Toward Constructive Optimisation: A new perspective on the regulation of recommender systems and the rights of users and society

How should we regulate systems designed to optimise digital environments and interactions? One needs to develop at least two critical perspectives to answer such a question. First, relative to what normative standards should optimisation be held? Second, how should regulation understand the tools of optimisation, such as ‘recommender systems’? This study develops an approach to both questions and integrates the corresponding perspectives into one answer. The study is divided into three main parts. In Part 1 a normative framework – centred around the values of self-development and self-determination – is elaborated as an interpretational resource to understand better how optimisation can be meaningful. When it comes to recommender systems, there is a need to move beyond naïve approaches, which implicitly assume that ‘the recommender system’ is an identifiable, discrete ‘unit’ that can be addressed and regulated as such. Instead, we propose to conceptualise and evaluate recommender systems through a so-called “stack approach”. This is the purpose of Part 2. The envisaged “stack approach” embraces the insight that beyond the surface interface level, recommender services are the result of different interactions, operations and layers, that are both social and technical in nature — software, hardware, infrastructure, organisational, design principles, and so on. All these parts work in concert to, ultimately, create particular tools, interfaces, and functionalities. Finally, Part 3 combines the normative framework of Part 1 and the stack approach of Part 2 for a critical analysis of the current approach to the regulation of recommender systems under the DSA, and for developing constructive suggestions of how to better account for the legitimate interests of users and society. Recommender systems should be regulated addressing every layer of the stack. Put simply, analysing and regulating the recommender system is not (only) about analysing and regulating the actual recommender engine, i.e., the software systems designed to fulfil optimisation logics, or the interface people interact with. The net should be cast wider. Optimisation goals determined by management, KPIs determined by business departments, performance reviews, hiring practices, data collection and analysis practices, iterative software design philosophies, UX/UI design choices, data training models, and so on, should all be incorporated into the bread and butter of recommender system regulation. This study, then, combines a more realistic, helpful approach to recommender systems as socio-technical artefacts with an original theoretical perspective on the normative standards we should hold optimisation systems to. In this report, we formulate a set of overarching recommendations that could guide future regulatory amendments. In an upcoming update and annex to this report, we will take up this exercise ourselves, and demonstrate how our model can be translated into concrete regulatory provisions. At the same time, we offer the stack approach as a toolkit to the reader: a starting point for reflection toward a more healthy and fair digital eco-system. In this context, it should be noted from the outset that the more realistic stack approach can be as enlightening as it can be overwhelming. The benefit of the approach is that it allows for a very wide, structural approach that cuts across the entire recommender value-chain or stack to show how a wide range of EU legislation can be used to regulate various elements of this ecosystem. The resulting analysis can, at the same time, also lead to what feels like a rather fragmented story – at least in terms of presentation. To further add to this enlightening complexity, the stack approach allows one to address separate layers of the stack individually, but one can also show how several layers (can) interact with one another in the regulatory context, or how ‘whole stack provisions’ address the entire stack. In short, the stack’s analytical modularity allows for a very all-encompassing mosaic approach that can address several analytical levels at the same time. Its inherent complexity is a feature, not a bug. This should be kept in mind when reading this exploratory study.