Being Considerate as a Pathway Towards Pluralistic Alignment for Agentic AI

TL;DR

This paper explores how agentic AI can achieve pluralistic alignment by being considerate of human values and perspectives, promoting harmony between AI objectives and diverse human interests.

Contribution

It introduces the concept of considerate behavior as a pathway to pluralistic alignment in agentic AI systems, emphasizing the importance of mindfulness towards others' wellbeing.

Findings

Being considerate promotes pluralistic alignment.

Considerate AI can better harmonize with human values.

The approach enhances AI's social compatibility.

Abstract

Pluralistic alignment is concerned with ensuring that an AI system's objectives and behaviors are in harmony with the diversity of human values and perspectives. In this paper we study the notion of pluralistic alignment in the context of agentic AI, and in particular in the context of an agent that is trying to learn a policy in a manner that is mindful of the values and perspective of others in the environment. To this end, we show how being considerate of the future wellbeing and agency of other (human) agents can promote a form of pluralistic alignment.