood of future outcomes based on the analysis of historical data.19
The terms ‘dynamic pricing’ and ‘algorithmic pricing’ nowadays are often used to describe the same process.20
Indeed price-setting algorithms allow for an effective real-time adoption of the price accord-ing to consumer’s former individual choices and acts.21 These self-improving pricing
algo-rithms enable their users to collect and use all sorts of demand-related data by taking into ac-count numerous variables to set the best price for a specific product for that customer at that time and to respond more quickly to changes in demand, supply and other factors than is hu-manly possible.22
Even before the invention of software-based pricing companies theoretically could have charged different customers with different prices. But the costs of processing all the relevant data that could play a role in regard to the adaption of the respective optimized price had been prohibitive.23 The widespread use of algorithms nowadays in e-commerce made dynamic
pric-17 Samuel B Hwang and Sungho Kim, ‘Dynamic Pricing Algorithm for E-Commerce’ in Tarek Sobh and Khaled Elleithy (eds), Advances in Systems, Computing Sciences and Software Engineering (Springer Netherlands 2006) 149; Salil K Mehra, ‘Antitrust and the Robo-Seller: Competition in the Time of Algorithms’ (2016) 100 Minnesota Law Review 1323 <http://www.minnesotalawreview.org/wp-content/uploads/2016/04/Mehra_ONLINEPDF1.pdf> 1336.
18 Hwang and Kim (n 17) 149. 19 Gal (n 3) para 8; OECD (n 1) 11.
20 See for example: Hwang and Kim (n 17) 149; Mehra (n 17) 1336. 21 Mehra (n 17) 1339.
22 ibid 1339; ‘Dynamic pricing: In-depth Guide to Improved Margins’ (2018) <https://blog.appliedai.com/dynamic-pricing/> accessed 2 June 2018.
23 Robert M Weiss and Ajay K Mehrotra, ‘Online Dynamic Pricing: Efficiency, Equity and the Future of E-Commerce’ (2001) 6 Va JL & Tech 1
<https://heinonline.org/HOL/Page?handle=hein.journals/vjolt6&id=267&div=14&collection=jour-ing not only possible, but also commercially feasible because companies are able to gather and process a lot of personal and habit-related information provided by Internet users at a minimal transaction cost.24 Similarly, algorithms allow companies to change prices with only
a minimal amount of time and effort.25
c) Examples for dynamic pricing in modern e-commerce
Initially, pricing algorithms were mostly used by online retail companies such as Amazon or eBay. Second in making massive (and above all very successful) use of pricing algorithmic software has been the travel industry, in particularly websites for booking flight tickets.26 But
also electricity suppliers soon saw the advantages of algorithmically adjusting prices based on supply and demand and employed price-setting algorithms on a large scale.27
However, the unsupervised use of pricing algorithms can lead to absurd results as shown by the skyrocketing price for a classic 20-year old biology textbook on fruit flies that attracted at-tention in 2011.28 Two 3rd party Amazon merchants had dynamic pricing models,
imple-mented by two different algorithms: while the first merchant’s algorithm set the price for the book about 0.998 times of the second merchant’s price, the second merchant’s algorithm set the price 1.271 time of the price of the first. In the end, this upward price loop made the com-peting algorithms setting the price at $23.6 million before the retailers noticed.
d) Dynamic pricing by taking competitors’ prices into account
As this thesis will focus on a potential collusion by pricing algorithm, it will only take as a ba-sis such algorithms that adjust prices exclusively by comparing the prices of their competitors. These competitive-response algorithms monitor pricing data from competitors and the impact of those prices on a company’s customers to set optimized own prices,29 as described in the
example of the Amazon book on fruit flies. This thesis will not consider algorithms that charge different prices to different buyers by taking into account other factors than competi-tors’ prices. Such behavior might be problematic in competition law regards, too, for example with respect to price discrimination. But it does not affect antitrust issues caused by collusive behavior.
nals> para 2ff. 24 ibid para 2. 25 ibid.
26 Hwang and Kim (n 17) 13. 27 Ittoo and Petit (n 9) 3.
28 Michael Eisen, ‘Amazon’s $23,698,655.93 book about flies’ (2011) accessed 18 May 2018. 29 (n 22); Hwang and Kim (n 17) 2.
II. Categories of possible collusive conduct by pricing algorithms
It is commonly accepted that modern pricing algorithms can not only work as a facilitating factor for collusion, but even could enable new forms of coordination that were not observed or not even possible before.30 This phenome is described with the term “algorithmic
collu-sion”.
Ezrachi and Stucke introduced four plausible ways in which algorithms could be used for or lead to collusion.31 In this chapter, these four scenarios will be described; however, the list is
not to be seen as exhaustive.32
1. Algorithms as “messenger”
In the first scenario, human beings agree on forming a cartel. Then they use the algorithmic software to assist in implementing this cartel: the algorithms monitor the agreed prices and policy the cartel by detecting deviation from the collusion.33
Clearly the use of algorithms facilitates the implementation of a cartel tremendously. The ability of powerful data mining algorithms to make predictions enables them, for instance, to distinguish between intentional deviations from the cartel agreement and natural reactions to changes in market conditions.34 Furthermore, the fact that deviations can be so easily detected
and immediately retaliated eliminates the profitability and thus the incentive of companies to deviate.35
Yet, the software is only used to execute the will of humans in their quest to collude.36 The
underlying “meeting of minds”, viz. the collusive agreement, is still concluded purely by hu-man beings.
2. “Hub and spoke” conspiracy
30 OECD (n 1) 19.
31 Ariel Ezrachi and Maurice E Stucke, ‘Artificial Intelligence & Collusion: When Computers Inhibit Competition’ [2015] accessed 9 April 2018.
32 See for example: here not discussed: the possibility of “signalling algorithms”, OECD (n 1) 29ff.; European Union, ‘Algo-rithms and Collusion - Note from the European Union’ (2017) <https://one.oecd.org/document/DAF/COMP/WD(2017)12/en/pdf> accessed 9 July 2018 para 27; Antonio Capobianco and Pedro Gonzaga, ‘Algorithms and Competition: Friends or Foes?’ [2017] Competition Policy International <https://www.com-petitionpolicyinternational.com/wp-content/uploads/2017/08/CPI-Capobianco-Gonzaga.pdf> accessed 27 May 2018 3. 33 Ezrachi and Stucke, ‘Artificial Intelligence & Collusion: When Computers Inhibit Competition’ (n 31) 8; European Union (n 32) para 37.
34 OECD (n 1) 22.
35 ibid 22, 27; Capobianco and Gonzaga (n 32) 3; European Commission, ‘Final report on the E-commerce Sector Inquiry’ (10 May 2017) <http://ec.europa.eu/competition/antitrust/sector_inquiry_final_report_en.pdf> accessed 20 July 2018 para 33. 36 Ezrachi and Stucke, ‘Artificial Intelligence & Collusion: When Computers Inhibit Competition’ (n 31) 8.
The second category of using algorithms involves the use of a single algorithm to determine the market price charged by numerous sellers.37
Several firms outsource the creation of a dynamic pricing algorithm to the same IT provider or populate their algorithm with the same data pool.38 This results in a cluster of similar
verti-cal agreements with many of the industries’ competitors.39 The developer of the algorithm (as
the “hub”) provides the competitors (the “spokes”) with the same or very close versions of one algorithm. The use of a similar algorithm and the same data pool facilitates alignment of price decisions and gives rise to a hub-and-spoke conspiracy.40
While this scenario might raise problems in regard to antitrust enforcement (in particular with regard to the intention of the software users to collude), it does not pose fundamental ques-tions relating to antitrust liability as it is still human being acting to collude.
3. Predictable agent
The scenario of a predictable agent refers to the prediction of behavior and the enhancement of market transparency through the use of algorithms.41
Each firm designs its machine unilaterally to deliver predictable outcomes and react in a given way to changing market conditions, without jointly agreeing on anything with its competitors. The companies only have an independent economic self-interest to develop and rely on the al-gorithm. The firms might be aware of likely developments of other algorithms used by the competitors,42 yet they did not intend to act in any illegal way.
However, this use of algorithms in this way could create anticompetitive effects through inter-dependent actions of the algorithms.43 Furthermore, the use of algorithms allows for quicker
reactions to changes in market conditions, a better monitoring of competitors’ prices and a more efficient sanction of deviations from a certain price level – in sum: a high market trans-parency.
In particular in oligopolistic markets, a high market transparency is one of the main factors that facilitate conscious parallelism (also called “tacit collusion”).44 Tacit collusion describes a
situation in which anti-competitive coordination is achieved without any need for an explicit
37 ibid.
38 OECD (n 1) 28; Ezrachi and Stucke (n 1) para 30.
39 Ezrachi and Stucke, ‘Artificial Intelligence & Collusion: When Computers Inhibit Competition’ (n 31) 8. 40 ibid 8; OECD (n 1) 28.
41 Ezrachi and Stucke, ‘Artificial Intelligence & Collusion: When Computers Inhibit Competition’ (n 31) 16. 42 ibid 8, 16.
43 ibid 8.
agreement, but in which competitors are able to maintain by recognizing their mutual interde-pendence.45 As tacit collusion does not entail any form of coordination between competitors,
but is the normal outcome of rational economic behavior of each firm under certain market conditions,46 it falls outside the scope of competition law. In its case Züchner v Bayerische
Vereinsbank47 for example, the Court stated that intelligent adaption to competitors’ existing
or anticipated conduct, without any prior direct or indirect contact between the competing firms, is not illegal under Article 101 TFEU.
However, between explicit and tacit collusion there is a grey area of practices that do not in-volve an express agreement between competitors, but nevertheless go beyond mere tacit col-lusion (so-called “concerted practices”).48 These practices include for instance the facilitation
of communication.
That is why the proof of an intent of the competitors to act anti-competitively is central in this scenario.49
4. Assessment of the first three scenarios
It can be said from the above-described that the use of algorithms, in general, increases mar-ket transparency and the frequency of responses to the competitors’ actions – both two very relevant factors for collusion.50
Under the first two scenarios, it might still be complex to detect the existence of an infringe-ment and prove such an infringeinfringe-ment, due to the use of algorithms.51 Nevertheless,
competi-tion agencies can rely on existing antitrust concepts in order to assess algorithms either on their own or as practices ancillary to a main infringement.52 All cases that fall under one of the
first two scenarios only involve joint conduct by human beings reaching familiar price-fixing agreements and then implementing them algorithmically.53 The assessment of the behavior is
not different from setting prices manually to implement a collusion in the offline world.54 For
these reasons, the first two scenarios do not raise fundamental legal issues in regard to liabil-ity.
45 OECD (n 1) 19. 46 ibid.
47 Züchner v Bayerische Vereinsbank (1981) C-172/80 ECR -02021 (CJEU). 48 OECD (n 1) 19 f.
49 Ezrachi and Stucke, ‘Artificial Intelligence & Collusion: When Computers Inhibit Competition’ (n 31) 8. 50 Capobianco and Gonzaga (n 32) 2.
51 OECD (n 1) 33; Capobianco and Gonzaga (n 32) 3. 52 OECD (n 1) 33; Capobianco and Gonzaga (n 32) 3.
53 Dylan I Ballard and Amar S Naik, ‘Algorithms, Artificial Intelligence, and Joint Conduct’ [2017] Competition Policy In-ternational <https://www.sheppardmullin.com/media/article/1649_CPI%20-%20Ballard-Naik.pdf> accessed 27 May 2018 1. 54 European Union (n 32) para 24.
The third scenario, however, raises difficult questions of how to deal with tacit collusion by these sophisticated algorithms as the algorithms act independently from and possibly contrary to the intention of their users. A widely debated question therefore is if and how the existing antitrust law could deal with this issue.55
5. Autonomous machine
Autonomous machines (that means: the above-described self-learning neural networks) are surely the most sophisticated way in which algorithms could be used in price determination. In this scenario, companies unilaterally create and use an algorithm to achieve a given target,56
for example, profit maximization or a sales increase. Through self-learning mechanisms the algorithm determines independently the optimal strategy to achieve this target.57 As the
algo-rithms determine their strategy autonomously, they could find out that the implementation of a joint strategy or even some form of explicit collusion is the most efficient one to achieve the target. In this way they are able to collude without the need for any human intervention.58
Without any human being in charge anymore, this scenario is very problematic in respect to the concept of current antitrust liability.
As a matter of course, deep learning machines could create a case of tacit collusion as well. They could conclude independently that they best achieve their given target by implementing a cooperative equilibrium with prices above the competitive pricing level. Experiments in simulated environments have shown that neural networks are particularly good in reaching such a cooperative outcome. Yet, this hypothesis remains to be proofed in real markets.59
As mentioned, these algorithms work as a “black box”. Accordingly, it is impossible to assess the process through which they found their output. Hence it could not be ruled out how they use, for instance, information that they received from other deep learning algorithms or hu-man beings active for a competing firm. However, recent studies have demonstrated that deep
55 See for example: Ezrachi and Stucke, ‘Artificial Intelligence & Collusion: When Computers Inhibit Competition’ (n 31) ; Capobianco and Gonzaga (n 32); Mehra (n 17); Monika Zdzieborska, ‘Brave New World of ‘Robot’ Cartels?’ (2017) <http://competitionlawblog.kluwercompetitionlaw.com/2017/03/07/brave-new-world-of-robot-cartels/> accessed 13 June 2018.
56 Ezrachi and Stucke, ‘Artificial Intelligence & Collusion: When Computers Inhibit Competition’ (n 31) 9. 57 ibid.
58 OECD (n 1) 31.
learning algorithms are able to coordinate with each other and implement their own way of communication, even though they were not coded to do so.60
These recent developments considered, it seems possible that pricing algorithms in the near future can achieve a sense of communication and mutual commitment which fully enables them to collude in every meaningful sense with the pricing algorithm of a competing firm.61
6. Underlying scenario of this paper
The recent debate on pricing algorithms has focused mainly on tacit collusion or “conscious” parallelism by pricing algorithms of which self-learning algorithms are certainly capable as well.
This paper, however, will take a deeper look into the upcoming developments in the field of artificial intelligence and take as basis deep learning algorithms that are able to communicate with each other, independently from human intervention or pre-coding. The paper will focus on some form of mutual commitment by algorithms that goes beyond mere tacit collusion. The underlying scenario is that two or more deep learning algorithms autonomously collude with each other in a more than only tacit manner. Today, it is not yet fully understood in what exact way such algorithms would find a commitment. This paper assumes that one day in the near future they will be able to somehow explicitly coordinate their behavior.
The exact way in which deep learning algorithms find their outcome (see the comparison with a “black box”62) is invisible from outside – a particularity of pricing by deep learning
algo-rithms that is taken into account over the course of the following investigation. However, this paper bases on a situation in which there is at least some evidence for any form of explicit communication between the algorithms.
60 See for example: Matt Reynolds, ‘Chatbots learn how to negotiate and drive a hard bargain’ (2017) <https://www.newsci-entist.com/article/mg23431304-300-chatbots-learn-how-to-drive-a-hard-bargain/> accessed 24 May 2018; Eike Kühl, ‘Eine Sprache macht noch keinen Terminator’ (2017) <https://www.zeit.de/digital/internet/2017-08/kuenstliche-intelligenz-sprache-lernen-facebook-chatbot> accessed 24 May 2018; Daniel Mutzel, ‘Google-KI entwickelt Verschlüsselung, die selbst Google nicht versteht’ (2016) <https://motherboard.vice.com/de/article/8q8wkv/google-ki-entwickelt-verschluesselung-die-selbst-google-nicht-versteht> accessed 24 May 2018.
61 See also: European Union (n 32) para 28; Ballard and Naik (n 53) 1,3. 62 See ‘Deep learning machines’, point B. I. 2.
C. Is an adoption of the current antitrust law necessary? I. Applicability of antitrust law to deep learning algorithms
Art. 101 TFEU prohibits “all agreements [...] and concerted practices [...] which have as their object or effect the prevention, restriction or distortion of competition.” Thus, the first step in order to apply antitrust law in general is to answer the question whether there has been an agreement.
The issue that will be dealt with in this chapter is whether the current interpretation of the term ‘agreement’ covers collusive conduct of a deep learning algorithm that acts au-tonomously of its human users. Only if the current definition of an agreement is sufficient to handle the challenges brought up by these algorithms, then other questions, such as the attri-bution of liability, can follow. Otherwise, the term ‘agreement’ as a fundamental concept of antitrust law needs to be revised in order to go on with further issues regarding collusion by algorithm.
a) Definition of the term ‘agreement’
Most jurisdictions require some sort of a direct or indirect contact showing that firms have not acted independently from each other – the so-called “meeting of the minds”.63
aa) European law definition
The Treaty on the Functioning of the European Union itself does not provide definition or fur-ther information about what should be considered as an ‘agreement’.
European legislation leaves it to the courts to determine the notion ‘agreement’: “[…] the no-tions of agreements, decisions and concerted practices are autonomous concepts of Commu-nity law covering the coordination of behavior of undertakings on the market as interpreted by the Community Courts.”64
According to the European Court, the concept of agreement “centers around the existence of a concurrence of wills between at least two parties, the form in which it is manifested being unimportant so long as it constitutes the faithful expression of the parties’ opinion.”65 The
63 OECD (n 1) 19.
64 Council Regulation (EC) No 1/2003 of 16 December 2002 on the implementation of the rules on competition laid down in Articles 81 and 82 of the Treaty, 16 December 2002 (European Council) para 8.
Court stated that an agreement reflects “a concurrence of wills between economic operators on the implementation of a policy, the pursuit of an objective, or the adoption of a given line of conduct on the market, irrespectively of the manner in which the parties’ intention to be-have on the market in accordance with the terms of the agreement is expressed.”66 The Court
hence mentioned as conditions for an agreement, firstly, the existence of a common will/a mental consensus/sense of mutual commitment between the parties, secondly, the relation of the will to the adoption of a particular form of conduct or measure67 on the market and,
thirdly, some form of manifestation of this will, usually through some kind of communication.
bb) U.S. definition
In order to highlight the requirements for an agreement in competition law under most juris-dictions, this paper will now take a very short glance at the definition of the term in the U. S. law.
Section 1 of the Sherman Antitrust Act uses multiple terms to refer to an agreement.68 It
de-clares illegal “every contract, combination in the form of trust or otherwise, or conspiracy”. Since ‘agreement’ is an essential ingredient of a contract, combination or conspiracy,69 also
the U. S. antitrust law focuses mainly on an agreement to determine an illegal collusion. Ac-cording to the Supreme Court, an agreement involves “a unity of purpose or a common design and understanding, or a meeting of minds.”70 Thus, also the U. S. competition law decides on
the existence of an agreement by considering whether there has been a “meeting of the minds” or “mutual assent”.71
b) Issue with autonomously acting deep learning algorithms
Usually, when assessing a possibly illegal conduct, the law judges the nature of this conduct through a ‘human’ prism.72 As discussed above, the focal point of intervention is an
agree-ment or understanding reflecting a concurrence of wills between human beings as colluding entities.73
66 OECD (n 1) 37; Bayer v Commission (n 65) para 173.
67 Christopher Bellamy and others, European Community law of competition (7. ed. Oxford Univ. Press 2013) para 2033.
68 OECD (n 1) 37.
69 Donald F Turner, ‘The Definition of Agreement under the Sherman Act: Conscious Parallelism and Refusals to Deal’ (1962) 75(4) Harvard Law Review 655 <https://www.jstor.org/stable/pdf/1338567.pdf> accessed 10 July 2018.
70 OECD (n 1) 37; American Tobacco Co. v. United States (1946) 328 U.S. 781-816 (US Supreme Court) 810.
71 Mehra (n 17) 1352.
72 Ezrachi and Stucke, ‘Artificial Intelligence & Collusion: When Computers Inhibit Competition’ (n 31) 7.
One of the most important issues nowadays in this regard is therefore whether we can apply the theories of agreement that require some sort of ‘will’ or ‘mental consensus’ as well as a form of communication between the parties to autonomous actions by algorithms.
aa) Critical voices against the application to algorithmic pricing
Some may argue that the current theories of agreement do not fit the actions by such algo-rithms and propose to develop a new, clearer definition of what is an agreement in regard to the need to address certain forms of “meeting of algorithms” (that is: algorithmic collusion).74
These voices mainly argue that the existing theories of agreement rely on explicit direct com-munication75 which is very different from the method that algorithms fix prices nowadays.
At the present stage, algorithms fix their respective prices for a certain good exclusively by observing market data and then adjust (possibly interdependently) the prices in response to their observations.
The much-discussed scenario of algorithmic tacit collusion is also based on this principle: whilst the algorithms implement a strategy that is based on “parallel consciousness” due to their ability to monitor market factors such as price changes effectively and thus create a highly transparent market environment, they still do not communicate explicitly with each other, but set their prices solely in accordance to market observations.
Yet, these algorithms still act according to a programmed order that is controlled by a human being. They might each figure out, independently, that the best way to achieve their pre-coded target is to coordinate with each other, in the sense to not undercut each other in a price com-petition, but to implement a certain price level with which each algorithm can reach optimal outcome according to the pre-coded target to achieve.
However, they cannot deviate from the pre-coded order and cannot implement “own” deci-sions that vary from the coded behavior. Thus, these algorithms are still under the control of a human being in some regard.
bb) The particular communication of deep learning algorithms
All the critical voices in the legal research community thus refer to price-fixing by algorithms that still act according to the order and under the control of a human being.
Deep learning algorithms, however, could be able to find a way of cooperation by explicitly communicating with each other, completely independent from their operators’ coding or con-trol. That is why the voices that call for a redefinition of the term ‘agreement’ are not at issue
74 OECD (n 1) 36ff; Capobianco and Gonzaga (n 32) 4; Mehra (n 17) from 1359.
in this scenario: they mainly criticize the lack of explicit direct communication when it comes to algorithmic collusion. But, as just mentioned, deep learning algorithms are indeed able to communicate explicitly with each other and in the here discussed scenario they actually do so.
cc) A ‘mind’ or ‘will’ of deep learning algorithms?
Another problem that could indeed occur with the raise of deep learning algorithms in price-setting is whether the term ‘agreement’ allows for a “meeting of minds” or “concurrence of will” by artificial – in other words: non-human – intelligence. The question whether artificial intelligence can have something like a “will” or “mind” (that enables it to agree on some-thing) is a very in-depth philosophical problem that arises in many fields at the interface be-tween artificial intelligence and law and that also raises a lot of other ethical issues, not only the legal one that is dealt with in this paper.
If one comes to the conclusion that algorithms do not have a will or mind in the classical sense of the term, then it might be necessary to redefine or rather broaden the term ‘agree-ment’ in order to cover collusion by algorithms that is not based on a classical meeting of the minds. However, before tackling this immense challenge, it seems useful to first investigate whether the current antitrust law provides any tools to bypass a redefinition by relying on the existent tools. Perhaps there is another possibility to apply antitrust law without a new inter-pretation of the term ‘agreement’.
c) The concept of concerted practices as a possible solution
As mentioned above, the problem with deep learning machines is that the user does not know how the algorithms achieve their optimized outcome because he can only observe the
out-come of the process. The operator as well as the regulatory body can only assess a potential
infringement of antitrust rules by such machines in respect of the outcome of their data pro-cessing.
However, the same is true in regard to human beings: for instance, when ‘human operators’ decide to agree on a certain level of prices for a specific good, the regulatory body sometimes cannot find a concrete proof of the “thinking process” of the human actors of how they found their exact collusive agreement, but merely notice the outcome. Solely from the human behav-ior, for examples, raising prices at the same time that a competitor does so, it is not possible to
judge whether this adoption is the effect of mere observation of the market realities and intel-ligent alignment or the result of collusion.
Thus, the antitrust law has developed some tools in order to determine an illegal conduct without having evidence of the thinking process.
aa) The European law concept of concerted practices
In the absence of direct proof for a formal agreement, the category of concerted practices can be applied.76 The term ‘concerted practices’ refers to a conduct that is not attributable to an
agreement or a decision but may nevertheless amount to an infringement of antitrust law.77
Hence, contacts between competitors that do not amount to an agreement can still be charac-terized as a concerted practice.78
(1) The term ‘concerted practice’
A concerted practice is “a form of coordination between undertakings which, without having reached the stage where an agreement properly so-called has been concluded, knowingly sub-stitutes practical cooperation between them for the risks of competition.”79
The concept inherent in the TFEU provisions on competition requires each undertaking to de-termine its market policy and “the conditions which it intends to offer to its customers” inde-pendently from its competitors.80 As the Court stated in Suiker Unie (one of the first cases in
which it had to examine the conditions for a concerted practice) the concept, however, does “strictly preclude any direct or indirect contact between such operators, the object or effect whereof is either to influence the conduct on the market of an actual or potential competitor or to disclose to such a competitor the course of conduct which they themselves have decided to adopt or contemplate adopting on the market.”81 A concerted practices thus involves any
di-rect or indidi-rect contact between competitors with the object or effect to influence the market conduct of the other firm.82
76 OECD (n 1) 37.
77 Richard Whish and David Bailey, Competition law (Eighth edition, Oxford University Press 2015) 117. 78 ibid 106.
79 ICI v Commission (Dyestuffs) (1972) 48-69 (CJEU) para 64.
80 European Commission, ‘Guidelines on the Applicability of Article 101 of the Treaty on the Functioning of the European Union to horizontal co-operation agreements’ (14 January 2011) para 60; Suiker Unie and Others v Commission (1975) C-40/73 ECR -01663 (CJEU) para 173.
81 Suiker Unie and Others v Commission (n 80) para 174. 82 Whish and Bailey (n 77) 118.
A concerted practice can violate Art. 101 TFEU either by restriction of competition by object or by restriction by effect. The crucial factor in either case is that the potentially infringing conduct removes uncertainty about competitors’ market behavior which restricts competi-tion.83
For a concerted practice to restrict competition by object, there has to be an exchange of infor-mation. In Anic, for instance, European propylene producers attended meetings where they discussed about strategies, quotas and commercial policies in general.84 In doing so, they
re-duced uncertainty about their respective competitors’ conduct and thus independent action on the market. 85 Another landmark case about concerted practices by object was T-Mobile.86 Five
mobile operators in the Netherlands had one legal meeting where, inter alia, confidential in-formation came up between participants over the course of the discussion.87 Although it was
unclear whether these information had any actual effect on retail prices of the mobile carriers, the Court decided that there was a concerted practice. If companies receive information from their competitors, it is presumed that they used this information in their decision-making and stop acting independently if this information is capable of removing uncertainties as regards to any business policy.88
A restriction by effect by means of a concerted practice involves any other action that reduces uncertainty on the market which is subject to an assessment of the legal and economic context of the conduct.
(2) Distinctive factors between tacit collusion and concerted practices
In principle, the requirement of independence does not deprive economic operators from their right to adopt themselves intelligently to the existing and anticipated conduct of their competi-tors.89 Under certain market conditions such rational economic behavior of independent firms
can lead to the above-described problem of tacit collusion.90 Tacit collusion falls outside the
scope of Art. 101 (1) TFEU whereas concerted practices are covered by Art. 101 TFEU. That
83 T-Mobile Netherlands (2009) C-8/08 ECR I-04529 (Opinion of Advocate General Kokott) para 52. 84 Anic (1999) C-49/92 P ECR I-04125 (CJEU).
85 ibid para 41.
86 T-Mobile Netherlands (2009) C-8/08 ECR I-04529 (CJEU). 87 ibid para 12.
88 ibid paras 31-43.
89 Suiker Unie and Others v Commission (n 80) para 174; Wood Pulp (1993) Joined cases 89/85, 104/85, 114/85, C-116/85, C-117/85 and C-125/85 to C-129/85 ECR I-01307 (CJEU) para 3.
is why a clear distinction between tacit collusion and a concerted practice is important, how-ever often not easy.
Parallel conduct may be consistent with both tacit collusion and concerted practices. This problem occurred in Wood Pulp.91 The Commission found wood pulp producers guilty in
vio-lating Art 85 of the Treaty of Rome92 by forming a cartel agreement and increasing prices
si-multaneously. The Court, however, ruled that the evidence was not sufficient to uphold all parts of the Commission’s decision. It stated that “parallel conduct cannot be regarded as fur-nishing proof of concertation unless concertation constitutes the only plausible explanation for such conduct.”93 A concerted practice cannot be assumed if “some parallel conduct may be
re-garded as constituting a rational response”94 to market conditions such as a high degree of
market transparency95 or oligopolistic tendencies.96
Hence, there must be a precise and consistent body of evidence in order to establish a con-certed practice.97 The mere fact of parallel market behavior is not conclusive.
Having said that, any contact with an exchange of information between competing compa-nies98 is sufficient as such evidence for a concerted practice since tacit collusion describes a
situation in which anti-competitive coordination is achieved without any need for contact be-tween the competitors and as pure rational economic reaction to market conditions.99 The
dis-tinction thus can be made by assessing whether there has been any contact between the com-petitors capable of having the effect to reduce independence on the market.
bb) Applicability of the concept of concerted practices to deep learning algorithms
As shown earlier in this paper, the requirement of an agreement by the potentially colluding algorithms may raise some difficult issues; therefore, it seems more reasonable to remain with the concept of concerted practices as a tool to deal with algorithmic collusion.
This paper assumes that there is no significant difference between a human mind and a ‘digi-tal mind’ in regard to finding a collusive outcome under antitrust law.
91 Wood Pulp (n 89). 92 Now: Art 101 TFEU. 93 Wood Pulp (n 89) para 3. 94 ibid para 126.
95 ibid. 96 ibid. 97 ibid para 127.
98 As described earlier, see ‘The term ‘concerted practices’, point C. I. 3. a) aa). 99 See ‘Predictable Agent’, point B. II. 3.
Deep learning algorithms are supposed to work like the human brain. Both find their opti-mized outcome through a chaotic, unpredictable process which cannot be monitored from out-side, and it seems possible that deep learning machines will be able to communicate with each other, just as human beings are communicating.
It is indeed possible (under the here discussed scenario) to assess whether there has been any contact between the algorithms that goes beyond mere surveillance of each other’s pricing be-havior; it could be found out whether the algorithms actively exchanged any information about price related factors such as respective demand or supply instead of just monitoring each other’s price-setting.
The question of whether algorithms are able to have a will or mind in antitrust sense is still open. However, the definition of agreement does not necessarily require the acting entity to have a mind or will in a highly philosophical sense. The purpose of this requirement of a will or mind is primary to distinguish illegal anti-competitive behavior from “accidentally” similar market conduct. If the requirements for a concerted practice are taken into account, there is a sufficient basis to draw this distinction.
Neither the European Commission nor the Court requires a clear distinction between an agree-ment and a concerted practice from a legal point of view because the only important distinc-tion is the one between collusive and non-collusive behavior.100 As the European institutions
hence consider both form of violations as equal and a concerted practice could be imple-mented by a deep learning algorithm, the current antitrust law as interpreted in reference to Art. 101 TFEU can thus be applied to these algorithms.
d) Interim conclusion
As shown above, there are options to apply current antitrust law to algorithmic conduct so that Art. 101 TFEU would cover collusion by deep learning machines. The concept of concerted practices does provide a profound tool to distinguish in individual cases whether there has been an illegal anti-competitive conduct. It is moreover applicable to collusion by deep learn-ing algorithms.
Hence, there is no adoption of existent antitrust law necessary in order to apply it to poten-tially colluding pricing algorithms.
However, the main issue of attribution of liability would still not be solved. The question whether antitrust law is applicable in general relates just to the existence of any anti-competi-tive illegal conduct. But this does not answer the question of who should be held liable and thus punished for this illegal behavior. Clearly, a machine cannot be held liable because there are no antitrust tools nor any theories about holding a non-human being accountable for its ac-tions.
The question to whom the liability then should be attributed and in what way will be dis-cussed in the second chapter.
II. Liability for employees and third parties
If one entity, whether it is a human being or any artificial actor, is acting on behalf of another one, there are three options to attribute liability:101 to the acting entity itself, to the deploying
entity or to no one.
However, the logic and basic concept on which current antitrust law has been tailored and jus-tifiedwould not allow to opt for the choice to attribute liability to no one and thus to do noth-ing after an infrnoth-ingement.102 Such an inaction after (possibly heavy) infringements of antitrust
law through explicit collusion would de facto provide impunity for anti-competitive conduct, potentially put in place through the intermediary of an algorithm.103 Thus, to avoid that
“com-panies […] escape responsibility for collusion by hiding behind a computer program”,104 the
third option cannot be considered as a realistic one.
The option to attribute liability to the algorithm itself would also not go along well with exist-ing concepts of (antitrust) law:105 “at present, computer programs are instrumentalities of the
persons who use them. If a program malfunctions, even in ways unanticipated by its designers or user, the legal consequences for the person who uses it are no different than the current stemming from the malfunction of another type of instrumentality. That a program malfunc-tions does not create capacity to act as a principal or agent.”106 Furthermore, the deterrence
objective of the attribution of liability, that is any form of sanction, does not affect algorithms as they are not considered to have any form of moral sense so far. So holding an algorithm
li-101 See also Mehra (n 17) 1366.
102 ibid 47.
103 ibid; OECD (n 1) 39.
104 Margrethe Vestager, ‘"Algorithms and Competition", Speech at the Bundeskartellamt 18th Conference of Competition, Berlin’ (2017) <https://ec.europa.eu/commission/commissioners/2014-2019/vestager/announcements/bundeskartellamt-18th-conference-competition-berlin-16-march-2017_en> accessed 15 June 2018.
105 Mehra (n 17) 1366.
able would not pursue the main objective of attributing liability itself and therefore would take the attribution of liability ad absurdum.
As a result, liability for an algorithm’s action can only be attributed to the using or designing entity. The aim of this chapter is to investigate whether the existent antitrust law provides tools or concepts according to which such an attribution is possible or whether we need an ad-justment of the current antitrust law in this regard.
1. How is liability for others designed in current antitrust law?
In the past, antitrust law evolved on the assumption of personhood among the actors it seeks to regulate.107 Physical persons act on behalf of legal entities for instance as employees,
out-side contractors, service providers or independent consultant agents.108
a) Design in legislation and similar sources
The only written source in EU law which deals with the design of current antitrust law in re-gard to liability of companies for other actors is to be found in the Communication from the Commission concerning guidelines for the application of Art. 101 to horizontal agreements.109
It states that “companies that form part of the same ‘undertaking’ are not considered to be competitors for the purpose of Art. 101 TFEU; when a company exercise decisive influence over another company, they form a single economic entity110 and, hence, are part of the same
undertaking”.111
The wording of this paragraph only covers “companies” and not employees. However, the principle laid down in the guideline applies as well to employees and third parties acting on behalf of the undertaking.
b) Design in European case-law
In the absence of any explicit comments in legislation concerning the liability for employees and other third persons for anti-competitive behavior, there is a huge case-law body in respect
107 Mehra (n 17) 1352.
108 Christopher Thomas, Gianni de Stefano and Dina Jubrail, ‘Antitrust Liability for Anti-competitive Behaviour by Em-ployees and Contractors’ (2016) <https://www.hoganlovells.com/en/blogs/focus-on-regulation/antitrust-liability-for-anti-competitive-behaviour-by-employees-and-contractors> accessed 28 May 2018.
109 European Commission, ‘Guidelines on the Applicability of Article 101 of the Treaty on the Functioning of the European Union to horizontal co-operation agreements’ (n 80).
110 See ‘The theory behind this design: the single economic entity doctrine, point C. II. 1. c).
111 European Commission, ‘Guidelines on the Applicability of Article 101 of the Treaty on the Functioning of the Euro -pean Union to horizontal co-operation agreements’ (n 80) para 11.
to the attribution of third persons to companies. Following are some of these judgements listed as examples.
aa) Employees – Suiker Unie, Jean Claude Becu and VM Remonts
For the first time, the Court held in Suiker Unie112 that employees form an integral part of
their employing undertaking since they are an economic unit with the undertaking when act-ing within their employment relationship.113
In Jean Claude Becu,114 a case that dealt mainly with the qualification of dock workers in the
port of Gent, the Court of Justice held that a person who performs work for and under the di-rection of an undertaking is, for the duration of that employment relationship, incorporated into the undertaking. He thus forms part of an economic unit with the undertaking and does not himself constitute an ‘undertaking’ within the meaning of EU competition law.115 Neither
can the workers be regarded as constituting an undertaking collectively.116
In VM Remonts,117 the Court reiterated that “an employee performs his duties for and under
the direction of the undertaking for which he works and, thus, is considered to be incorporated into the economic unit comprised by that undertaking”.118 Any anti-competitive conduct of the
employee is thus attributable to the undertaking to which he belongs and that undertaking is, as a matter of principle, held liable under EU competition law for that conduct.119
In both the cases, liability was basically attributed by regarding the employees as part of the undertaking.
bb) Third persons – FNV Kunsten Informatie and VM Remonts
FNV Kunsten Informatie en Media v Staat der Nederlanden120 dealt with a Dutch trade union
for, inter alia, self-employed orchestra substitute workers. The Court ruled that an agreement with an association of self-employed persons would infringe Art. 101 TFEU as they are ‘un-dertakings’ within the meaning of Art. 101 TFEU “for they perform their activities as inde-pendent economic operators in relation to their principal”.121 This rule does not apply if the
re-ality of the matter is that the self-employed persons are in fact ‘false self-employed’ which is
112 Suiker Unie and Others v Commission (n 80). 113 ibid para 539.
114 Jean Claude Becu (1999) C-22/98 ECR I-05665 (CJEU).
115 ibid para 26. 116 ibid 27.
117 VM Remonts (2016) C-542/14 (CJEU). 118 ibid para 23.
119 ibid para 24.
120 FNV Kunsten Informatie en Media v Staat der Nederlanden (2014) C-413/13 (CJEU). 121 ibid para 26f.
to be decided on a case-by-case basis. The Court provides some decisive factors as to assess the concrete character of the work-relationship such as the question of who bears the financial risks.122 Only if the self-employed persons are in a situation comparable to that of employees
(‘false self-employed persons’),123 they are regarded as part of the undertaking and, thus, their
actions are attributed to the employing company.
In VM Remonts,124 the Court ruled under which conditions a company is held liable for
collu-sion through independent service providers supplying the company with its services. The Court decided that an undertaking is held liable under Art. 101 TFEU for anti-competitive conduct of an external service provider only if, inter alia, the service provider was in fact act-ing under the direction or control of the undertakact-ing – a situation which is, again, very compa-rable to that of an employee of the company concerned.
In these two cases, liability is attributed to the controlling company on the basis of economic rational.
c) The theory behind this design: the single economic entity doctrine
As mentioned above, Art. 101 (1) TFEU does not apply to agreements between two or more legal persons that form a so-called single economic entity.125 This concept suggests that
enti-ties associated with each other through common control operate as a single economic unity. Thus, these entities comprise one single undertaking. All infringements of antitrust law com-mitted by one of these forming entities (for instance by an employee) are therefore attributed to the undertaking.
This concept of attribution of liability is based on the following considerations:
Firstly, the concept of single economic entity is based on the allocation of economic force: an employee does not exert a competitive force on the market that is separate from that of the employing organization.126 In general, all the economic actions of the employee are
consid-122 ibid para 33-37. 123 ibid para 31. 124 VM Remonts (n 117).
125 See for example: Whish and Bailey (n 77) 95.
126 David Bailey and Okeoghene Odudu, ‘The single economic entity doctrine in EU competition law’ [2014] Common Market Law Review 1721 <http://www.kluwerlawonline.com/document.php?id=COLA2014136> 1735.
ered as acts of his company. If for example the employee sells or purchase any good on behalf of the company, it is exclusively the company that acquires rights and obligations.127
The second consideration is the aspect of control over the employee: the company has deci-sive influence over the employee128 and determines the overall policy of the employee’s
ac-tions. The employee does not enjoy autonomy in determining his course of action in the mar-ket but carries out the instructions issued to him by the controlling company.129 It is assumed
that there are effective mechanisms preventing the employees from acting against the interests of their employer.130 The company is seen as being able to give orders to its employees and
di-recting their conduct.131
Furthermore, it must be borne in mind that, as a rule, the company benefits financially and economically from an infringement of EU competition law. The attribution of liability (and thus, the imposition of a fine) to the company seems justified when the benefits of the em-ployee’s illegal activities accrues to the company.132 This applies all the more because the
main purpose of an antitrust fine is to absorb the financial benefits a company mad due to a cartel agreement.133
d) Interim conclusion
The general principle under EU competition law is thus that companies are held liable for competitive behavior of their employees and other persons under their control. Any anti-competitive conduct by an employee is attributable to their employer as the company is held liable “as a matter of principle”.134 This attribution is implemented in a very strict manner: It
is not necessary that there has been any action by or knowledge on the part of the principal managers of the company;135 attribution of liability is not excluded if the natural person who
entered into an agreement did not have authority to do so.136 Companies are even held
respon-127 ibid 1754. 128 ibid 1756.
129 Settled case-law, for instance: ICI v Commission (Dyestuffs) (n 79) paras 133f; Viho Europe BV v Commission (1996) Case C-73/95 P ECR I-05457 (CJEU) para 16.
130 Bailey and Odudu (n 126) 1744. 131 ibid 1745.
132 ibid 1753; see also Dow Chemicals v Commission (2012) T-77/08 (CJEU) para 101. 133 Bailey and Odudu (n 126) 1753.
134 Thomas, Stefano and Jubrail (n 108).
135 ibid.
sible if they can show that they have used their best efforts to prevent such behavior or if the employee was acting contrary to explicit instructions.137
2. Can these concepts be applied to deep learning machines?
Based on the current stand of law, algorithms are simply considered as tools, implying that their decision can be directly attributed to their human operators.138 Thus, the liability for
ac-tions of algorithms would be attributed to their companies just as the liability for “human” tools, that is employees.
However, it has to be discussed whether this still counts if algorithms take their decisions completely independently from their human users. The question is thus whether the above-de-scribed legal mechanism can also apply for liability for deep learning algorithms.
a) Against the applicability to algorithms
Some state that “imputing robo-sellers’ actions to the human that program or deploy them might work in some cases but would likely lead to highly unpredictable results.”139 The result
of such an assumption is that the concepts of existent antitrust law would not be applicable to algorithms.
The basic problem of deep learning algorithms is their operation that is often simplistically described as a “black box” process. They process raw data, resembling the human brain, and deliver output without revealing the relevant features that were behind the decision process. Hence, the processing of data and its outcome is somehow unpredictable. Some argue that the determination of liability for illegal behavior requires an assessment whether any illegal ac-tion could have been anticipated or predetermined by the individuals who benefit from the al-gorithm.140
Surely, there are cases in respect to the use of deep learning algorithms in which the operators could not have anticipated or predetermined the illegal conduct of an algorithm, even if they kept an eye on compliance with antitrust law when giving the algorithm instructions and im-plemented safeguard mechanisms. Hence, in these cases companies using algorithms could not be held liable for anti-competitive conduct of algorithms, at least not according to existing concepts.
137 Thomas, Stefano and Jubrail (n 108).
138 OECD (n 1) 39. 139 Mehra (n 17) 1362.
