When AI Sounds More Empathetic Than Humans: Experimental Evidence from Indonesian EFL Students

Large language models (LLMs) increasingly participate in emotionally sensitive communication, yet it remains unclear how their empathic language is perceived by English as a Foreign Language (EFL) learners. This study investigated whether Indonesian EFL students perceived LLM-generated responses as more empathetic than human-generated responses and examined the linguistic characteristics shaping those perceptions. A mixed-methods within-subject experimental design was employed with 100 undergraduate students in the English Education Study Program at STKIP PGRI Bandar Lampung, Indonesia. Participants evaluated paired human- and LLM-generated responses to ten standardized emotional vignettes under blind-source conditions. Perceived empathy was assessed across emotional understanding, validation, warmth, perspective-taking, supportiveness, and contextual appropriateness, while semi-structured interviews with 18 purposively selected participants explored the linguistic features influencing their judgments. Quantitative findings indicated that LLM-generated responses received higher overall empathy ratings than human-generated responses. Qualitative analysis showed that emotional acknowledgment, validation, reassurance, supportive guidance, lexical warmth, and non-judgmental wording contributed strongly to perceptions of empathy. Human responses, however, were more frequently associated with authenticity and contextual specificity, whereas some LLM responses were perceived as polished but formulaic. These findings suggest that perceived empathy in human–AI communication is shaped largely by linguistic performance and reader interpretation rather than by evidence of genuine emotional experience. The study extends research on artificial empathy into the EFL context and highlights the importance of critical AI literacy in helping pre-service English teachers distinguish persuasive empathic language from authentic human emotional understanding. It also demonstrates value of blind comparisons for examining interpersonal meaning in AI-mediated second-language communication.