Artificial intelligence and collusive behaviors: How algorithms may affect collusion in the market?

Nội dung text: Artificial intelligence and collusive behaviors: How algorithms may affect collusion in the market?

1. HỘI THẢO KHOA HỌC QUỐC GIA ĐỊNH HÌNH LẠI HỆ THỐNG TÀI CHÍNH TOÀN CẦU VÀ CHIẾN LƯỢC CỦA VIỆT NAM 6 9. 1Mai Nguyen Dung* Abstract With Artificial Intelligence (AI) development, some may argue that AI’s effects increase competition since firms enter new areas outside their previous core business. However, some also warned that it could lead to super firms’ creation that could negatively affect the whole economy. In that context, competition law, as an instrument to maintain market competition, has to alter and develop to govern the consequences that AI can cause. At the end of the day, how can it impact competition in the market? This paper would draw a broad picture of the interaction between AI and collusive behaviors by doing qualitative research, as well as analyzing and synthesizing theories method. The result is, while AI helps tacit collusion in strengthening some factors that promote collusions, such as market transparency and interaction, it is still ambiguous to understand the impact of AI on the number of firms and barriers to entry. Besides, it is noted that AI may cause some adverse effects in reaching a collusive agreement, especially by its innovation and cost asymmetry consequences. We may think about some competition law approaches to these factors from these analyses, such as some ex-ante competition tools and rethinking about some related concepts. Keywords: Artificial intelligence, collusion, tacit collusion, competition law. 1. Introduction 1.1. Background AI plays an essential role in our lives, not only in the digital perspective but also in economic, education, transportation, or even medication. In the industrial sector, AI can affect the market in many ways. Brynjolfsson, Rock, and Syverson argued that AI is a general-purpose technology, that is, a technology that can affect an entire economy and * University of Economics Ho Chi Minh city | Email: dungnm@ueh.edu.vn 1037
2. HỘI THẢO KHOA HỌC QUỐC GIA ĐỊNH HÌNH LẠI HỆ THỐNG TÀI CHÍNH TOÀN CẦU VÀ CHIẾN LƯỢC CỦA VIỆT NAM open up new opportunities (Agrawal et al., 2019, p. 4). With the use of AI, in the long run, society will be wealthier. One could argue that AI’s effects are increasing competition since firms entering new areas outside their previous core business (Stevenson, 2019, p. 195). However, some also warned that it could lead to super firms' creation that could have detrimental effects on the broader economy (Szczepański, 2019, p. 2). In that context, competition law, as an instrument to maintain market competition, has to alter and develop to govern the consequences that AI can cause. At the end of the day, how can it impact competition in the market? We can imagine the effects of AI and antitrust in both negative and positive ways. On the one hand, AI could make independent decisions on its behalf by automatically investigating any market when it finds that the market has some competitive concerns. However, on the other hand, AI itself also raises some anticompetitive concerns, leading to collusion or abusing market’s dominant position. Among all anticompetitive behaviors, collusive agreements are something that most jurisdictions are trying to fight against. It occurs when rival firms agree to work together; for example, fix selling prices for one product in a relevant market. It is referred to as one of the most harmful anticompetitive practices known to competition law and is a prime target of competition authorities (Whish & Bailey, 2018, p. 520). There are many negative impacts of collusion: high prices for consumers, high entry barriers, low incentives for innovation, and increased productivity. This paper will link AI and collusion, and as a result, a new term as ‘collusive algorithm’ would be formed. In the first part, this paper will introduce some fundamental concepts of AI and its underlying technologies: algorithms, machine learning (ML), and deep learning (DL). The second part will discuss the idea of collusion in general, tacit collusion in particular, and the economic part of them: why collusion exists and which factors determine collusion stability. The third part will deal with the impact of algorithms and AI on the likelihood of collusion from both the supply side and demand side. The last part will give some concluding remarks on this paper’s topic. 1.2. Purposes Through this paper, the author’s primary purpose is to understand better novel algorithmic collusive agreements that use AI. It consists of his research field and interest in competition law and computational law. In reaching the goal, this paper is intended to provide an overview of AI and its underlying technology, thereby learning how it can affect the competition law in general and collusion in particular from a view of law and economics. In doing this research, the paper would base on some articles about collusion 1038
3. HỘI THẢO KHOA HỌC QUỐC GIA ĐỊNH HÌNH LẠI HỆ THỐNG TÀI CHÍNH TOÀN CẦU VÀ CHIẾN LƯỢC CỦA VIỆT NAM by Edward Chamberlin, Nicolas Petit, Ariel Ezrachi, Maurice E. Stucke, or some institutions such as OCED, OECD Committee, and Autorité de la concurrence & Bundeskartellamt (the competition authorities in France and Germany). 1.3. Methodologies This paper focuses on these main methods: ▪ Qualitative method: the qualitative method is the primary method in this article. ▪ Analyzing and synthesizing theory method: from all academic sources such as books, reports, legal journals, and other legal and economic materials. This paper would use this method to explain some technological and legal concepts (for instance, deep learning, machine learning, collusion, and so on). ▪ Normative approach: this method (primarily general normative research) will be used in Part 3: The economics of collusion and tacit collusion and Part 4: The economic impacts of AI on collusion. ▪ Case study: some cases would be mentioned, i.e., the case of Asus (Case AT.40465), Trod and GBE, EGEL and Dyball, or some pieces of evidence in Germany, the Netherlands. 2. The general concept of AI All of the fundamentals of AI, ML, and DL are based on algorithms; hence, it is essential to understand this term’s concept. ‘Algorithm’ is not a new word, as it has a long history and can be traced back to the 9th century by a Persian Muslim mathematician. Besides, it seems to be that the term ‘algorithm’ does not have a clear consensus on its definition. In the broadest sense of interpretation and the past, the word ‘algorithms’ referred to all definite procedures for solving problems or performing tasks (McFadden, 2017). However, since this paper only focuses on algorithms, which are the basis for AI, it just targets computer science. Accordingly, it is related to “any well-defined computational procedure that takes some value, or set of values, as input and produces some value, or set of values, as output”, for instance, “a sequence of computational steps that transform the input into the output” (Cormen et al., 2009, p. 45). There are many ways to classify the types of algorithms. Some may divide them into supervised learning, unsupervised learning, semi-supervised learning, reinforcement learning, transduction, and learning to learn. In the technological aspect, AI systems can be referred to as “self-training structures of ML predictors that automate and accelerate human tasks” (Taddy, 2019, p. 62). In other words, it is the field that studies the synthesis and analysis of computational agents that act intelligently. According to this definition, two terms need to be interpreted carefully, which are ‘intelligently’ and ‘computational’. ‘Intelligently’ can be seen as (i) 1039
4. HỘI THẢO KHOA HỌC QUỐC GIA ĐỊNH HÌNH LẠI HỆ THỐNG TÀI CHÍNH TOÀN CẦU VÀ CHIẾN LƯỢC CỦA VIỆT NAM an appropriate behavior or conduct in specific circumstances to fulfill its goals; (ii) flexible in changing environments and changing goals; (iii) learns from experience; and (iv) makes appropriate choices given its perceptual and computational limitations. On the other side, a ‘computational’ agent is an agent who makes decisions that can be explained in terms of computation. That is, the decision can be broken down into primitive operations that can be implemented in a physical device. In AI, it is carried out in ‘hardware’ (Mackworth & Poole, 2010, p. 3). Many people are confused with ML and AI, and some also think they are the same technology; in fact, they are pretty different. ML can do incredible things, but it is basically limited to predict a future that looks mostly like the past. In contrast, an AI system can solve complex problems that humans have previously resolved. In order to do this, it breaks these problems into a bunch of simple prediction tasks, each of which can be attacked by a ‘dumb’ algorithm. AI uses instances of ML as components of the more extensive system. These ML instances need to be organized within a structure defined by domain knowledge, and they need to be fed data that helps them complete their allocated prediction tasks. So, in essence, an AI system is ML-driven (Taddy, 2019, p. 63). ML is implanted in every aspect of AI, and ML algorithms are the building blocks of AI within a broader context. As can be seen, ML is simply a way of achieving AI (McClelland, 2017). On the other hand, DL is a subset of ML, which teaches computers to do what comes naturally to humans: learn by example. DL increases the original ML algorithms’ accuracy, using multiple layers in a neural network to analyze data in many different abstractions. It was trained by using a massive labeled data pool without the need for manual feature extraction. It has become popular in recent decades thanks to the rapid increase in the amount of digital data that humans have created, in addition to the increased processing power of computers while the cost has dropped. DL has driven progress in various fields such as object perception and machine translation, which used to be very challenged by AI researchers (Brownlee, 2019). As analyzed above, these technologies are become more ‘intelligent’, thanks to the development of powerful processors, Big Data, Big Analysis, and digitalization. This knowledge is necessary to understand the upcoming analysis of collusion and collusion using algorithms in this paper’s forthcoming parts. We can see that AI is a new general- purpose technology, that is, a technology that can affect an entire economy and open up new opportunities (Agrawal et al., 2019, p. 4). It is ‘the invention of a method of invention’, and plays a crucial role in the market. Therefore, to competition agencies, to conclude whether we should accept AI unconditionally or cautiously or in which way we should regulate AI, it is obvious to carefully analyze this technology in the light of 1040
5. HỘI THẢO KHOA HỌC QUỐC GIA ĐỊNH HÌNH LẠI HỆ THỐNG TÀI CHÍNH TOÀN CẦU VÀ CHIẾN LƯỢC CỦA VIỆT NAM competitive concerns in the market and protecting the consumers against the dominant market players. 3. The economics of collusion and tacit collusion 3.1. The economics of collusion In economics, collusion can be referred to as “a non-competitive, secret, and sometimes illegal agreement between rivals which attempts to disrupt the market’s equilibrium” (Young, 2019). Usually, companies will operate independently and compete with each other to attract customers. However, under some circumstances, companies would collude to maximize profits and gain unfair conditions. Markets are defined as where, at a minimum, undertakings in an industry that produce goods or provide services that are reasonably good substitutes for one another. In the output markets, when pursuing the goal of maximizing profits, the behavior of all firms is the same: they must choose the output so that the marginal cost and the marginal revenue of the final output unit are equal. Two extremes can take place in a market: perfect competition and monopoly. In the perfect competition market, one or several businesses cannot affect the price of products or services since companies are constantly competing with each other. At the opposite extreme, there is only one company in the market, which is limited only by the demand curve in determining prices. However, in practice, all actions lie in the middle. Most industries exist many firms, often of drastically varying sizes, with some or all having market power, which is the ability to raise prices higher than the price of competitors and still in buoyant demand. Such situations are known as an oligopoly or imperfect competition. In this case, the behaviors of a few firms would affect others, and these firms recognize that their actions will be taken into account by other firms. For example, when one seller reduces its prices, other sellers will act in the same way. As such, firms in an oligopoly recognize their interdependence in the market (Marshall & Marx, 2012, p. 4). Also, since the monopoly outcome is not a noncooperative equilibrium, if firms can cooperate and agree to limit their production or prices, the profits of all can be increased. That is the case of collusion. The sustainability of collusion depends on many essential factors. Regarding the scope of this paper, information exchange and punishment will be discussed as they are relevant to the application of AI. First, the transfer of information is a vital feature of collusion, directly supporting the enforcement of any collusive agreements and structures (Church & Ware, 2000, p. 308). Detection cheaters would be more natural if firms can observe the price and output of their rivals. There is theoretical evidence support that information exchange would promote more significant collusion and is detrimental to 1041
6. HỘI THẢO KHOA HỌC QUỐC GIA ĐỊNH HÌNH LẠI HỆ THỐNG TÀI CHÍNH TOÀN CẦU VÀ CHIẾN LƯỢC CỦA VIỆT NAM society. With communication, the probability that a price cut will trigger a punishment is significant (Awaya & Krishna, 2016, p. 307). Second, since collusion is a definite game, firms compete against each other continuously. Any firms in the collusion would recognize and act on their mutual independence and maximize joint profit. However, firms also have incentives to cheat because they would ideally earn the total profits in the market instead of sharing them with other players. As a result, the burden of enforcement falls on the firms, and it depends on the detection and punishment of deviators: “The stronger, the swifter, or the more certain punishment, the more likely a collusive agreement is sustainable” (Church & Ware, 2000, p. 328). 3.2. The economics of tacit collusion In the oligopolist environment (a phenomenon that exists somewhere between monopoly and perfect competition), all firms recognize their mutual interdependence and the advantages of coordination. Hence they will anticipate its price that would be in line with their competitors. In the end, an undertaking will use a course of behaviors in the knowledge that it is mutually beneficial if all players in the market act as an undertaking. These firms do not need to enter into any formal arrangement, i.e., a contract or an agreement. In terms of effects, it is likely as if there were collusion in the market, but there is no actual collusion in terms of forms. It is the case of ‘tacit collusion’, which is a market failure. It is contrary to the monopolists or atomistic firms; that is when firms choose their best response without acknowledging other players’ strategies. When saying about tacit collusion and oligopoly, we often discuss the oligopoly problem (that is, the tacit collusion occurs in an oligopoly). Chamberlin stated that one would realize that his move has a considerable effect upon his rivals in an oligopolist. The prices all move together, and it will be the monopoly price (Chamberlin, 1929, p. 89). In light of the basic knowledge of microeconomics about game theory, tacit collusion can be understood as a classic prisoner’s dilemma: two players who cannot communicate with each other will choose for their self-interests and betray other players. However, in a repeated game in which oligopolists can indeed communicate to each other, and there is an increase in the discount factor, the long-term profits of collusion will exceed the short-term profits earned by betraying each other. Hence, in this kind of game, any oligopolist that decides to deviate faces the risk of retaliation in subsequent periods and may thus be driven out of such industry. It is a clear example of a grim trigger. In order to reach that equilibrium, four conditions should be satisfied: (i) oligopolists must share a common understanding of the price at which collusion should unveil; (ii) there must be a credible threat against deviators, (iii) there is a monitor mechanism to detect any cheaters, and (iv) the sustainability of tacitly collusive prices is conditioned on the 1042
7. HỘI THẢO KHOA HỌC QUỐC GIA ĐỊNH HÌNH LẠI HỆ THỐNG TÀI CHÍNH TOÀN CẦU VÀ CHIẾN LƯỢC CỦA VIỆT NAM oligopolists’ ability to discourage production by external firms (Petit, 2012, p. 7). It is noted that boundaries between tacit collusion and parallel behavior in oligopolistic markets are not so clear (Polemis & Oikonomou, 2018, p. 26). As the same with explicit collusion, tacit can cause adverse effects on consumer welfare since it can have a negative impact on allocative efficiency and productive efficiency. 4. The economic impacts of AI on collusion Many elements affect the stability and sustainability of collusion. For instance, initiating an agreement will be more challenging if firms have different costs. If firms have high average costs, they would prefer their products to be sold at a high price, lower aggregate output, and vice versa. In this case, the two firms in a duopoly cannot divide output equally since the firm with lower costs has more incentives to cheat than the high-cost firm (Church & Ware, 2000, p. 319). We can imagine AI would impact and change the structure of the industrial organization in many relevant characteristics. That leads to some market that is more prone than others to have long-lasting collusion. This part is aimed to examine these factors. First, the relative number and size of participating companies are among the most critical characteristics that impact collusion risk (Ivaldi et al., 2003, p. 12). Suppose a market has several participants in a collusive agreement more than the number of outside the agreement, and such participants have a significant market share. In that case, there will be an enormous potential for collusion to create market power. It will be more difficult in tacit collusion when firms mainly observe their rivals to adjust behaviors. A large number of firms could obviously lead to more complicated to identify a focal point for coordination, and each player in collusion may receive a smaller share of profit. Simultaneously, the incentives to cheat will increase. In this way, AI can reduce the coordination costs by “facilitating the processing of information needed to implement coordination, thereby increasing the stability of collusion” (Autorité de la concurrence & Bundeskartellamt, 2019, p. 18). However, the OECD Paper added that though the market concentration plays an important role, it is not a must condition for algorithmic collusion to occur (OECD, 2017, p. 21). Second, barriers to entry. The collusion will be challenging to sustain if there are low entry barriers to a particular market. Players in the market would choose triggering strategies, hence reduce the profit of collusion. However, the OECD claimed that the effects of AI on entry barriers are unclear since many industries are different and characterized by natural entry barriers. Also, with AI's prediction ability, incumbent firms can use killer-acquisition to prevent entries (OECD, 2017, p. 21). There would be more 1043
8. HỘI THẢO KHOA HỌC QUỐC GIA ĐỊNH HÌNH LẠI HỆ THỐNG TÀI CHÍNH TOÀN CẦU VÀ CHIẾN LƯỢC CỦA VIỆT NAM effective pricing strategies due to algorithms, but on the other hand, the vast data that AI needs can lead to a rise in entry barriers. Third, frequency of interactions. It is one of the essential factors that AI may cause to the market. The more frequently firms interact with each other, the more comfortable they are to collude since they can respond more quickly with any cheated behaviors. Hence, a cheating firm cannot take advantage of a long time, and their incentives to cheat will decrease (Ivaldi et al., 2003, pp. 20–21). With the faster processing and predicting ability, and AI can increase interaction frequency and increase collusion stability. Hence, the price will be updated in real-time, thus allowing an immediate punishment to deviators. For instance, ML algorithms could make dynamic pricing faster and less costly. Associated with the characteristic of market transparency, the sooner the competitors acknowledge the deviated conduct, the faster they will punish and retaliate. In vertical agreements, AI also can be used. Specifically, that is the case of retail price maintenance (RPM). There are three applications of AI in such agreements: (i) detecting any deviations from the minimum sales price, (ii) increasing price transparency through price monitoring software, thus limiting the incentive to deviate, and (iii) spreading high prices to other retailers who do not engage in the RPM scheme. In the case of Asus, the European Commission (EC) held that this manufacturer had infringed the EU Competition law by imposing fixed or minimum resale prices. Asus used software to observe and monitor online retailers. If retailers did not follow RPM, they would be sanctioned or blocked of supply. Notably, pricing algorithms were adapted by these retailers, which “may multiply the impact of price movements [ ], avoid price erosion across, potentially, its entire (online) retail network” (European Commission, 2018, p. 14). As another example, in some markets, both online and offline, seem to a rapid increase in the use of pricing algorithms. In 2013, Amazon conducted more than 2.5 million price changes every day, which is 50 times more than Walmart or Best buy did in a month (Profitero, 2013). In the offline environment, the use of AI leads to more data can be gathered, but it is challenging to observe the price in these markets due to the lack of transparency. Nevertheless, some markets witness a rise in algorithmic pricing; for instance, the retail petrol providers already used algorithms to collude tacitly and potentially raised prices for consumers (CMA, 2018, p. 19; Schechner, 2017). Fourth, market transparency. When the market lacks transparency, players cannot quickly identify deviators since there is not enough information. In other words, collusion is more difficult to sustain when prices are not observable. Besides, the delay in obtaining data on costs and output also plays a vital role because if rivals receive the information 1044
9. HỘI THẢO KHOA HỌC QUỐC GIA ĐỊNH HÌNH LẠI HỆ THỐNG TÀI CHÍNH TOÀN CẦU VÀ CHIẾN LƯỢC CỦA VIỆT NAM promptly, they will have time to trace and punish deviants (OECD Competition Committee, 2010, p. 148). Market transparency is seen as one benefit of AI on the market. In this way, companies can collect data easily and quickly to foster collusion. Therefore, real-time data can be automatically analyzed and converted into action. (Autorité de la concurrence & Bundeskartellamt, 2019, p. 18). Moreover, when more and more firms join the ‘algorithmic-driven’ game, the market would be more transparent, result in an increase in the chance of collusion and reducing strategic uncertainty. In a paper, Ezrachi and Stucke used a gas station as an instance, as it can silently adjust the price to be compatible with other competitors in the same area. For example, in Germany, an oligopoly of five firms, which has a significant market share in the off- motorway petrol station market. In enhancing competition in this market, the German authorities required these stations to report their prices to the government’s transparency unit any price changes in real-time and then inform the consumers, aiming for them to find the cheapest station nearby. However, as a result, this policy has the opposite effect while increasing the petrol price, rather than lowering it. Therefore, instead of several hours for a price change, which algorithms, it only takes a few seconds to monitor and punish any deviators. The competitors will have less incentive to discount (Ezrachi & Stucke, 2020). Also, tacit collusion is not illegal and will not trigger any intervention from the authorities. They presented some experiments and concluded that firms have incentives to use algorithms to facilitate tacit collusion in the market. Moreover, in the Hub-and- spoke scenario, which can be defined as using a single pricing algorithm to determine the market price, which is charged by numerous users, the data points and values, the likelihood, and their dynamic pricing strategies for alignment increases. An instance is the petrol market in Rotterdam (the Netherlands), where several petrol stations used the same provider for an advanced analytic tool to determine petrol prices (Schechner, 2017). Fifth, asymmetries between companies may lead to some negative impacts on the stability and likelihood of collusion. The different costs can affect the incentives of the firms to cheat. Moreover, the asymmetry in discount factors plays an essential role in determining whether punishment in the future can be efficient in preventing cheating today. If the discount factor is high, a firm might think about colluding since their earnings in the future are more valuable than today, and vice versa. Besides, firms or individuals may differ in their willingness to engage in illegal activities (Church & Ware, 2000, p. 325). Firms are also different in capacity constraints, affecting the amount that it gains from undercutting its rival and retaliatory power (OECD, 2017, p. 21). 1045
10. HỘI THẢO KHOA HỌC QUỐC GIA ĐỊNH HÌNH LẠI HỆ THỐNG TÀI CHÍNH TOÀN CẦU VÀ CHIẾN LƯỢC CỦA VIỆT NAM In 2016, the UK Competition & Markets Authority (CMA) fined Trod and GBE for their participation in an anti-competitive agreement and/or concerted practice by having a concurrence of wills and coordination of conduct between parties “not to undercut each other in certain specified circumstances on prices for posters and frames sold by both Parties on Amazon UK” (CMA, 2016). They used automated, repricing, and configuring software to implement their arrangement. In the same year of 2016, the UK regulator of the energy sector (Ofgem) found EGEL and Dyball agreed not to target each other’s customers by sharing customer lists. To effectively corporate, they had designed, implemented, and maintained common software and algorithms that allowed the acquisition of certain customers to be blocked and customer lists to be shared (Ofgem, 2019). In this way, we can see the impact of algorithms on the asymmetries between companies, frequency of interactions, and market transparency. Sixth, product differentiation. The other factor is related to the fact that firms often try to differentiate their products. They may compete for market shares via other means, such as product quality or advertisement, which makes the tacit collusion become not sustainable, as the firm has less to fear from retaliation from its rivals. Under the use of AI, collusion may be challenging to sustain from asymmetries in preferences and product heterogeneity. This technology allows firms to differentiate their products, leading to difficulties in finding a ‘focal-point’ (Autorité de la concurrence & Bundeskartellamt, 2019, p. 23). However, with AI’s ability, it helps companies analyze and react to other players more efficiently and sophisticatedly, which can increase collusion stability. Seventh, innovation. Firms are more challenging in reaching and sustaining an agreement in the market that there is product innovation rather than the market in which products are stable (Church & Ware, 2000, p. 323). The firm will choose to betray to earn short-term profit rather than cooperate in obtaining long-term one. In the digitalization era, AI can reduce the stability of collusion by promoting innovation and developing untraditional business models. We can see that many companies have spent billions of dollars on Research and Development (R&D) activities in recent years, and a significant amount of money has been used to design the best algorithms. In the enforcing and implementation period, AI helps firms overcome difficulties in the initiation phase since it can use its abilities to analyze other players' behaviors through potentially superior capabilities to deal with such industry. With ML and DL, companies can compare and evaluate trade-offs when deciding whether or not they should collude and how they collude tacitly. 1046
11. HỘI THẢO KHOA HỌC QUỐC GIA ĐỊNH HÌNH LẠI HỆ THỐNG TÀI CHÍNH TOÀN CẦU VÀ CHIẾN LƯỢC CỦA VIỆT NAM 5. Conclusion Dealing with the research question: “How can AI impact collusion in the market?”, this paper began with explaining and evaluating AI and its impact on the market. It is necessary to understand the fundamentals of this technology. First, AI is a broad term but primarily refers to self-learning and automation, which does not involve human beings. In most situations, the terms algorithm and AI are the same since AI is generally an advanced development algorithm. ML is a subset of AI that emphasized the ability to predict the future based on past data. DL is another essential subset, but it is more advanced since it is powered by multiple layers of neural networks that are able to learn by examples; hence, it increases the accuracy of the original ML algorithms. We can see that AI brings both benefits and concerns to the market. For instance, the increase in transparency and reduction of transaction costs are two typical examples. On the other hand, it can tacitly collude with competitors, thus harming the consumers and reducing their welfare. Then, it is essential to explain the economic concepts of collusion and tacit collusion. In the implicit case, instead of communicating with others, firms will observe others’ behavior, adjust and adapt their conduct. A firm may choose to co-operate with others in a repeated game rather than deviate and be sanctioned. It is then vital to connect between technology and economics. The stability and long-lasting of any collusive behaviors depend on many factors. While the impacts of AI on the number of firms and entry barriers are ambiguous, it is essential to note that market transparency will increase with AI use. In this case, competitors can easily track others’ pricing and strategies. They can quickly identify deviators, foster collusion, as well as reduce strategic uncertainty. Moreover, AI increases the frequency of interactions, which helps firms behave collusively by rapidly identifying any deviators, increasing the collusion’s sustainability. However, in some aspects, AI impacts collusion negatively by raising the innovation and cost asymmetries. It should be noted that the likelihood that tacit collusion using AI systems will occur in the future is very high, thanks to the advanced development in several technologies, such as supercomputers, blockchain, IoT, and Big Data. We can expect the future in which supercomputers can execute ML or DL quickly with no delay. Moreover, with supercomputing, the God view scenarios seem to become real in the antitrust perspective, as these computers of firms will reach the same outcome intelligently and simultaneously. They would be ‘smarter’ to collude and find a focal point without human intervention tacitly. 1047
12. HỘI THẢO KHOA HỌC QUỐC GIA ĐỊNH HÌNH LẠI HỆ THỐNG TÀI CHÍNH TOÀN CẦU VÀ CHIẾN LƯỢC CỦA VIỆT NAM The rising of blockchain, together with AI, also created a threat to the market, as it is able to prevent, monitor, and correct deviant behaviors efficiently. Public blockchains let firms access a massive data pool and enhance their ability to observe others’ practices. Besides, AI and blockchain can solve the conflicts between members of an agreement, as they avoid human biases by automating governance. In terms of correcting deviant behaviors, blockchain and AI may coordinate and ensure a fair balance through a redistribution of extra profits (Schrepel, 2019, pp. 144–149). IoT and Big Data also contribute to potential collusive practices as it provides more tools to trace and track other competitors’ behaviors, thus reducing the incentive to betray. Firms now acknowledge that their conduct would be observed immediately. Moreover, it enables machine-to-machine communication, leading to the Digital Eye scenario. Also, the data that IoT devices collect would be a considerable resource to train the AI system and algorithms. Big Data also plays a vital role. As illustrated, AI needs data: the more data it learns, the more ability, intelligence, and accuracy it performs. With more information, the transparency of the market will be increased. The increasing popularity of algorithmic collusion and its dangerous consequences have led to a necessity in addressing this emerging phenomenon. References Agrawal, A., Gans, J., & Goldfarb, A. (2019). Introduction. In A. Agrawal, J. Gans, & A. Goldfarb (Eds.), The Economics of Artificial Intelligence: An Agenda (p. 4). The University of Chicago Press. Autorité de la concurrence & Bundeskartellamt. (2019, November). Algorithms and Competition. Bundeskartellamt. lgorithms_and_Competition_Working-Paper.pdf?__blob=publicationFile&v=5 Awaya, Y., & Krishna, V. (2016). On Communication and Collusion. American Economic Review, 106(2), 285–315. Brownlee, J. (2019, December 20). What is Deep Learning? Machine Learning Mastery. Chamberlin, E. H. (1929). Duopoly: Value Where Sellers Are Few. The Quarterly Journal of Economics, 44(1), 63–100. Church, J. R., & Ware, R. (2000). Industrial organization: A strategic approach (International ed). Irwin McGraw-Hill. CMA. (2016, August 12). Decision of the Competition and Markets Authority—Online sales of posters and frames—Case 50223. final-non-confidential-infringement-decision.pdf 1048
13. HỘI THẢO KHOA HỌC QUỐC GIA ĐỊNH HÌNH LẠI HỆ THỐNG TÀI CHÍNH TOÀN CẦU VÀ CHIẾN LƯỢC CỦA VIỆT NAM CMA. (2018, October 8). Pricing algorithms—Economic working paper on the use of algorithms to facilitate collusion and personalised pricing. a/file/746353/Algorithms_econ_report.pdf Cormen, T. H., Leiserson, C. E., Rivest, R. L., & Stein, C. (2009). Introduction to Algorithms, Third Edition (3rd ed.). The MIT Press. European Commission. (2018, July 24). Case AT.40465—ASUS - Antitrust procedure Council Regulation (EC) 1/2003. Ezrachi, A., & Stucke, M. E. (2020). Sustainable and unchallenged algorithmic tacit collusion. NW. J. Tech. & Intell. Prop., 17, 217. Ivaldi, M., Jullien, B., Rey, P., Seabright, P., & Tirole, J. (2003). The Economics of Tacit Collusion—Final Report for DG Competition, European Commission [Final Report for DG Competition, European Commission]. _en.pdf Mackworth, A. K., & Poole, D. L. (2010). Artificial Intelligence and Agents. In Artificial Intelligence: Foundations of Computational Agents (pp. 3–42). Cambridge University Press; Cambridge Core. Marshall, R. C., & Marx, L. M. (2012). The economics of collusion: Cartels and bidding rings. MIT Press 01503 McClelland, C. (2017, December 4). The Difference Between Artificial Intelligence, Machine Learning, and Deep Learning. Medium. between-artificial-intelligence-machine-learning-and-deep-learning-3aa67bff5991 McFadden, C. (2017, January 30). The Origin of Algorithms We Use Every Single Day. Interesting Engineering. OECD. (2017). Algorithms and Collusion: Competition Policy in the Digital Age. www.oecd.org/competition/algorithms-collusion-competition-policy-in-the-digital-age.htm OECD Competition Committee. (2010). Information Exchanges between Competitors under Competition Law. Ofgem. (2019, July 26). Infringement by Economy Energy, E (Gas and Electricity) and Dyball Associates of Chapter I of the Competition Act 1998 with respect to an anti-competitive agreement. Ofgem. _e_gas_and_electricity__dyball_associates_infringement_of_chapter_i_ca98_doorstep_sale s_redacted_decision_document_26_july_2019.pdf Petit, N. (2012). The Oligopoly Problem in EU Competition Law. SSRN Electronic Journal. 1049
14. HỘI THẢO KHOA HỌC QUỐC GIA ĐỊNH HÌNH LẠI HỆ THỐNG TÀI CHÍNH TOÀN CẦU VÀ CHIẾN LƯỢC CỦA VIỆT NAM Polemis, M. L., & Oikonomou, A. (2018). Tacit collusion or parallel behaviour in oligopolistic markets? The two faces of Janus. European Competition Journal, 14(1), 1–37. Profitero. (2013, December). Profitero Price Intelligence: Amazon makes more than 2.5 million daily price changes. Profitero. amazon-com-makes-more-than-2-5-million-price-changes-every-day/ Schechner, S. (2017, May 8). Why Do Gas Station Prices Constantly Change? Blame the Algorithm. The Wall Street Journal. prices-constantly-change-blame-the-algorithm-1494262674 Schrepel, T. (2019). Collusion by Blockchain and Smart Contracts. Harv. J.L. & Tech., 33, 117. Stevenson, B. (2019). Artificial Intelligence, Income, Employment and Meaning. In A. Agrawal, J. Gans, & A. Goldfarb (Eds.), The Economics of Artificial Intelligence: An Agenda (p. 194). The University of Chicago Press. Szczepański, M. (2019, July). Economic impacts of artificial intelligence (AI). Europarl. 67_EN.pdf Taddy, M. (2019). The Technological Elements of Artificial Intelligence. In A. Agrawal, J. Gans, & Avi Goldfarb (Eds.), The Economics of Artificial Intelligence: An Agenda (p. 61). The University of Chicago Press. Whish, R., & Bailey, D. (2018). Competition Law (Ninth Edition). Oxford University Press. Young, J. (2019, May 11). Collusion. Investopedia. 1050