Faculty of law blogs / UNIVERSITY OF OXFORD

Harmful Signals: Cartel Prohibition and Oligopoly Theory in the Age of Machine Learning


Stefan Thomas


Time to read

4 Minutes

In my paper ‘Cartel Prohibition and Oligopoly Theory in the Age of Machine Learning’ I deal with the application of the cartel prohibition in the light of alleged legal gaps resulting from the surge of algorithmic pricing. I focus on Article 101 TFEU, yet I retain a close tie to the jurisprudence and scholarship on Section 1 of the US Sherman Act. Instead of the currently established set of criteria for distinguishing between illicit collusion and legitimate oligopoly conduct, I suggest an effects-based approach in order to fill possible regulatory gaps.

The classical legal approach for distinguishing between illicit collusion and legitimate oligopoly conduct is to rely on criteria that relate to the means and the form of how rivals interact. While firms are deemed to have the right to adapt intelligently to their rivals’ conduct, they shall be prohibited from entering into a ‘form of coordination’ which ‘knowingly substitutes practical cooperation between them for the risks of competition’, as the European Court of Justice puts it. One may emphasize the need to demonstrate an anticompetitive intent among firms to establish illicit collusion. The management of oligopolistic firms might engage in public or private price announcements, for example, to achieve a new collusive equilibrium. This will then qualify as illicit collusion if the managers have thereby ‘knowingly’ substituted practical cooperation for the risks of competition.

From this line of reasoning, it is sometimes concluded that the cartel prohibition of Article 101 TFEU or Section 1 of the Sherman Act can be unable to capture collusion if it is achieved by autonomously acting computers relying on machine learning capabilities. Firms might achieve supracompetitively inflated prices without having specific knowledge or intent. Pricing algorithms might reach a new equilibrium on their own. It is even conceivable that computers with machine learning capabilities reach collusive equilibria although their designers have not intended this.

Against this backdrop, I contend in my paper that it is not appropriate to distinguish legitimate oligopoly conduct from illicit collusion by relying on criteria, such as the ‘knowingly’ criterion or other factors that relate to the means and form or anticompetitive intent. This traditional approach fails to acknowledge that there is, in fact, no categorical difference in the way illicit collusion on the one hand and tacit collusion on the other hand work. Tacit collusion can be characterized as a non-cooperative game in terms of game theory in the same way as concerted practices (and even legally unenforceable cartel agreements) can be considered a non-cooperative game. A firm reacts to its rivals, and the rivals then react to the observed conduct of this firm intelligently.

Therefore, I argue that it is unconvincing to try to find something idiosyncratic about the means or form used for illicit concerted practices, which is purported to be absent in ‘mere’ tacit collusion cases. The law puts different labels on what is ultimately the same economic phenomenon. I therefore suggest in my paper that the relevant criterion should be whether the release of informational signals by two or more rivals creates or sustains a supracompetitive equilibrium, the consumer harm of which is not offset by concomitant consumer benefits. This allows distinguishing between legitimate and illicit collusion based on an effects-analysis guided by the consumer welfare standard. It makes other criteria, which rely on the means and form of communication and the inner sphere of natural persons, dispensable.

The idea submitted here is not equal to the approach, as suggested in some scholarly writings (eg here and here) on Section 1 of the Sherman Act, of condemning oligopolistic pricing as per se anticompetitive under the cartel prohibition, namely, to equate tacit collusion with agreements or concerted practices. The differentiated approach submitted here avoids some of the conceptual problems that come with that idea. It does not force companies to act as if they do not know what they actually know about their competition.

Rather, the present suggestion is to identify singular elements of communication, ie ‘informational signals’, which must be checked for their propensity to create a consumer harm. If and to the extent that an informational signal creates harm, a firm must refrain from releasing that signal. The relevant counterfactual for the identification of illicit coordination therefore is not a situation in which tacit collusion does not occur at all. Rather, it is the hypothetical market outcome as it would present itself if the potentially harmful informational signal, which is being analyzed, were absent. Some informational signals, such as publicly available price lists, might be indispensable for consumers to make informed choices and plan their purchases ahead. Even though these price lists might facilitate collusion, they can create benefits to consumers which offset a potential consumer harm. They should therefore not be prohibited. If price announcements are being made, however, which do not create benefits to consumers in a given case although they enable oligopolists to collude, this informational signal should be prohibited. This might be the case for non-binding announcements that are being made a long time before price changes are supposed to take place, or for a private exchange of information, if and to the extent that a sufficient consumer benefit cannot be established. If such a harmful signal can be identified, according to the approach submitted here, it will not be decisive whether natural persons or computers with machine learning capabilities have released it. Neither will it be relevant whether natural persons intended to influence prices or whether the designer of an algorithm had any anticompetitive intent.  

The advantage of the approach submitted here is that it renders the application of the cartel prohibition more robust in cases where direct human involvement in the coordination process is limited, absent, or hard to detect. This can help to close the regulatory lacuna that is alleged to exist with respect to algorithmic pricing under the cartel prohibition. Also, it reconciles the law with the consumer welfare approach in that it allows for interventions if consumers suffer irrespective of further criteria. Negligence or intent can be relevant for the imposition of a fine or an antitrust damages claim. Yet the lack of an anticompetitive intent, under this approach, will not hinder a prohibition decision. At the same time, the release of information which increases market transparency will be unobjectionable if and to the extent it creates consumer benefits.

Stefan Thomas is a Professor of Civil, Commercial and Competition Law at the University of Tübingen, Germany.


With the support of