Algorithms are usually described as finite sequences of defined, computer-implementable instructions, typically to solve a class of problems or to perform a computation (source: Wikipedia). The Study focused on whether the existing (EU and national) Competition Law framework adequately addresses potential threats for competition. The Study was published in November 2019. It is available under this link.
The study particularly focuses on algorithms used for dynamic price setting, as these are most likely to be detrimental for competition. It reiterates that Competition Law in the EU and in most of the EU Member States prohibits only agreements and concerted practices in addition to abuse of a dominant position. Therefore, in the absence of dominance, a violation of competition law normally necessitates some kind of collaboration between to independent parties. Companies that merely adapt their behaviour intelligently by monitoring the market and their competitors are not liable under Competition Law; "parallel behaviour" ie. simply following a market, is permissible. Below we examine some of the findings.
1. Algorithms as facilitators of anti-competitive practices
Firstly, the Study examined situations in which “traditional” anticompetitive practices resulting from prior contact between humans already exists and algorithms are used only to facilitate, support or monitor this practice or its effects.
Example: Trod and GBE, two poster traders in the UK, agreed not to undercut each other's online prices. They used pricing algorithms to enforce and monitor their agreement. The CMA mainly relied on evidence not directly related to the algorithms (such as e-mail) and fined the two companies. (CMA Decision of 12 August 2016 (Case 50223))
Generally, these scenarios do not pose new, particular Competition Law concerns: The algorithm is merely a means of supporting a conduct which is clearly anti-competitive already.
2. Algorithm-driven collusion involving a third party
Secondly, the Study examined scenarios in which a third party, e.g. an external consultant or software developer, provides the same algorithm or somehow coordinated algorithms to competitors. Different to the first scenario, there is no direct communication between the competitors. However, a certain degree of alignment, even if unintended, could arise from the actions of the third party.
The Study distinguishes between alignment at the level of the algorithm (code level) and an alignment at the level of the input factors (data level). Alignment at code level could arise, for example, if the parties involved were to delegate strategic decisions (such as pricing) to a common third party who takes these decisions using an algorithm. Alignment at data level could involve competitors using an algorithm as a means for an (undue) information exchange.
Example: Several travel agencies in Lithuania used an online booking platform operated by a third party named Eturas. Eturas applied a common cap on discounts applicable to services provided through it; the cap was communicated to the agencies through an internal messaging system in the form of an amendment to the platform's terms and conditions. It was implemented by Eturas using technical means.
Overall, cases falling under this 2nd category are much more difficult to assess from a Competition Law angle than the 1st category. Potential competition concerns in such situations will particularly depend on the content of the algorithmic alignment. Does it e.g. relate to current or future prices or quantities (which would be critical) or to less relevant parameters?
3. Collusion induced by the parallel use of individual algorithms
The third scenario analysed by the Study involves algorithms which are unilaterally designed and implemented, ie. each company uses a distinct pricing algorithm. There is no prior or ongoing communication or contact between the respective companies’ human representatives. Still, the fact that several or even all competitors rely on pricing algorithms might facilitate an alignment of their market behaviour, resulting from a mere interaction of computers.
Example: Competitors A, B, C and D each use algorithms to ensure the "best" pricing based on a multitude of factors including production costs, quantities in stock, competitor prices and expected future pricing. Eventually, the algorithms in place may align the price levels in such a way that they are (almost) identical. Does this involve a violation of Competition Law?
The Study is hesitant to deliver a clear-cut answer to the third scenario. So far, it is also unclear whether from a technical perspective a collusion "by chance" between different algorithms can arise. From a legal point of view, this scenario points back to the very basic observation that mere parallel behaviour and absent any evidence of collusion is permissible, even if algorithms allow for adaptation within milliseconds.
As algorithms are becoming more and more vital in every-day business, companies will need to closely monitor the regulators' approach to algorithms and – where necessary – adapt their business strategy accordingly. The 2nd example (Eturas) particularly demonstrates that companies have to be increasingly careful when relying on third party data or IT infrastructure, as this may "inadvertently" expose them to Competition Law risks. To that end, companies (including their legal department or external advisors) will need to understand exactly how the respective algorithm service operates and which data is at stake, dealt with or (potentially) manipulated.
A topic not touched on by the Study is whether companies should be entitled to access data harvested by larger companies to ensure a level playing field with those larger companies (or at least avoid a competitive disadvantage). So far, it is unclear under which circumstances such claims for data might be warranted. However, the German legislator has recently introduced a draft bill (so-called 10. GWB-Novelle) which establishes an explicit claim for access to data against dominant companies.