The Normative Value of Legal Analytics. Is There a Case for Statistical Precedent?

Abstract

This work contributes to the harmonisation of the quantitative methodologies of data science, computer science and statistics with the qualitative methodology of the law. It gives a layered answer to the research question whether the outcome of a machine learning algorithm with case law as an input can have normative value. The thesis argues first that the outcome of a machine learning algorithm with case law as an input is not ‘law’ as we know it today. Neither is it a fact in a court case, nor a secondary source of law. The thesis claims furthermore that for methodological reasons, such an outcome is to be considered as a ‘sui generis’ concept, a concept of its own kind, with which courts can, and even should, engage in adjudication. In addition, it is argued that modelling with machine learning can have an implicit normativity through the definition of the purpose of the algorithm, its design and the choices that are made by the software engineers. In the first part, the work introduces several building blocks that inform the following parts. The second part is a critical analysis of 9 experiments with mainly supervised machine learning algorithms, with case law as an input. The final part discusses the use of the outcome of such algorithms in court cases.