CYIL vol. 11 (2020)

CYIL 11 (2020) ARTIFICIAL INTELLIGENCE AND COMPETITION LAW… also be enabled to “calculate” the individual price in real time, thus making first-degree

discrimination feasible. 3. Machine Learning

Concerning the “internal” working of an algorithm, we may distinguish between two main categories. Relatively simple algorithms are “fixed”; the procedures they perform are set by their authors in advance and they do not “evolve”. Conversely, self-learning algorithms employ artificial intelligence (machine learning), and thus develop their own criteria based on their “experience”. 23 Crucially, the undertakings using such algorithms are not aware of the parameters the algorithms use, they can only observe the outputs; this is typical, especially for the most advanced form of machine learning, the “deep learning”, which emulates the layered neuron networks characteristic of a human brain. As summarised by the UK Study : a more advanced algorithm could be left to decide what data it considers is most relevant to meeting its objective (such as profit maximising). The algorithm would then act as a “black box” so that even the employees who instruct the algorithm would not know which variables it was using to set a particular price, and may not be aware of whether any increase in profit was due to attracting additional customers, charging higher prices to loyal customers, or tacit coordination. 24 The question whether the machine learning algorithm may be “understood” will be crucial for this article, and it is therefore useful to distinguish the algorithms based on their “interpretability”. On the one hand, there is a broad category of algorithms that may be described as “descriptive”; the procedures these algorithms execute are discernible from their code. 25 On the other, there are “black-box” algorithms, whose code allows to understand their goal, but not their strategy to reach it. 26 III. Algorithms and Competition Law In this article, we are not going to discuss whether pricing algorithms may facilitate anticompetitive conduct; this question has been answered by numerous authors to the affirmative. 27 We will therefore only briefly outline the situations in which the setting of price by a pricing algorithm may amount to a breach of competition law. We will instead focus on the attribution of liability for such conduct, specifically, whether the notion of an undertaking, as understood by competition law, is broad enough for these purposes. In order to do so, this chapter will first address the basic characteristics of the notion of an undertaking for the purposes of attributing liability for an anticompetitive conduct; 25 Franco-German Study , p. 11: “ A descriptive algorithm typically has at least partly predefined ways to observe the “state” of the world, often including the competitive environment, as for example competitors’ prices. It then analyses this state, possibly using more or less sophisticated statistical and analytical methods, and potentially also including some learning elements. […] Finally, it defines certain predefined rules to determine its reaction, for example by matching the lowest price ”. 26 Ibid , p. 12. 27 Seee.g.EZRACHI,A.,STUCKE,M.E.ArtificialIntelligence&Collusion:WhenComputersInhibitCompetition. University of Illinois Law Review , 2017. Available at: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2591874 (1 June 2020), p. 1798. 23 For further details, see e.g. OECD Report on Algorithms , p. 9 et seq . 24 Franco-German Study , p. 10.

243