Dis/similarities in the design and development of legal and algorithmic normative systems: the case of Perspective API

Abstract

ABSTRACT For several decades now, legal scholars and other social scientists have been interested in conceiving of technologies as regulatory media and comparing their normative affordances to law’s regulatory characteristics. Recently, scholars have also started to explore the normative nature of Machine Learning and Artificial Intelligence systems. Most of this scholarship, however, adopts a largely theoretical perspective. This article takes a different approach and attempts to provide the discussion with a more factual grounding. It does so by investigating the construction of one particular Machine Learning system, the content moderation system Perspective API developed by Google. Its open-source development and a voluminous trove of publicly available documentation render Perspective API a virtually unique resource to study the inner logics of Machine Learning systems development. Based on an in-depth analysis of these logics, the article fleshes out similarities and dissimilarities concerning the normative structure of algorithmic and legal systems regarding four different subjects: practical constraints, evaluative diversity, modes of evolution and standards of evaluation. The article then relates the case of Perspective to the European Union’s proposal for an Artificial Intelligence Act and shows how the study’s insights might help in readying the Act for the realities of contemporary multi-party AI development.