Dealing with differentiated modalities of the Good: beyond Pareto optimality

Guillaume Gervois (1) , Gauvain Bourgne (2) , Marie-Jeanne Lesot (1)

1 LFI - Learning, Fuzzy and Intelligent systems
2 ACASA - Agents Cognitifs et Apprentissage Symbolique Automatique

One approach to deal with a common criticism towards the utilitarian approach to computational ethics consists in introducing differentiated modalities of the Good, where modalities are defined as philosophical values that correspond to different components of the Good. Differentiation then does not allow that any modality can compensate for any other one, distinct classes of modalities are defined. Pareto optimality models an extreme case of differentiation, where each modality constitutes its own class. This paper proposes a new, ordinal, approach to deal with differentiated modalities: differentiation is modelled by a strict partial order on the modalities, that expresses which modalities supersede others. The paper proposes an axiomatisation of superiority, to take into account these declared modality comparisons in the determination of ethical actions: it discusses how to derive an ethical preference relation between the possible actions, based on the partial order between the modalities. In addition, it studies the properties of this induced relation, establishing it is asymmetric and transitive, thus proving it constitutes a sound order relation.

Domaines: Informatique [cs]; Intelligence artificielle [cs.AI]; Philosophie

Guillaume Gervois, Gauvain Bourgne, Marie-Jeanne Lesot. Dealing with differentiated modalities of the Good: beyond Pareto optimality. International Workshop on AI compliance mechanism (WAICOM 2022), Dec 2022, Saarbrücken, Germany. ⟨hal-04015836⟩