Is Empathic AI Possible?

Abstract

Efforts to develop “empathic AI” often assume a universal definition for empathy—a singular, objective understanding of “human” experiences. This implies that achieving empathic AI is only a matter of optimization difficulty, where researchers should find the “correct” way to align AI systems with human-like empathy—essentially an engineering problem. In this chapter, we challenge this assumption. We begin by emphasizing the multifaceted nature of empathy. Then, we explore how variations in (1) the multiple forms of empathy and (2) the many groups we are embedded within present significant challenges in creating a universally applicable empathic AI. We discuss the difficulties of codifying empathy (an inherently context-dependent phenomenon) into AI systems, as well as the recent evidence for cultural biases and moral stereotypes in widely used Large Language Models (an important class of AI systems), highlighting the broader ethical, epistemic, and possibly existential issues inherent in designing machines that “understand” or “feel” others as humans do. We conclude by advocating for a more contextualized approach to empathic AI, one that is culture-aware, context-sensitive, and pluralistic, moving beyond the reductionist notion of “humans” as a monolith. We do not intend to address the question in our chapter’s title with a simple yes-or-no answer; instead, we advocate for asking questions such as “What kind of empathy?” and “Empathy for whom?” as researchers and engineers move toward developing more empathic generative language models.