Engineering Kindness: Building a Machine with Compassionate Intelligence

Introduction

Repeated interactions with the artifacts we create have a rub-off effect. They are changing us. Recent co-discoveries across a number of different fields support the idea that positive emotion has many good side effects so at this juncture in AI, it is time to consider the idea of designing compassionate intelligence (CI). The paper outlines a generalized software requirements description and describes the EM-2, a software agent with CI. The work represents first steps towards building machines (robots and software agents) that have a stake in us, by programming with compassionate intelligence based on both state of mind and state of heart. We offer the generalized software requirements description for anyone who wants to pursue this direction of programming in AI/Robotics. This direction is significant not just for AI but for user interfaces, healthcare, education, and design in other fields. We illustrate the CI ideas with code snippets and architectures of software agent, EM-2.

What does it meant to give an AI system compassionate intelligence or have a stake in us? Simply, we mean the ability to program an AI system that makes decisions and actions that take into account positive regard for others and self. We describe software agent EM-2 as a first step to an AI program that represents and reasons not only with logical concepts of mental state but state of the heart. EM-2 components leverage the sensory analysis and multi-agent functions of prior agent systems, however in the paper we focus on elements of EM-2 related to CI. At the programming level this includes multi-level representations and predicates of feelings of self and others as well as logical concepts and the logical concepts about feelings. The work presented here covers several AI concepts and software agent systems: cooperative multi-agent systems technology, emotion oriented programming, common sense knowledge, affective inference, default reasoning and belief revision. We do not presume the agent to “have” feelings, nor do we address this issue here. Rather we address the computational aspects of creating a reasoning apparatus that uses representation of these concepts to accomplish compassionate decision making. Essential to the system is a pro-social agential stance – this includes but is not limited to a) agents do not lie about concepts or feelings and there is common sense knowledge of positive emotion, society and culture.

The motivation for this work is based on a growing body of recent discoveries from social and cognitive neuroscience, psychoneuroimmunology, and genetics indicating humans benefit greatly from compassionate experiences. User experience studies, neuroplasticity and genetic plasticity studies indicate we are literally changed by repeated interactions with objects and relations in our environment. With our growing symbiotic relationship with gadgets (phones, cars, robotic assistants, browsers, appliances, IoT, etc.) there is significant biological imperative to intentionally design for humane and pro-social AI systems and interfaces. It’s not to say all AI needs such features, but that we have a choice when it makes sense to do so.The technical components of the paper are divided into three parts. In Part 1 we present a generalized software requirements guide for building AI systems with compassionate intelligence and then detail the EM-2 implementation of the first three design requirements, namely: a) agent philosophy of mind and architecture that supports it b) a representation and reasoning calculus that supports notions of combining emotion, standard logic, self and other and c) an inferencing and learning component that supports a) and uses b) to make decisions and take action based on consideration of self and other. In Part 1I we increase our exploration of the issues of combining affect and logical fact components in multiple agents by deepening the descriptions of algorithms and code snippets given in Part 1 with a machine theoretic architecture and description of the introspective machines for beliefs and feelings of EM-2, including the predicates for introspection and meta-cognition, conflict resolution and naming. Part 1II of the paper discusses social and technical issues surrounding this topic such as human-level AI, compassionate intelligence, etc.