On 10 December 2020, the Dutch Authority for Consumers & Markets (ACM) announced that it has launched a pilot investigation into the role of algorithms in commercial interactions between market players and the impact on market behaviour these algorithms may have. The ACM intends to put this knowledge to use in order to inform market parties on what they can expect in case the ACM would start an investigation into the functioning of their algorithms. The ACM is conducting the pilot in cooperation with Muziekweb, an online music library. On the same day, the ACM has published a position paper on the monitoring of algorithmic applications. This position paper is a starting point from which the ACM intends to further develop this type of monitoring. The paper provides general guidance on investigations into potential infringements in which algorithmic applications play a role.
In the position paper, the ACM describes what algorithmic applications are and in which cases they are relevant for the ACM’s competition law enforcement. The ACM subsequently explains how it can investigate algorithmic applications in practice and discusses some challenges it may face in practice. With this publication, the ACM contributes to the national and international debate on the application of artificial intelligence by companies and its supervision by regulators.
Trial with Muziekweb
Muziekweb uses algorithms to give visitors tailored music advice. In the trial that the ACM has set up with Muziekweb, the ACM investigates why the algorithm, for example, recommends rock music to one person and jazz to another. The ACM is not conducting an investigation into Muziekweb; in this trial, it only intends to map out what types of information it requires to be able to conduct research into the operation of algorithms. This knowledge is important for future investigations into infringements of the antitrust rules by market parties in which algorithms play a role.
The role of algorithms
The notion has grown that algorithmic applications increasingly play a (decisive) role in activities that have a direct impact on people, businesses and society as a whole. Companies are increasingly making use of algorithmic applications in their operational management. For example: for an optimal design of production processes, routing of deliverers, individualised offers, the dynamic adjustment of offers or for dynamic pricing. This brings many benefits to society. The application of algorithms can also bring benefits for market parties, such as better matching of supply and demand.
However, there are also many concerns in society about algorithmic applications that are used by companies to deliver their products and services to consumers. For example, in case companies use algorithms to guide consumers online in their choices and purchases, here such guidance may not always be in the interests of the consumer. There are also doubts as to whether there is still sufficient visibility and control over the operation and effects of complex algorithms. Hence, there is currently a great deal of attention in politics and society for the consequences of algorithmic applications for people, businesses and society as a whole.
Enforcement and algorithms
For the ACM, algorithmic applications are relevant to the extent that they play a role in activities affecting its enforcement areas, such as activities that have an impact on consumers or market operators. Algorithmic applications can, for example, steer supply and demand on the energy market, lead to price discrimination or the forming of cartels between market operators, or guide consumers towards purchasing decisions that are against their own interests.
The ACM has a team of fifteen algorithm and data experts at its disposal, which can investigate the role of algorithms in a possible infringement, the purpose and principles underlying the design and/or implementation of the algorithms and the input for and outcomes of the algorithms.
The ACM may investigate the use of algorithmic applications by market players in gathering information or evidence about a possible infringement. The scope of this investigation will depend on the circumstances and the elements that need to be proven.
The ACM may use the following powers to investigate algorithmic applications, based on the Dutch General Administrative Law Act (GALA):
- Company visits (‘dawn raids’) (art. 5:15 GALA)
- Requests for information (art. 5:16 GALA)
- Claiming access to business information and documents (art. 5:17 GALA)
The ACM already has some experience in using these investigative powers to gather digital evidence. The ACM also has experience in analysing algorithmic applications or studying their use. In the past, the ACM has carried out such investigations in the context of suspected infringements of competition law, the misleading of consumers or breaches of the Telecommunications Act.
The ACM will not in all cases need to directly investigate the actual (technical) functioning of algorithmic applications. Other information, such as (internal) correspondence, documentation and statements, may also provide sufficient evidence to support the case. However, in research into algorithmic applications, three main research questions can be distinguished:
- What is the role of an algorithm in the activity being investigated and which processes are involved (procedural transparency)?
- What is the behaviour of the algorithm (interpretability)?
- What is the functioning of the algorithm (technical transparency)?
Challenges for enforcement
The ACM will probably face various challenges when conducting research into algorithmic applications. Self-learning algorithms and their related data may be variable. Self-learning algorithms adapt themselves and may therefore behave differently from one moment to the next. However, the context in which they operate is volatile, both in terms of input for the system in operation and the behaviour linked to the outcome of the algorithm. Algorithms may also involve, for example, personal data that are aggregated but cannot be stored for privacy reasons. In case the ACM investigates a structural infringement that has been committed in the past, this volatility of (self-learning) algorithms and/or the related data can be a challenge.
Moreover, algorithmic applications do not operate in isolation. They are part of an ICT environment with links to (sub)systems and datasets that may belong to or fall under the responsibility of external parties. Particularly where it comes to applications that make use of complex algorithms, companies regularly use algorithmic applications from third parties. These applications can run locally or have remote addresses (cloud applications). Nevertheless, the fact that algorithmic applications from external parties are being used does not mean that the ACM cannot carry out any research into these algorithms. Third parties are also under an obligation to cooperate with a request by the ACM to access business information and documents in relation to an investigation into the conduct of their customers. Where several third parties are involved in the development or operation of algorithmic applications under investigation, this may complicate the ACM’s investigation. However, the starting point remains that the company that purchases the services and products from third parties for its business activities remains responsible for its own business activities.
Another challenge could be that, during investigations, the ACM is regularly confronted with data stored abroad by an external party. The ACM takes the view that such data can be reviewed or copied in an investigation if one of the following three criteria, or a combination thereof, is satisfied. The company under investigation must (i) be the owner of the data, (ii) manage the data, and/or (iii) be a user of the data.
The ACM states that experience shows that – depending on the situation – the copying of data stored abroad by an external party during an investigation works very well. This principle could also be applied when researching algorithmic applications. In case the algorithmic application can be copied as a whole in the form of separate files or of a (virtual) system, such files or systems can be considered similar to the copying of a data collection.
It is generally known that many algorithmic applications process large amounts of personal data. Research into the operation and behaviour of such algorithmic applications may require that the ACM obtains access to (parts of) such personal data. In that case, the ACM, as data controller needs to comply with the requirements of the General Data Protection Regulation (GDPR). Therefore, the ACM will need to identify in advance whether personal data will need to be processed during an investigation into algorithmic applications and, if so, what measures the ACM should take in that specific case to do so lawfully with sufficient safeguards.
As becomes clear from the foregoing, the ACM’s research into algorithmic applications is still at an early stage and there are quite some challenges that the ACM might face when investigating the role of algorithms in companies’ market activities. It will be interesting to see how the pilot will develop and whether the ACM will indeed be successful in implementing research strategies that can disclose the role of algorithmic applications in potential infringements of competition law.