Learning to Classify Morals and Conventions: Artificial Intelligence in Terms of the Economics of Convention

Abstract

Artificial Intelligence (AI) and its relation with societies has become an increasingly interesting subject of study for the social sciences. Nevertheless, there is still an important lack of interdisciplinary and empirical research applying social theories to the field of AI. We here aim to shed light on the interactions between humans and autonomous systems and analyse the moral conventions, which underly these interactions and cause moments of conflict and cooperation. For this purpose we employ the Economics of Convention (EC), originally developed in the context of economic processes of production and management involving humans, objects and machines. We create a dataset from three relevant text sources and perform a qualitative exploration of its content. Then, we train a combination of Machine Learning (ML) classifiers on this dataset, which achieve an average classification accuracy of 83.7%. A qualitative and quantitative evaluation of the predicted conventions reveals, inter alia, that the Industrial and Inspired conventions tend to co-exist in the AI domain. This is the first time, ML classifiers are used to study the EC in different AI-related text types. Our analysis of a larger dataset is especially beneficial for the social sciences.