LLMs, legal contestability and scientific falsifiability

In this chapter Hildebrandt investigates the link between the contestability that is key to constitutional democracies on the one hand and the falsifiability of scientific theories on the other hand, with regard to large language models (LLMs). Legally relevant decision-making that is based on the deployment of applications that involve LLMs must be contestable in a court of law. The current flavour of such contestability is focused on transparency, usually framed in terms of the explainability of the model (explainable AI). In the long run, however, the fairness and reliability of these models should be tested in a more scientific manner, based on the falsifiability of the theoretical framework that should underpin the model. This requires that researchers in the domain of LLMs learn to abduct theoretical frameworks, based on the output models of LLMs and the real world patterns they imply, while this abduction should be such that the theory can be inductively tested in a way that allows for falsification. On top of that, researchers need to conduct empirical research to enable such inductive testing. The chapter thus argues that the contestability required under the Rule of Law, should move beyond explanations of how the model generates its output, to whether the real world patterns represented in the model output can falsify the theoretical framework that should inform the model.