Algorithms, Machine Learning, and Collusion

This paper discusses the question whether self-learning price-setting algorithms are able to coordinate their pricing behaviour to achieve a collusive outcome that maximizes the joint profits of the firms using these algorithms. While the legal literature generally assumes that algorithmic collusion is indeed possible and in fact very easy, the computer science literature on cooperation between algorithms as well as the economics literature on collusion in experimental oligopolies indicate that a coordinated and in particular tacitly collusive behaviour is in general rather difficult to achieve. Many studies have shown that some form of communication is of vital importance for collusion if there are more than two firms in a market. Communication between algorithms is also a topic in artificial intelligence research and some recent contributions indicate that algorithms may learn to communicate, albeit in a rather limited way. This leads to the conclusion that algorithmic collusion is currently much more difficult to achieve than often assumed in the legal literature and is therefore currently not a particularly important competitive concern. In addition, there are also several legal problems associated with algorithmic collusion, for example, questions of liability, of auditing and monitoring algorithms as well as enforcement. The limited resources of competition authorities should rather be devoted to more pressing problems as, for example, the abuse of dominant positions by large online-platforms