Agent to Agent · Human–AI communication; empathic AI
Should Machines Be Empathic? An Integrative Review of Empathy Appropriateness in Large Language Models
Empathy is appropriate when it preserves the user's primary goal, and costly when it displaces it.
- Stage
- Conceptual paper · propositions developed
- Last activity
- 2026-08
- Collaborators
- Chi-Ching Juan
An LLM that responds to a debugging question by first acknowledging how frustrating the bug must be is doing something recognizably empathic and possibly counterproductive. This paper argues that appropriateness is not a property of the empathic move itself but of its relationship to what the user came to accomplish.
Using Multiple Goals Theory, we treat every message as carrying a hierarchy of goals and define empathy as appropriate when it preserves that hierarchy — recognizing affect in service of the primary goal rather than in competition with it. The framework yields five propositions and, more usefully for design, a diagnostic: the question is never how much empathy to express, but whether this expression advances or displaces the goal the user actually has. Emotional support escapes the critique entirely when support is the primary goal, which is why blanket prescriptions in either direction fail.
The record
- Idea
- Shift the debate from whether LLMs can display empathy to when empathy is appropriate, using Multiple Goals Theory to explain how empathic recognition can support or displace the user’s primary communicative goal.
- Research Question
- When is empathy appropriate in an LLM response, and what communicative principle should govern whether and how an LLM uses empathy?
- Key Proposition
- Empathy is appropriate when it preserves and advances the user’s primary communicative goal; it becomes inappropriate when socioemotional engagement competes with, redirects, or substitutes for that goal. Emotional support is not displacing when it is itself the primary goal.
- Data
- Conceptual synthesis of research on LLM empathy capability, perceived empathy, interaction context, human–AI differences, and the costs of excessive or misdirected empathy; no new empirical dataset in the current paper architecture.
- Analysis Result
- Conceptual result, not an empirical finding: the paper develops goal-hierarchy preservation, distinguishes goal-preserving empathy from goal displacement, and derives five propositions covering goal contingency, recognition need, facilitation of primary-goal pursuit, displacement costs, and emotional support as a legitimate primary goal.