Functional vs. phenomenological empathy in large language models: Rethinking artificial empathy through experimental evidence

Abstract

Although large language models (LLMs) are increasingly used in emotionally sensitive contexts, it remains unclear whether their responses reflect genuine empathy or functional simulation. This study employed a mixed-methods experimental design to compare empathetic responses from human participants (n = 100) with those generated by LLMs under three conditions: text-only, anthropomorphic cue, and multimodal with cue. Using ten standardized emotional vignettes, 200 independent raters evaluated all responses with the Consultation and Relational Empathy (CARE) measure. Quantitative results showed that LLM-generated responses were consistently rated as more empathetic than human responses, with empathy scores increasing across enhanced design conditions. Qualitative analysis using the Empathic Communication Coding System (ECCS) revealed that LLMs relied heavily on surface validation and supportive strategies, while demonstrating limited contextual probing and a lack of experiential grounding compared to human responses. These findings highlight a critical distinction between perceived empathy and phenomenological empathy. While LLMs can effectively simulate empathic communication and achieve high ratings due to consistency and alignment with social expectations, they lack the contextual depth and experiential understanding that characterize human empathy. This study contributes to the literature by moving beyond performance-based evaluations and clarifying the structural differences between human and artificial empathy. The results have important implications for the use of LLMs in healthcare, education, and counseling, where perceived empathy may enhance user experience but also raises concerns about over-trust and anthropomorphic attribution.

CRediT Statement

Ahmad Mugableh: Conceptualization, Methodology, Software, Validation, Formal Analysis, Investigation, Writing – Original Draft, Visualization, Supervision, Project Administration; Hissah Mohammed Alruwaili: Resource, Data Curation, Writing – Reviewing & Editing.

Disclosure Statement

The authors reported no potential conflicts of interest.

Generative AI Statement

The authors used an AI-assisted language tool during the preparation and revision of the manuscript to support language refinement, clarity improvement, formatting consistency, and editorial revision. The AI tool was not used to generate original data, conduct statistical analyses, interpret findings independently, or make scholarly decisions regarding the study’s design, methodology, results, or conclusions. All conceptual development, data analysis, interpretation, and final manuscript decisions were performed and verified by the authors, who take full responsibility for the manuscript's content.