ARTICLES Antitrust , Vol. 33, No. 1, Fall 2018. © 2018 by the American Bar Association. Reproduced with permission. All rights reserved. This information or any portion thereof may not be copied or disseminated in any form or by any means or stored in an electronic database or retrieval system without the express written consent of the American Bar Association. response to my price increase, my algorithm would then drop my price to the cost of the product, or even below the What Do We Know cost. The low price “hurts” both your and my revenue. The algorithm would keep this “low price” regime for a period of About Algorithmic time and then repeat the process of raising and then lower - ing prices if you do not raise your prices as well. After sever - al rounds of interaction, it is possible that you realize that my Tacit Collusion? algorithm appears to be sending you a signal: raise price with me or suffer financial losses. At that point, you might decide BY AI DENG to reciprocate my price increase, given your and my interest in long-term profitability. This outcome is just as likely if you also use a pricing algorithm that tries to maximize your long- term profit. Notice that during the entire interaction, there are no traditional communications between us. We do not ECENT YEARS HAVE SEEN LEGAL even need to know each other as long as all the conditions are scholars and antitrust agencies express interest met and intended learning is somehow achieved. 3 Note the in and concerns with algorithmic collusion. As “reward-punishment” element in my algorithm, a point to António Gomes, Head of the Competition which I will return. RDivision at the Organization for Economic Co- Many have argued that the threat of algorithmic tacit col - operation and Development (OECD), stated in a recent lusion is real and poses even greater challenges for anti trust interview, developing artificial intelligence (AI) and machine enforcement than human coordination and collusion. learning that enable algorithms more efficiently to achieve a Maurice E. Stucke and Ariel Ezrachi postulate that AI, which collusive outcome is “the most complex and subtle way for enables computers to make decisions and learn through expe - companies to collude, without explicitly programming algo - riences autonomously, “can expand tacit collusion beyond rithms to do so. ”1 price, beyond oligopolistic markets, and beyond easy detec - The type of algorithms that are capable of collusion (tac - tion. ”4 This sentiment is echoed by Michal S. Gal, who dis - itly or explicitly) by themselves, without human interfer - cusses “tacit collusion among algorithms, reached without ence, may sound far-fetched. Indeed, as one scholar puts it, the need for a preliminary agreement among them. ”5 Dylan “AAI [Antitrust and Artificial Intelligence] literature is the I. Ballard and Amar S. Naik also argue, “Joint conduct by closet ever our field came to science-fiction.” It has also been robots is likely to be different—harder to detect, more effec - stated that, regarding algorithms the “possibility of enhanced tive, more stable and persistent. ”6 The background note by tacit collusion . remains theoretical. ”2 the OECD Secretariat also states that even though “[i]t is still While these statements remain true, there is growing not clear how machine learning algorithms may actually experimental evidence that an algorithm can be designed to reach a collusive outcome . once it has been asserted that tacitly collude. The possibility of tacit collusion through an market conditions are prone to collusion, it is likely that algorithm is, in fact, not hard to see in some stylized cases. algorithms learning faster than humans are also able through For example, assume that you and I are the only two online high-speed trial-and-error to eventually reach a cooperative sellers of a homogeneous product and we know that our pro - equilibrium. ”7 curement costs are similar. Because our prices are posted These concerns naturally make one wonder what we online, we also know each other’s pricing. Suppose I use an should do about the possibility of algorithms reaching a col - algorithm that not only monitors your price but also sets my lusive outcome without companies even intending that result. own price accordingly. For skeptics, the first questions are probably “Is it even The way my algorithm works is as follows. First, it would possible?” and “Is there any real evidence that a machine raise and then keep my price high until you also change your could ever achieve that in actual markets? ”8 In fact, some have price. If it turns out that you do not raise your price in expressed doubt as to the plausibility of autonomous algo - rithmic collusion. For example, the Competition Bureau of Ai Deng, Ph.D., is a Principal at Bates White Economic Consulting and a Canada recently pointed out the lack of evidence of such lecturer at Johns Hopkins University. This article is partially based on his autonomous algorithmic collusion while recognizing the con - remarks on the panel, “Pricing Bots: Are R2D2 and C3PO Colluding?” at stantly evolving technology and business practices. 9 One sen - the ABA Section of Antitrust Law 2018 Spring Meeting. The author thanks ior U.S. Department of Justice (DOJ) Antitrust Division Joseph Harrington and Paul Johnson for discussions on this topic. The official recently stated that “concerns about price fixing views expressed in this article are those of the author only and do not through algorithms stem from a lack of understanding of the necessarily reflect the opinions of Bates White or its clients, or Johns Hopkins University or its affiliates. Heather Dittbrenner provided excellent technology, and that tacit collusion through such mecha - editorial assistance. nisms is not illegal without an agreement among partici - pants. ”10 88 · ANTITRUST Indeed, the existing legal principle states that an agreement each, an outcome strictly worse than the “cooperative” out - requires a “conscious commitment to a common scheme come. It is not surprising that cartel members face a similar designed to achieve an unlawful objective. 11 But in the con - type of incentive problem. They are both better off if they text of this debate, the evaluation of the plausibility of tacit cooperate (say, raise prices or reduce output). But at the same algorithmic collusion becomes an important exercise, even time, if I know that my competitors are raising prices, I have before we see concrete evidence that it has passed from the - an incentive to lower my prices to steal the business and oretical possibility to marketplace reality. Moreover, insights increase my revenue. Since cartel members are usually smart about how algorithms may or may not come to collude are enough not to write the cartel agreement into a contract, they invaluable in focusing attention on the key legal and eco - have to find ways to enforce their agreement. nomic questions, policy dilemmas, and practical real-world A critical point is that solving this incentive problem is key evidence. to the success of a cartel, whether it is humans or algorithms. In this article, I survey and draw lessons from the literature In other words, the use of an algorithm does not magically on A I12 and on the economics of algorithmic tacit collusion. remove this fundamental incentive problem that a cartel As we will see, while we do not have all the answers to all the faces. Of course, unlike the “one-shot” situation in the stan - questions that algorithmic collusion presents, a good under - dard Prisoners’ Dilemma, competitors interact with each standing of this literature is a crucial first step to better under - other repeatedly in the market. It turns out that in repeated standing the antitrust risks of algorithmic pricing and devis - interactions, there is “more hope” that firms can learn to ing better antitrust policies to mitigate those risks. 13 cooperate. In fact, repeated interaction is an important rea - son that tacit collusion emerges in the stylized example dis - Cartels’ Incentive Problem cussed earlier in the article. To better understand the problems a cartel must solve to sustain an agreement to restrict competition (e.g., raise prices The AI Literature or reduce output), it is instructive to look at the well-known Is there any evidence that computer algorithms can (tacitly) Prisoners’ Dilemma. Imagine two accomplices of a crime are collude? Empirically, we have not seen an actual case that being interrogated in separate rooms and they cannot com - involves tacitly colluding robots. The most well-known case municate. They must decide whether to confess to the crime was the Topkins case prosecuted by the DOJ in 201 5, but in and hence expose the other accomplice. Table 1 shows the that case computer algorithms were used as a tool to imple - consequences of their decisions. ment a cartel agreement among humans. 14 Interestingly, there has also been some theoretical and Table 1: A Prisoner’s Dilemma: Understanding the Incentive experimental evidence that certain algorithms could lead to Problem of a Cartel tacit collusion. One algorithm that has been found to be conducive to cooperative behavior in experimental settings is Prisoner B the so-called tit-for-tat algorithm (TFT). 15 This strategy starts Not confess Confess with cooperation, but then each party will just copy exactly (Cooperate) what the opponent did in the previous period in repeated Not confess (Cooperate) (–1, –1) (–3, 0) interaction. Intuitively, if two opponents start by cooperat - Prisoner A Confess (0, –3) (–2, –2) ing, then the very definition of the TFT algorithm dictates their continued cooperation. But will competitors have an incentive to deviate from cooperation? The answer is that The two rows and two columns in Table 1 represent the they might not.