21

Competition Law and AI

Thomas Cheng

I Introduction
The basic premise of this handbook is that the advancement and proliferation of
artificial intelligence (AI) are such that the impact of AI on the law can no longer be
ignored. Nowhere is this more true than in competition law, where the discussion
about the impact of algorithms on the enforcement of competition law has been
raging for years since the publication of Virtual Competition by Ariel Ezrachi and
Maurice Stucke in 2016.1 The book has sparked a hotly contested debate about the
technological feasibility of algorithmic collusion and what to do about it. The topic
has taken over journals and conferences. Article after article has been written about
it, and it is impossible to attend a conference where not at least a few panels are
devoted to issues raised by algorithms.2 The possible impact of algorithms on com-
petition law enforcement remains to be fully explored. Two topics have garnered
the most attention thus far: algorithmic collusion and personalised pricing.
Collusion is widely acknowledged as the cardinal sin in competition law. Per se
treatment, or summary condemnation, of price fixing and other collusive behaviour
is one of the few issues over which a global consensus exists in a field of law that is
otherwise mired in controversy over the appropriate treatment of business practices.
This consensus, however, is confined to express collusion, whereby firms collude
by reaching an agreement among themselves. The legality of tacit collusion, under
which firms achieve a collusive outcome through intelligent adaptation to market

1
Ariel Ezrachi and Maurice Stucke, Virtual Competition (Harvard University Press 2016).
2
Michal S Gal, ‘Algorithms as Illegal Agreements’ (2019) 34 Berkeley Technology Law Journal 67;
Ulrich Schwalbe, ‘Algorithms, Machine Learning, and Collusion’ (2019) 14 Journal of Competition
Law & Economics 568; Christophe Samuel Hutchinson, Gulnara Fliurovna Ruchkina and Sergei
Guerasimovich Pavlikov, ‘Tacit Collusion on Steroids: The Potential Risks for Competition Resulting
from the Use of Algorithm Technology by Companies’ (2021) 13 Sustainability 951; Burton Ong,
‘The Applicability of Art.101 TFEU to Horizontal Algorithmic Pricing Practices: Two Conceptual
Frontiers’ (2021) 52 International Review of Intellectual Property and Competition Law 189; Ana Pošćić
and Adrijana Martinović, ‘EU Competition Law in the Digital Area: Algorithmic Collusion as a
Regulatory Challenge’ (2021) EU and Comparative Law Issues and Challenges Series (ECLIC) 1016.

472

https://doi.org/10.1017/9781108980197.025 Published online by Cambridge University Press

 Competition Law and AI 473

conditions and competitors’ behaviour without resorting to direct communication,

remains unsettled, at least within the academic community.3 Algorithmic collusion
refers to the possibility that the use of algorithms by businesses to set prices for their
products and to monitor competitors’ pricing practices may facilitate collusion. In
its most extreme, and some may argue most fanciful, form, algorithmic collusion
encompasses autonomous collusive conduct undertaken by algorithms without
any human intervention beyond the initial adoption of the algorithm. Autonomous
algorithmic collusion raises difficult questions concerning the attribution of con-
duct by algorithms to firms and reopens the longstanding debate about the legality
of tacit collusion.
Ezrachi and Stucke have been the most fervent crusaders against autonomous
algorithmic collusion, arguing that it may signify the end of competition as we know
it.4 Other commentators dispute the technological feasibility of such collusion and
assert that algorithms in their present form are incapable of handling the complex-
ities and variability of real-world markets with a multitude of competitors and con-
stantly changing market conditions.5 Even if these other commentators are correct,
there remains the possibility that regulation may be necessary in the future if algo-
rithms acquire greater technological capability and become able to collude among
themselves. This chapter will attempt to address the appropriate legal treatment of
autonomous algorithmic collusion in light of the current evidence of its technical
feasibility and various theoretical considerations.

II Algorithms

A What Is an Algorithm?
Algorithms have become increasingly widely adopted by online merchants on both
sides of the Atlantic. One-third of the 1,600 best-selling products on Amazon in the
United States (US) in 2015 were sold through algorithms.6 These products were found
to have a tendency of having higher prices and sales volume.7 A European Union
(EU) sector study published in 2017 found that two-thirds of online retailers in the
3
Alexander Stewart-Moreno, ‘EU Competition Policy: Algorithmic Collusion in the Digital Single
Market’ (2020) 1 York Law Review 49, 67–68; Francisco Beneke and Mark-Oliver Mackenrodt,
‘Artificial Intelligence and Collusion’ (2019) 50 International Review of Intellectual Property and
Competition Law 109, 118–119; Guan Zheng and Hong Wu, ‘Collusive Algorithms as Mere Tools,
Super-Tools or Legal Persons’ (2019) 15 Journal of Competition Law & Economics 123, 134.
4
Ezrachi and Stucke (n 1) 31.
5
Schwalbe (n 2); Ashwin Ittoo and Nicolas Petit, ‘Algorithmic Pricing Agents and Tacit Collusion: A
Technological Perspective’ in Hervé Jacquemin and Alexandre de Streel (eds), L’intelligence artifici-
elle et le droit (Larcier 2017).
6
Le Chen, Alan Mislove and Christo Wilson, ‘An Empirical Analysis of Algorithmic Pricing on Amazon
Marketplace’ Proceedings of the 25th International Conference on World Wide Web (2016) 1337.
7
Timo Klein, ‘Autonomous Algorithmic Collusion: Q-Learning under Sequential Pricing’ (Tinbergen
Institute Discussion Paper, 2020) 11 <www.papers.tinbergen.nl/18056.pdf>.

https://doi.org/10.1017/9781108980197.025 Published online by Cambridge University Press

 474 Thomas Cheng

EU use pricing algorithms that automatically adjust their prices based on competi-
tors’ prices observed by the algorithm.8
Before examining how algorithms may affect competition, it is important to
understand what they are. The Organization for Economic Co-operation and
Development (OECD) defines an algorithm as ‘a sequence of rules that should
be performed in an exact order to carry out a certain task. Thus, an algorithm is an
instance of logic that generates an output from a given input, whether it is a method
to solve a mathematical problem, a food recipe, or a music sheet’.9 In computer sci-
ence terms, an algorithm can be understood as ‘a programmed procedure for solving
a mathematical problem in a finite number of steps that frequently involves repe-
tition of an operation’.10 According to Ulrich Schwalbe, the two defining features
of an algorithm are finiteness and definiteness. Finiteness refers to the fact that ‘an
algorithm always terminates after a finite number of steps’11 and definiteness to the
fact that ‘each step in the algorithm has to be precisely defined and the actions to
be carried out must be rigorously and unambiguously specified’.12 Algorithms hence
may not possess the same flexibility, creativity, and inductive analytical ability of the
human brain. It ‘thinks through’ and analyses issues by repeating the same logical
steps thousands, if not millions, of times until a pattern or answer emerges. What
it lacks in the way of human intelligence, it makes up for with virtually unlimited
capacity for repetition and sheer computational speed.

B Classifications of Algorithms
There are a number of ways to classify algorithms. They can be categorised by input
parameters, by function, by interpretability, and by learning method.13 In terms of input
parameters, an algorithm may vary depending on ‘data size, type, or level of detail’.14

1 Functional Classification
In terms of function, at least as far as digital retailers are concerned, algorithms can
be classified as monitoring, data collection, pricing, customer tracking and person-
alisation, and signalling algorithms.15 The most relevant ones for our purpose are

8
European Commission, ‘Final Report on the E-Commerce Sector Inquiry’ (2017) COM 229 5.
9
‘Algorithms and Collusion: Competition Policy in the Digital Age’ (OECD, 2017) 8 <www.oecd.org/
competition/algorithms-collusion-competition-policy-in-the-digital-age.htm>.
10
Pieter Van Cleynenbreugel, ‘Article 101 TFEU’s Association of Undertakings Notion and Its Surprising
Potential to Help Distinguish Acceptable from Unacceptable Algorithmic Collusion’ (2020) 65 The
Antitrust Bulletin 423, 426.
11
Schwalbe (n 2) 575.
12
Ibid.
13
Pošćić and Martinović (n 2) 1019.
14
Ibid.
15
Ibid.

https://doi.org/10.1017/9781108980197.025 Published online by Cambridge University Press

 Competition Law and AI 475

monitoring, pricing, and signalling algorithms. Monitoring algorithms, as their name

suggests, allow a firm to monitor the market, in particular competitors and custom-
ers, through the use of scraping.16 These algorithms are especially relevant in the
context of algorithmic collusion by helping competitors monitor each other’s prices
and permitting them to retaliate almost instantly against a defecting cartel member.17
Pricing algorithms help firms to ‘optimise pricing strategies by reacting faster to
changes, thereby incurring lower costs than human agents’.18 The first-generation
pricing algorithms were more straightforward and merely followed simple pricing
instructions.19 The second-generation ones are more complex. They do not follow
pre-set rules and instead react to changing market conditions.20 Firms can enter their
firm-specific data such as cost structure and distribution channels, and the algorithm
will generate a pricing decision.21 One of the key advantages of pricing algorithms is
speed. They dramatically speed up the process of price updates. It used to take weeks
if not months for a large brick-and-mortar retailer to change the prices of all the prod-
ucts in a store. An online retailer can now do the same in a matter of seconds.22
Signalling algorithms mostly serve a more nefarious purpose. They allow firms
to signal their pricing intentions to their competitors. These algorithms help firms
collude, either expressly or tacitly, by implementing ‘instantaneous price changes
in the middle of the night, which allows a company to give a glimpse of its future
prices to competitors equipped with sophisticated algorithms capable of decoding
these stealthy price announcements without consumers even knowing about it’.23
Algorithms can pursue price negotiations among themselves through these signals,
obviating the need for direct communication.24 Burton Ong analogises these algo-
rithms as ‘the digital equivalents of unilateral price announcements made in offline
markets, creating pricing focal points around which all market players converge
towards rather than making their own independent pricing decisions’.25

2 Classification by Interpretability
In terms of interpretability, the two main types of algorithms are black box and white
box algorithms. White box algorithms, also known as descriptive algorithms, are

16
Hutchinson, Ruchkina and Pavlikov (n 2) 953.
17
Ong (n 2) 197.
18
Stewart-Moreno (n 3) 55.
19
Lea Bernhardt and Ralf Dewenter, ‘Collusion by Code or Algorithmic Collusion? When Pricing
Algorithms Take Over’ (2020) 16 European Competition Journal 312, 317.
20
Ibid.
21
Ibid. 316.
22
Ariel Ezrachi and Maurice E Stucke, ‘Artificial Intelligence & Collusion: When Computers Inhibit
Competition’ (2017) 2017 University of Illinois Law Review 1775, 1780.
23
Hutchinson, Ruchkina and Pavlikov (n 2) 957.
24
Ibid. 955.
25
Ong (n 2) 197.

https://doi.org/10.1017/9781108980197.025 Published online by Cambridge University Press

 476 Thomas Cheng

designed as transparent and clear code blocks, in contrast to black-box algo-

rithms which are very much impenetrable. The white-box algorithms are almost
completely visible and understandable to humans with suitable knowledge and
equipment. Therefore, one can retrace steps leading to a certain price decision.26

In contrast, black box algorithms work

in the same way as human thought processes, which cannot be accurately inferred.
The existence of a black box is bound to prevent users from effectively controlling
all the outcomes when using machine learning algorithms, and obstruct the courts
from determining the intent of users by reverse engineering.27

The interpretability of algorithms is highly pertinent to the issue of the attribution

of algorithmic conduct to the firm. If it is impossible for the firm deploying an algo-
rithm to understand whether and how the algorithm has learned to collude with
other algorithms, which some commentators have suggested would be the case with
black box algorithms,28 an argument can be made that the firm should not be held
liable for the subsequent collusion autonomously adopted by algorithms. Some
commentators, however, have disputed whether an algorithm can truly be a com-
plete black box indecipherable to its owner.29 Others have argued that businesses
have little incentive to adopt a black box algorithm as they would want to under-
stand the basis for a particular pricing decision ‘to obtain better market insights’.30
In any case, even if an algorithm is truly a black box, there remains an issue of
whether the firm should bear responsibility for adopting a black box algorithm that
ultimately engages in collusion. It cannot be the case that a firm can outsource its
most important competitive function to an entity over which it exercises no control
and then claim to be absolved of all responsibility for any subsequent illegal conduct
even though it benefits from such conduct.

3 Classification by Learning Method

Lastly, in terms of learning method, the main distinction is between algorithms that
are capable of machine learning and those that are not. Adaptive algorithms ‘are,
essentially, sets of rules that dictate optimal responses to specific contingencies’31 and
‘must therefore be instructed to coordinate on one of many possible outcomes’.32
They are fixed in their capabilities and cannot improve autonomously. In essence,

26
Bernhardt and Dewenter (n 19) 335.
27
Zheng and Wu (n 3) 129–30.
28
Matteo Courthoud, ‘Algorithmic Collusion Detection’ (2021) 5 <www.matteocourthoud.github.io/
files/Algorithmic_Collusion_Detection.pdf>.
29
Gal (n 2) 108.
30
Beneke and Mackenrodt (n 3) 129.
31
Emilio Calvano and others, ‘Algorithmic Pricing: What Implications for Competition Policy?’ (2019)
55 Review of Industrial Organization 155, 158.
32
Ibid. 159.

https://doi.org/10.1017/9781108980197.025 Published online by Cambridge University Press

 Competition Law and AI 477

these algorithms do what they are programmed to do under specific instructions by

the programmer and do not stray beyond them. These algorithms present fewer con-
cerns for the purpose of algorithmic collusion because they ‘cannot collude unless
they are designed by their programmers to do so’.33 These algorithms do not follow
a competitor’s prices or retaliate against a defecting rival unless they are instructed
to. The programmer’s intent to collude will be plain to see. Collusion facilitated by
these algorithms does not require adaptations or novel approaches under existing
competition law. They are mostly deployed in the Messenger scenario described by
Ezrachi and Stucke, which will be explained subsequently.
Learning algorithms, in contrast, are capable of machine learning, which allows
their functions to improve over time. Machine learning is a subfield of artificial
intelligence which creates algorithms capable of learning from data, experience,
and experimentation.34 Unlike adaptive algorithms, learning algorithms do not fol-
low static programming instructions but instead ‘build a decision process by learning
from data inputs’.35 A key advantage of these algorithms is that they learn through
experimentation. They modify themselves over time to improve their performance
in light of what they have learned from past experiences.36 There is no need
to specify a model of the market, estimate the model, and solve for the optimal strat-
egy. The programmer instead chooses only which variables the strategy should be
conditioned on, how frequently the program must experiment, and how much weight
to give to more recent experience relative to the cumulated stock of knowledge.37

This is especially important in the context of algorithmic collusion because an algo-

rithm that is based on specified models would have much greater difficulty pursuing
algorithmic collusion. The programmer would need to build a specific model for
the market at issue and adjust the model any time market conditions change. The
model would likely be highly sensitive to its assumptions, which significantly limit
its applicability and adaptability.
Learning algorithms can be further classified based on the type of machine learn-
ing on which it relies, which includes supervised learning, unsupervised learning,
and reinforcement learning. Supervised learning refers to machine learning con-
ducted under human supervision. Under supervised learning, ‘an algorithm is pres-
ented with [labelled] example data and associated target values to predict correct
target values after training when confronted with new data’.38 In contrast, no human
supervision is involved in unsupervised learning. Under unsupervised learning,
the algorithm ‘attempts to identify hidden structures and patterns from unlabelled

33
Ibid.
34
OECD (n 9) 9.
35
Gal (n 2) 78.
36
Schwalbe (n 2) 576.
37
Calvano and others, ‘Algorithmic Pricing: What Implications for Competition Policy?’ (n 31) 160.
38
Schwalbe (n 2) 576.

https://doi.org/10.1017/9781108980197.025 Published online by Cambridge University Press

 478 Thomas Cheng

data’39 ‘by deducing structures or patterns in the input data to extract general rules’.40
Reinforcement learning is the most advanced form of machine learning among the
three. Under reinforcement learning, an algorithm ‘learn[s] to take actions in an
unknown but fixed environment to maximize some sort of cumulative reward’.41 The
algorithm is not trained directly to determine the best cause of action in a given envi-
ronment but is geared ‘to maximize the expected sum of discounted future rewards’.42
The reinforcement learning process does not identify the correct input-output com-
binations or rectify unsuccessful action.43 Instead, it emphasises the maximisation
of long-term returns through experimentation. Because of the reliance on repeated
experimentation, the learning process could be very long. In some of the experi-
mental studies on collusion by a type of reinforcement learning algorithm known as
the Q-learning algorithm, it was found that the algorithm would need hundreds of
thousands of rounds to settle on the final collusive outcome.44

III Algorithmic Collusion

A What Is Collusion?
Before delving into algorithmic collusion, it is important to clarify the meaning of
collusion as both an economic and a legal concept. Collusion involves competi-
tors agreeing with each other to coordinate their competitive actions to raise their
profits beyond what would be possible under unfettered competition.45 The OECD
defines collusion as ‘a joint profit maximisation strategy put in place by compet-
ing firms that might harm consumers’.46 The three distinguishing features of col-
lusion are hence (1) coordination among competitors that (2) raises their profits to
a supra-competitive level, thereby (3) harming consumers. Even though the para-
mount objective of competition law is to protect consumer welfare, not every kind
of collusive conduct that harms consumers is proscribed by competition law. Tacit
collusion, which is generally tolerated by competition law, can inflict as much harm
on consumers as does an express cartel. An argument can be made that the costs of
regulating tacit collusion can outweigh its benefits as doing so would be tantamount
to requiring a firm to ignore its competitors’ pricing decisions.47

39
OECD (n 9) 9.
40
Schwalbe (n 2) 577.
41
Ibid.
42
Ibid.
43
Ibid. 578.
44
Emilio Calvano and others, ‘Artificial Intelligence, Algorithmic Pricing, and Collusion’ (2020) 110
The American Economic Review 3267; Emilio Calvano and others, ‘Algorithmic Collusion with
Imperfect Monitoring’ (2021) 79 International Journal of Industrial Organization 102712.
45
Stewart-Moreno (n 3) 51.
46
OECD (n 9) 19.
47
Stewart-Moreno (n 3) 67–68.

https://doi.org/10.1017/9781108980197.025 Published online by Cambridge University Press

 Competition Law and AI 479

Both sides of the Atlantic emphasise that collusion requires an agreement among
competitors. In the US, the case law speaks of ‘a meeting of minds’,48 ‘a unity of pur-
pose or a common design and understanding’,49 or ‘a conscious commitment to a
common scheme designed to achieve an unlawful objective’.50 These metaphors are
all taken to require that there be evidence of a conscious agreement or mutual under-
standing among the parties to behave anticompetitively. The EU case law also refers to
‘a concurrence of will’.51 Francisco Beneke and Mark-Oliver Mackenrodt sum up the
EU position on agreement succinctly: ‘[t]he concept of agreement requires an expres-
sion or the joint intention of the undertakings to conduct themselves on the market in
a specific way’.52 Again, what sets an agreement under EU law apart from unilateral
conduct, which falls outside the scope of Article 101 of the Treaty on the Functioning
of the European Union (TFEU), is ‘the element of communication between rivals’.53
This emphasis on the existence of direct communication among the colluding firms
to constitute illegal collusion prompted Joseph Harrington to observe that under com-
petition law, ‘[c]ollusion is not unlawful’,54 ‘what is illegal is communication among
firms intended to achieve an agreement where an agreement is mutual understanding
between firms to limit competition’.55 Louis Kaplow has echoed this observation.56
Economists understand collusion somewhat differently. In a definition adopted
by many other economists, Harrington defines collusion as ‘when firms use strate-
gies that embody a reward–punishment scheme which rewards a firm for abiding
by the supra-competitive outcome and punishes it for departing from it’.57 To him,
the lynchpin of collusion is the reward-punishment scheme. This is because supra-
competitive prices can be achieved with or without collusion.58 The latter happens
in an uncompetitive oligopolistic market where firms do not compete vigorously
with each other. Prices may exceed the competitive level due to the lack of com-
petitive pressure, even though the competitors are not colluding with each other. A
reward-punishment scheme is critical to collusion because it ties ‘a firm’s current
conduct with rival firms’ future conduct’.59 It is this causal relationship, not supra-
competitive prices, that defines collusion.60

48
American Tobacco Co v United States, 328 U.S. 781, 809 (1946).
49
Ibid.
50
Monsanto Co v Spray-Rite Service Corp, 465 U.S. 752, 764 (1984).
51
Case T-41/96 Bayer v Commission, ECLI: EU: T2000:242, para. 69.
52
Beneke and Mackenrodt (n 3) 112.
53
Hutchinson, Ruchkina and Pavlikov (n 2) 956.
54
Joseph E Harrington Jr, ‘Developing Competition Law for Collusion by Autonomous Artificial
Agents’ (2019) 14 Journal of Competition Law & Economics 331, 340.
55
Ibid. 346.
56
Louis Kaplow, ‘Direct versus Communications-Based Prohibitions on Price Fixing’ (2011) 3 Journal of
Legal Analysis 449, 449–50.
57
Harrington Jr (n 53) 336.
58
Ibid. 334.
59
Ibid. 336.
60
Ibid.

https://doi.org/10.1017/9781108980197.025 Published online by Cambridge University Press

 480 Thomas Cheng

It seems that both in law and in economics, collusion is not condemned solely due
to its consequences of supra-competitive prices and consumer harm. There must be
evidence that the parties to the collusive scheme are in fact acting in concert. There
is hesitation to regulate any conduct that inflicts consumer harm because regulation
entails its own costs.61 What is required in law is evidence of direct communication,
which signifies an agreement or mutual understanding. Economists focus on the
existence of a reward-punishment scheme, which motivates parties to refrain from
cutting prices. This implicitly highlights the importance of incentives: what are the
reasons and motivations for firms to pursue and sustain supra-competitive prices?
This is unsurprising as economics has always focused on incentives for human and
firm behaviour.
The slightly different emphases between law and economics indicate their dispa-
rate attitudes towards tacit collusion. The insistence on an agreement and evidence
of direct communication means that tacit collusion does not generally fall within
the prohibition of collusion under competition law. Meanwhile, the existence of a
reward-punishment scheme does not require an agreement or direct communica-
tion between competitors. Firms may pursue price wars to punish a defecting rival
even when there is no express agreement among them. Economists hence do not
distinguish between express and tacit collusion. The existence of an express agree-
ment among the colluding firms makes no difference to them.
Economists have identified a number of structural characteristics that are condu-
cive to collusion. These are not pre-requisites for collusion in the sense that collu-
sion is impossible in their absence. But experience has taught us that markets that
share these characteristics are more likely to experience collusion. This is impor-
tant for our purpose because in trying to determine whether and how algorithms
facilitate collusion, the focus is on how algorithms render market characteristics
even more favourable to collusion. These structural characteristics include (1) a
concentrated market with few competitors; (2) symmetric competitors in the sense
of similarity in cost structure; (3) homogeneous products; (4) barriers to entry; (5)
market transparency; (6) stable demand; and (7) small and frequent purchases by
customers.62
A concentrated market makes it easier to collude because it is obviously easier
to coordinate the conduct of three firms as opposed to thirteen firms, for example.
Cost symmetry and product homogeneity mean that competitors are more similar
to each other, which improves their chance of reaching terms of coordination. High
barriers to entry reduce the likelihood that the collusive scheme will be undermined
by a new entrant. A transparent market and stable demand allow the colluding firms
to monitor each other’s compliance with the collusive scheme more effectively. And

61
Christopher R Leslie and Mark A Lemley, ‘Categorical Analysis in Antitrust Jurisprudence’ (2008) 93
Iowa L Rev 1207.
62
Ai Deng, ‘What Do We Know about Algorithmic Tacit Collusion?’ (2018) 33 Antitrust 88, 92.

https://doi.org/10.1017/9781108980197.025 Published online by Cambridge University Press

 Competition Law and AI 481

finally, small and frequent purchases by customers reduce the incentives for the col-
luding firms to cheat. The colluding firms are presumed to have greater incentives
to cheat when customers are prone to make big and lumpy purchases, which means
each instance of defection can be highly profitable.

B What Is Algorithmic Collusion?

The entire controversy regarding algorithmic collusion is premised on the idea that
algorithms facilitate collusion or can even consummate collusion with other algo-
rithms autonomously without the need for human intervention. The latter will be
referred to as autonomous algorithmic collusion for the rest of this chapter. The
arguments are perhaps more straightforward when algorithms are merely used to
facilitate collusion. If human agents agree to collude and algorithms merely facili-
tate it, the use of algorithms should not affect the legality of the underlying collusive
scheme. Just as the law draws no distinction between different means of commu-
nication among fellow colluding firms, be it by post, by telegraph, by email, by
WhatsApp, or even by human messengers, and condemns all cartels regardless, the
fact that a collusive scheme is consummated with the help of algorithms should
make no difference. It may also be possible to pursue the reliance on algorithms as
a facilitating practice under US antitrust law or as a concerted practice under EU
competition law. This would be the most promising route for regulating algorithmic
collusion if, for some reason, it is impossible to prosecute the cartel directly.
Autonomous algorithmic collusion, however, is a much thornier issue. Much
of the controversy is concerned with its technical feasibility. A number of promi-
nent commentators, such as Ariel Ezrachi, Maurice Stucke, and Michal Gal, have
argued that algorithms are well capable of achieving tacit collusion.63 A number of
experimental studies, most notably by Emilio Calvano and co-authors, have dem-
onstrated that algorithms, specifically Q-learning algorithms, are capable of tacit
collusion in certain experimental settings after a long period of experimentation.64
Meanwhile, opponents such as Nicholas Petit, Ashwin Ittoo, and Ulrich Schwalbe
maintain that autonomous algorithmic collusion remains a remote possibility and
there is no need for competition law to be concerned about it at the moment.65
Joseph Harrington sums up the issue the best:
Can [algorithms] learn to collude in a simple setting? Yes. With two [algorithms],
two prices, and a fixed environment, simulations show that collusion is more com-
mon than competition. Can [algorithms] learn to collude in an actual market set-
ting? We do not know, and I am skeptical of anyone who thinks they know. As we

63
Ezrachi and Stucke (n 1); Gal (n 2).
64
Calvano and others, ‘Artificial Intelligence, Algorithmic Pricing, and Collusion’ (n 44); Calvano and
others, ‘Algorithmic Collusion with Imperfect Monitoring’ (n 44).
65
Schwalbe (n 2); Ittoo and Petit (n 5).

https://doi.org/10.1017/9781108980197.025 Published online by Cambridge University Press

 482 Thomas Cheng

cannot dismiss the possibility that [algorithms] are able to learn to collude in actual
markets, it is prudent to find an appropriate legal response should they be able to
do so.66

Even if it can be conclusively shown that under the current state of technology,
algorithms are incapable of tacit collusion, advancement in technology may allow
them to do so in the near future. There is certainly no harm for the competition
law community to anticipate the problem and engage in a thorough discussion to
come up with suitable solutions. The competition law community has been too
slow to react to the rise of the Big Tech. It should not repeat the same mistake with
autonomous algorithmic collusion. The remainder of this chapter will proceed on
the premises that there is some plausible, but by no means definitive, evidence that
autonomous algorithmic collusion is feasible and that it is fruitful to start the discus-
sion now even in light of the uncertain evidence.
There are two types of autonomous algorithmic collusion. The first type involves
direct communication between algorithms, which qualifies it as express collusion
and is clearly illegal under US antitrust law and EU competition law. The only
issue would be whether such collusion among algorithms should be attributed to
the firms deploying them. Studies have shown that some algorithms can learn to
communicate with each other in order to achieve their purpose.67 Other algorithms
have the capability to decipher each other’s software code, which could arguably
constitute another form of algorithmic communication.68 The case would be par-
ticularly strong if there is evidence that the code was intentionally exposed and ren-
dered decipherable to third parties by the firm in the first place.
The second type is accomplished through intelligent and independent adap-
tation to competitors’ conduct by algorithms with no direct communication
between them. This is algorithmic tacit collusion. Tacit collusion is generally
taken to be legal under both US antitrust law and EU competition law. This, how-
ever, has not deterred some commentators from advocating a different approach
for algorithmic tacit collusion on a variety of grounds. The question is whether a
change in the current approach to tacit collusion is called for specifically in the
algorithmic context.
The foregoing discussion suggests that there are three main dimensions along
which to analyse and understand algorithmic collusion: the existence of direct
communication among the colluding firms or algorithms, the degree of algorith-
mic autonomy, and the extent of collusive human intent. If there is evidence of
direct communication among the colluding firms or algorithms, there is little dispute
regarding the legality of the conduct. If the collusive outcome is achieved through

66
Harrington Jr (n 53) 346.
67
Schwalbe (n 2) 596; Calvano and others, ‘Algorithmic Pricing: What Implications for Competition
Policy?’ (n 31) 166.
68
Gal (n 2) 87.

https://doi.org/10.1017/9781108980197.025 Published online by Cambridge University Press

 Competition Law and AI 483

intelligent and independent adaptations by the competitors, absent any direct com-
munication, the conduct constitutes tacit collusion and is currently regarded as legal.
The degree of algorithmic autonomy runs the gamut, from almost complete
autonomy with practically no human involvement after the initial selection and
deployment of a particular algorithm to minimal autonomy where the algorithm is
a mere instrument to consummate a collusive scheme concocted by human agents.
The legality of the collusive conduct presents no novel issues where the degree of
algorithmic autonomy is minimal. It is not much more than a good old human car-
tel with a sprinkle of algorithms on top. Where the degree of algorithmic autonomy
is strong or almost complete, the issue arises as to whether the algorithm’s conduct
can be fairly attributed to the firm. It would not be possible to hold the firm liable
for the collusive scheme absent attribution if the scheme is autonomously consum-
mated by algorithms unless strict liability is contemplated for any illegal conduct
subsequently perpetrated by the algorithm after it has been created and adopted.
This would be tantamount to imposing a duty on firms to take action to prevent
algorithms from entering into a collusive scheme later on.
The degree of collusive human intent, which can be viewed as a corollary of the
degree of algorithmic autonomy, can range from a complete absence of such intent
to the presence of a full collusive intent. There is a complete absence of collusive
human intent in the case of autonomous algorithmic collusion, where the human
agents who adopt the algorithms have no intention that the algorithms will collude.
In fact, they may have no knowledge of the algorithmic collusion. There is full col-
lusive intent where the collusive scheme is intended by and consummated among
human agents, and algorithms merely play a facilitative role.
In general, it is easier to condemn a collusive scheme where evidence of direct
communication is found. Both sides of the Atlantic require the existence of an agree-
ment in order to do so and an agreement can be readily proven where direct com-
munication among the colluding firms can be shown. The lack of evidence of direct
communication may require us to infer the existence of an agreement on other bases.
Condemnation is easier to justify in the presence of collusive human intent. A collu-
sive agreement can be more readily established when there is evidence of an intent to
collude. Such an intent will be harder to find if there is a high degree of algorithmic
autonomy, meaning that the collusive scheme is largely the result of the autonomous
decisions made by algorithms. Where algorithmic autonomy is almost complete and
the extent of human involvement is minimal, prohibition is only possible either by
holding the algorithms directly liable, which would necessitate the conferment of
legal personhood on algorithms as advocated by some commentators,69 or attributing
the conduct by algorithms to the firm deploying them. This of course would only be
possible if the underlying conduct is deemed to be illegal, which would require a
resolution of the debate about the legality of tacit collusion.

69
Zheng and Wu (n 3) 151.

https://doi.org/10.1017/9781108980197.025 Published online by Cambridge University Press

 484 Thomas Cheng

C Ezrachi and Stucke’s Classification

Any discussion about the legal treatment of algorithmic collusion ought to begin
with the classification put forward by Ezrachi and Stucke in Virtual Competition.
They describe four scenarios of algorithmic collusion: Messenger, Hub and Spoke,
Predictable Agent, and Digital Eye. Messenger ‘concerns the use of computers to
execute the will of humans in their quest to collude and restrict competition’.70
Under this scenario, human agents have reached an agreement to collude and use
algorithms to assist in the execution of the collusive scheme. The existence of collu-
sion among human agents is not in doubt. There is express collusion, which should
entail direct communication among the colluding firms and indicate the existence
of collusive human intent. The degree of algorithmic autonomy is minimal. The
collusive scheme is not consummated by algorithms, which are used as mere instru-
mentalities. It is perhaps a bit of a misnomer to call this algorithmic collusion. This
is merely algorithm-assisted human collusion. As suggested earlier, the involvement
of algorithms does not alter the nature of the underlying conduct, and the illegality
of the collusion is undisputed.
Hub and Spoke refers to a situation where ‘competitors use the same (or a single)
algorithm to determine the market price or react to market changes. In this scenario,
the common algorithm, which traders use as a vertical input, leads to horizontal
alignment’.71 The key feature of an algorithmic hub and spoke cartel is the use of
a common pricing algorithm to determine the prices charged by each cartel mem-
ber.72 This sets it apart from the Messenger scenario where the algorithm merely
executes the orders of a human agent.73
The degree of algorithmic autonomy is higher because algorithms are no longer
deployed as mere instrumentalities to execute human will. The extent of human
involvement is probably no more than the choice of the common algorithm, but
it should still exceed that in the case of autonomous algorithmic collusion. In hub
and spoke arrangements, the human agents may have intended the hub algorithm
to pass on information to each other and to help coordinate pricing among them.
Price coordination may not be the completely unintended outcome, as in the case
of tacit collusion, at least as in the case of Digital Eye.
The existence of direct communication cannot be taken for granted, and thus
the existence of a collusive scheme needs to be established on other bases. In many
cases, the parties may have intentionally chosen the same algorithm with full aware-
ness of each other’s choices. In the context of a hub and spoke arrangement, the
case law on both sides of the Atlantic tends to assume that the ‘spoke’ parties to

70
Ezrachi and Stucke (n 22) 1782.
71
Ibid. 1787.
72
Rob Nicholls and Brent Fisse, ‘Concerted Practices and Algorithmic Coordination: Does the New
Australian Law Compute?’ (2018) 26 Competition & Consumer Law Journal 82, 95.
73
Ibid.

https://doi.org/10.1017/9781108980197.025 Published online by Cambridge University Press

 Competition Law and AI 485

the arrangement communicate indirectly through the hub. Such communication

is assumed if it can be shown that the spokes are all aware that they are part of the
arrangement and coordinated prices are observed among the spokes. Illegality can
be quite easily established if it can be shown that ‘there is an intent to achieve pro-
hibited conduct, that is, if competitors act with knowledge about the potential pro-
hibited conduct’.74 As former Commissioner of the US Federal Trade Commission
Maureen Ohlhausen wittily observes, ‘If it isn’t ok for a guy named Bob to do it, then
it probably isn’t ok for an algorithm to do it either’.75 The case law on both sides of
the Atlantic such as Interstate Circuit,76 Toys ‘R Us,77 and the Apple e-books case78
in the US and Eturas79 in the EU firmly establish the illegality of algorithmic hub
and spoke arrangements where the parties involved have knowledge of each other’s
involvement.
The third and fourth scenarios described by Ezrachi and Stucke, Predictable
Agent and Digital Eye, are both premised on tacit collusion. Both scenarios are
likely to involve learning algorithms. Adaptive algorithms are unable to collude
autonomously. There are, however, fine distinctions between them. Both scenarios
constitute tacit collusion in the sense that an agreement or understanding among the
parties is absent. Under Predictable Agent, firms adopt their own algorithms inde-
pendently, but with an awareness of competitors’ adoption of similar algorithms.80
Ezrachi and Stucke are careful to emphasise that the firms have not agreed and do
not intend to collude.81 In fact, they are presumed not even to have agreed to adopt
similar algorithms. Ezrachi and Stucke are not clear as to whether the firms are
aware of the collusive potential of their adoption of similar algorithms. Under the
appropriate market conditions, the adoption of algorithms, especially similar ones,
by most firms in the market may facilitate tacit collusion and bring about higher
prices.82 And it may be possible to pursue the adoption of colluding algorithms as
facilitating practices. As Ezrachi and Stucke observe, the Predictable Agent scenario
‘raises challenging questions as to the ability to condemn the creation or strengthen-
ing of conscious parallelism through a sophisticated algorithm’.83
Given that both Predictable Agent and Digital Eye are premised on tacit collu-
sion, it may not be immediately obvious how they differ from each other. According
to Ezrachi and Stucke, under Digital Eye, competitors adopt algorithms that are set

74
Pošćić and Martinović (n 2) 1025.
75
Maureen Ohlhausen, ‘Should We Fear the Things That Go Beep in the Night?’ (23 May 2017) 10 <www​
.ftc.gov/system/files/documents/public_statements/1220893/ohlhausen_-_concurrences_5-23-17.pdf>.
76
Interstate Circuit v United States, 306 U.S. 208 (1939).
77
Toys ‘R’ Us, Inc v Federal Trade Commission, 221 F.3d 928 (7th Cir. 2000).
78
United States v Apple Inc, 791 F.3d 290 (2d Cir. 2015).
79
Case 74/14, ‘Eturas’ UAB v Lietuvos Respublikos konkurencijos taryba ECLI:EU:C:2016:42.
80
Ezrachi and Stucke (n 22) 1783.
81
Ibid. 1790.
82
Ibid. 1789.
83
Ibid. 1795.

https://doi.org/10.1017/9781108980197.025 Published online by Cambridge University Press

 486 Thomas Cheng

with the goal of profit maximisation.84 The algorithms (here they are most likely to
be reinforcement learning algorithms) independently determine through experi-
mentation and machine learning that consciously parallel pricing behaviour leads
to the highest profit and adopt this course of action.85 Ezrachi and Stucke emphasise
that tacit coordination here is ‘the outcome of evolution, self-learning, and indepen-
dent machine execution’.86
One possible distinction between Predictable Agent and Digital Eye may seem
to be that under the former, algorithms improve the suitability of market conditions
for tacit collusion, while the Digital Eye scenario is not premised on facilitation of
tacit collusion at all. Algorithms simply get on with parallel pricing behavior. This
distinction, however, may be difficult to draw in reality because pricing algorithms
that achieve parallel pricing probably also improve market conditions for tacit collu-
sion. It is very difficult to know where facilitation of collusion ends and actual tacit
collusion begins. It is also not clear how useful this distinction is. Under Predictable
Agent, after having improved market conditions for tacit collusion, it is the very
same algorithms that take advantage of these conditions to achieve parallel pricing.
The meaningful distinction between Predictable Agent and Digital Eye hence can-
not be the different roles played by the algorithms to achieve the collusive outcome.
Ezrachi and Stucke also suggest that Digital Eye ‘increases the complexity of iden-
tifying intent and distinguishing between the operation of the machine and that of its
designer’.87 The main difference between Predictable Agent and Digital Eye as far as
intent is concerned seems to be that under the former, competitors are aware of each
other’s adoption of similar algorithms that could lead to a collusive result, whereas
under the latter, the decision to adopt an algorithm seems to be made entirely inde-
pendently. This perhaps provides a more meaningful distinction between the two
scenarios. What makes Predictable Agent more suspect is the awareness on the part
of the human agents of the collusive potential of the adoption of the same algorithm.
Given that consciously parallel pricing behaviour results under both scenarios, there
is no difference in the role of the algorithm that can possibly justify disparate legal
treatment of the two scenarios. In contrast, the degree of human awareness of the
likelihood of collusive outcome may have a bearing on two legal issues: whether it
is possible to find an agreement among the firms to adopt similar algorithms and
whether it is justified to attribute the algorithm’s actions to the firms.
The lack of agreement among the firms under both Predictable Agent and Digital
Eye suggests that there is no direct or probably any kind of communication among
the firms, unless the adoption of similar algorithms can be treated as signalling.
Under Predictable Agent, if the firms have no awareness that their parallel adoption

84
Ibid. 1783.
85
Ibid.
86
Ibid. 1795.
87
Ibid. 1797.

https://doi.org/10.1017/9781108980197.025 Published online by Cambridge University Press

 Competition Law and AI 487

of similar algorithms may facilitate collusion, there will be no collusive human

intent. And since firms adopt algorithms in order to maximise profit under Digital
Eye, they are clearly devoid of a collusive intent. The degree of algorithmic auton-
omy under both Predictable Agent and Digital Eye is higher than that in the pre-
vious two scenarios because with tacit collusion, it is the intelligent adaptation by
algorithms to each other’s actions that results in collusion. The algorithms are not
instructed by human agents to take the actions that may ultimately result in tacit
collusion, be it monitoring of rivals, retaliating against defectors, or simply setting
profit-maximising prices.

D Autonomous Algorithmic Collusion

The illegality of Messenger and Hub and Spoke is not in doubt. Thus, much of the
debate concerning algorithmic collusion focuses on Predictable Agent and Digital
Eye, both instances of autonomous algorithmic collusion. The controversy regard-
ing autonomous algorithmic collusion in some ways resurrects the longstanding
debate, at least within academic circles, about the legality of tacit collusion.

1 Traditional Debate about Tacit Collusion

The longstanding debate about tacit collusion harkens back to Donald Turner
and Richard Posner, who actually later changed his views about the issue. Turner
believes that tacit collusion or conscious parallelism is the natural and inevitable
consequence of an oligopolistic market with high market transparency and homo-
geneous product.88 In such a market, firms will naturally refrain from price cutting
and follow each other’s price increases to reach supra-competitive prices. For the
sake of consistency, tolerance of monopolistic pricing means that competition law
should also condone tacit collusion.89 Moreover, a prohibition of tacit collusion
would require firms to behave irrationally and avoid maximising profit.90
Richard Posner asserts that tacit collusion should be prohibited just the same as
express collusion. Both constitute an illegal agreement under the Sherman Act as
they entail a meeting of the minds or a mutual understanding.91 What separates
tacit collusion from express collusion is only a matter of evidence. Express collusion
can be proved by documentary evidence, while tacit collusion relies on economic
evidence.92 According to Posner, ‘a seller communicates his ‘offer’ by restricting

88
Donald Turner, ‘The Definition of Agreement under the Sherman Act: Conscious Parallelism and
Refusals to Deal’ (1962) 75 Harvard Law Review 655, 665.
89
Ibid. 668.
90
Ibid. 669.
91
Richard A Posner, Antitrust Law (2nd edn, The University of Chicago Press 2021) 94.
92
Richard A Posner, ‘Oligopoly and the Antitrust Laws: A Suggested Approach’ (1969) 21 Stanford Law
Review 1562, 1576.

https://doi.org/10.1017/9781108980197.025 Published online by Cambridge University Press

 488 Thomas Cheng

output, and the offer is ‘accepted’ by the actions of his rivals in restricting their
outputs as well’.93 It is fair to say that the courts have largely taken Turner’s side of
the debate. Justice Stephen Breyer, when he was still a judge on the United States
Court of Appeals for the First Circuit, asserts that courts have upheld conscious
parallelism ‘not because such pricing is desirable (it is not), but because it is close
to impossible to devise a judicially enforceable remedy for ‘interdependent’ pricing.
How does one order a firm to set its prices without regard to the likely reactions of
its competitors?’94
This debate was recently rekindled by Louis Kaplow, who asserts that the cur-
rent approach to the concept of agreement under Section 1 of the Sherman Act
is misguided. It focuses on the form of the agreement based on the existence of
communication between firms instead of a more economic approach to collusion
under oligopoly theory.95 He notes that ‘successful interdependent coordination
that produces supra-competitive pricing leads to essentially the same economic
consequences regardless of the particular manner of interactions that generate this
outcome’.96 The same amount of consumer harm results regardless of whether the
collusion is express or tacit. He also dismisses the commonly invoked argument
that it is difficult to craft an appropriate remedy for tacit collusion because courts
would need to either set prices for firms or enjoin them to deviate from the profit-
maximising price. He retorts that what deters firms from infringing the law is not
injunctive relief but the expectation of punishment in the form of fines and dam-
ages awards.97 It is important, however, to point out that Kaplow himself stops short
of advocating the outright prohibition of tacit collusion, arguing that ‘the question
is an empirical one in which the prevalence of social harm under the various stan-
dards must be compared as well as their costs of administration’.98
It was said earlier that both Predictable Agent and Digital Eye are premised on
tacit collusion. There is, however, an important qualitative difference between
them. Under Predictable Agent, firms adopt a similar algorithm with the awareness
and perhaps the expectation that competitors will follow suit. And if the firms are
aware that the algorithms may improve market conditions to facilitate tacit collu-
sion, one can argue that the intent of the firms is not the purely innocuous intent
to maximise profit. This is different from the archetypal tacit collusion that Turner,
Posner, and Kaplow have in mind, where firms may truly only intend to maximise
profit. They are independently adopting a course of conduct that they know com-
petitors are likely to follow and that may facilitate tacit collusion. A distinction is

93
Ibid.
94
Clamp-All Corp v Cast Iron Soil Pipe Institute, 851 F.2d 478, 484 (1st Cir. 1988).
95
Kaplow (n 55) 449–450.
96
Louis Kaplow, ‘On the Meaning of Horizontal Agreements in Competition Law’ (2011) 99 California
Law Review 683, 686.
97
Kaplow (n 55) 475.
98
Ibid. (n 95) 814.

https://doi.org/10.1017/9781108980197.025 Published online by Cambridge University Press

 Competition Law and AI 489

thus drawn between setting a profit-maximising price, which is part and parcel of
every business, and pursuing conduct that facilitates collusion, which a firm is by
no means compelled to do. Their awareness that their action may facilitate tacit
collusion taints their intent and bolsters the case for applying closer scrutiny of the
Predictable Agent.

2 What Is Different about Algorithmic Tacit Collusion?

Digital Eye requires us to directly confront the issue of the legality of tacit collusion.
This author is sympathetic to Posner’s and Kaplow’s arguments in support of pro-
hibiting tacit collusion. Express and tacit collusion can inflict the same consumer
harm. Moreover, the current approach to tacit collusion, with its focus on the lack
of direct communication between the parties, is overly formalistic, which is incon-
sistent with the emphasis of substance over form under competition law. Without
trying to settle the debate in the offline context, this author believes that there are
good arguments for taking a stricter stance against algorithmic tacit collusion for
a variety of reasons. First, one of the premises for treating tacit collusion leniently
in the offline context is that it should be relatively rare. As noted by Beneke and
Mackenrodt, ‘[m]ost scholars circumscribe this to a rare set of circumstances that
include highly concentrated industries, homogeneous goods, symmetric cost struc-
tures across firms, and price transparency, among others’.99
A number of commentators have observed that algorithms may allow collu-
sion in a wider variety of market structures.100 There are reasons to believe that
the immense and ever-improving technical capabilities of algorithms to monitor
rivals’ prices, to signal pricing intentions to other algorithms, and to enact frequent
price changes will turn autonomous algorithmic collusion into a more common
phenomenon. Tacit collusion may no longer be limited to highly concentrated oli-
gopolistic markets. It can also ‘be achieved with a large number of participants. The
increased transparency of the internet, the high reaction speed of various IT-systems
and algorithm-based price adjustments are thereby all decisive factors that enable
rivals to tacitly collude on markets with many market players’.101
Algorithms have been said to facilitate the following critical aspects of tacit collu-
sion: reaching terms of coordination among firms;102 rapid detection and retaliation,
hence reducing the incentive to cheat;103 communication among competitors;104

99
Beneke and Mackenrodt (n 3) 118.
100
Peter Georg Picht and Gaspare Tazio Loderer, ‘Framing Algorithms: Competition Law and (Other)
Regulatory Tools’ (2019) 42 World Competition 391, 406; Kaylynn Noethlich, ‘Artificially Intelligent
and Free to Monopolize: A New Threat to Competitive Markets around the World’ (2019) 34
American University International Law Review 923, 940; Ong (n 2) 205; Zheng and Wu (n 3) 142.
101
Giulia Sonderegger, ‘Algorithms and Collusion’ (2021) 42 European Competition Law Review 213, 216.
102
Gal (n 2) 82.
103
OECD (n 9) 21, 27.
104
Gal (n 2) 87.

https://doi.org/10.1017/9781108980197.025 Published online by Cambridge University Press

 490 Thomas Cheng

and the elimination of irrational behaviour to prevent unnecessary disruption to

the collusive scheme.105 Algorithms can communicate in myriad ways. They can
engage in repeated rounds of price signalling until a price is agreed upon.106 They
can try to decipher each other’s software codes.107 In fact, it has been said that ‘neu-
ral networks can indeed learn how to encrypt and decrypt messages, as well as how
to apply those operations selectively to ensure goals related to confidentiality’.108 It
has been argued that algorithms can also create new hurdles for collusion.109 Yet,
on balance, there are good reasons to believe that algorithms should make tacit
collusion more prevalent and more attainable. While we may have condoned tacit
collusion in the brick-and-mortar world on the ground of its rarity, the same ratio-
nale may no longer apply once technology is sufficiently advanced to popularise
algorithmic tacit collusion.
One may object to the foregoing conclusion, arguing that firms should not be
held accountable if the algorithm learns on its own to engage in tacit collusion
with no human involvement. This argument would have even greater validity in
the case of black box algorithms, whose decision-making rationale and processes
cannot be deciphered by human agents. After all, how can one be held liable
for an action that the person is in no position to prevent? An analogy could be
drawn between an algorithm and an employee. A firm can equally argue that if
an employee decides on her own to engage in collusion without being instructed
by her superiors to do so, the firm should not be held liable. A firm equally has no
means to know what an employee intends to do until after the fact if she has never
articulated her intentions. Competition law, however, has always held firms liable
for their employee’s conduct regardless of whether the employee is properly autho-
rised to act for the firm.110
By the same logic, firms should be held liable for their algorithm’s conduct.
Moreover, firms can no longer claim to be unaware of the collusive potential of algo-
rithms in light of the increasing amount of literature on the issue. They have been
forewarned. They should be expected to monitor their algorithms closely. It hardly
makes sense that consumers should be made to suffer the adverse consequences of
inadequately supervised algorithms while firms enjoy their many efficiencies and
advantages. If algorithms cannot be properly constrained from harming consum-
ers, it is apt to question whether firms should be allowed to use them.111 While an

105
Noethlich (n 99) 941.
106
Ariel Ezrachi and Maurice E Stucke, ‘Sustainable and Unchallenged Algorithmic Tacit Collusion’
(2020) 17 Northwestern Journal of Technology and Intellectual Property 217, 246.
107
Gal (n 2) 87.
108
Schwalbe (n 2) 596.
109
Ibid. 574; OECD (n 9) 23; Gal (n 2) 92.
110
Alison Jones, Brenda Sufrin and Niamh Dunne, EU Competition Law: Text, Cases, and Materials
(Oxford University Press 2019) 166.
111
Harrington Jr (n 53) 350.

https://doi.org/10.1017/9781108980197.025 Published online by Cambridge University Press

 Competition Law and AI 491

outright ban of such algorithms may be a step too far, firms should be held account-
able for collusion autonomously consummated by their algorithms when such a
possibility is known.

IV Conclusion
Algorithmic collusion will probably continue to be hotly debated in the competi-
tion law community for years to come. Until the controversy regarding the techni-
cal feasibility of autonomous algorithmic collusion is settled once and for all, there
will be commentators who insist that such collusion remains the stuff of science
fiction, and that there is no need for competition law to pay heed to it. Such an atti-
tude, however, is retrogressive, and competition law should take a pro-active stance
towards algorithmic collusion. If it is made clear to programmers that autonomous
algorithmic collusion will not be tolerated and that the indecipherable nature of
black-box algorithms will not be accepted as a defence, algorithm designers will
have the appropriate incentives to create and hone their algorithms to minimise the
possibility of autonomous algorithmic collusion. Given the technical complexity of
understanding an algorithm’s decision-making process, the best way to minimise
algorithmic collusion remains to tackle it at the design stage instead of trying to
go after it once suspected collusion arises. This may require ex ante regulation as
opposed to ex post enforcement under competition law.

https://doi.org/10.1017/9781108980197.025 Published online by Cambridge University Press