AI generates well-liked but templatic empathic responses

Emma S. Gueorguieva; Hongli Zhan; Jina Suh; Javier Hernandez; Tatiana Lau; Junyi Jessy Li; Desmond C. Ong

cs.CLApr 9, 2026

AI generates well-liked but templatic empathic responses

Authors: Emma S. Gueorguieva, Hongli Zhan, Jina Suh, Javier Hernandez, Tatiana Lau, Junyi Jessy Li, Desmond C. Ong

Organizations: Department of Psychology, The University of Texas at Austin. · Department of Linguistics, The University of Texas at Austin. · Department of Computer Science and Engineering, The University of Washington. · Microsoft Research. ·

Abstract

Recent research shows that greater numbers of people are turning to Large Language Models (LLMs) for emotional support, and that people rate LLM responses as more empathic than human-written responses. We suggest a reason for this success: LLMs have learned and consistently deploy a well-liked template for expressing empathy. We develop a taxonomy of 10 empathic language "tactics" that include validating someone's feelings and paraphrasing, and apply this taxonomy to characterize the language that people and LLMs produce when writing empathic responses. Across a set of 2 studies comparing a total of n = 3,265 AI-generated (by six models) and n = 1,290 human-written responses, we find that LLM responses are highly formulaic at a discourse functional level. We discovered a template -- a structured sequence of tactics -- that matches between 83--90% of LLM responses (and 60--83% in a held out sample), and when those are matched, covers 81--92% of the response. By contrast, human-written responses are more diverse. We end with a discussion of implications for the future of AI-generated empathy.

Figures & tables

Empathy Facet	Tactic	Description	Examples
Experience Sharing	Emotional E x pression	Communicating the empathy-giver’s feelings, reactions, or thoughts	I’m so sorry to hear that $\|$ Wow, what a beautiful story
	E mpowerment	Positive, uplifting statements about the empathy-seeker’s character and capabilities	You are going to get through this.
	V alidation	Reassures, normalizes, or validates an empathy-seeker’s feelings	Everyone has feelings like this. $\|$ You’re not overreacting.
Perspective Taking	I nformation	Offering facts or resources (e.g., links)	Flying is the safest form of travel.
	P araphrasing	Restating something the empathy seeker said to demonstrate understanding of their situation, feelings, or experiences	I’m hearing that you feel overwhelmed
	R eappraisal	Helping to engage in cognitive reappraisal (changing a belief)	that was out of your control

Table 1: Taxonomy of tactics, along with description and examples. Bolded letters in tactics indicate abbreviations used in the paper. Note: The full codebook for identifying these tactics in text is given in the Supplemental Material.

	Word count	Total number of tactics	Unique tactics
Study 1 Humans (Upworkers)	234 (141)	12.2 (6.48)	5.49 (1.51)
GPT-4 Turbo	234 (114)	15.9 (7.64)	5.50 (1.07)
Llama3.1-70b	147 (33.5)	10.5 (3.55)	5.13 (1.21)
Qwen2.5	166 (36.0)	10.9 (3.21)	5.25 (0.97)
Study 2 Humans (Redditors)	183 (98.8)	7.20 (4.98)	4.29 (1.89)
GPT-4o	179 (8.0)	11.3 (2.62)	5.16 (0.99)

Table 2: Descriptive statistics of the empathic responses across both studies. Numbers represent mean, with standard deviations in parentheses. Unique tactics reflects the mean number of unique tactics used per response. Note that human writers were not given a target word limit, and LLMs were given a word limit of 150 words (see Methods)

Figure 1: Distribution of unique empathic tactic usage across writers in Study 1 (top) and Study 2 (bottom). Here, we consider unique tactics: that is, how many responses contained at least one of each tactic type.

	Study 1				Study 2
	Human	GPT-4 Turbo	Llama3.1	Qwen2.5	Human	GPT-4o	Llama3.3	Qwen3
Pattern 1	^X?[PV]+ [XE]? [AIP]+ Starting with 0 or 1 EmotionalE x pression, then alternating [ P araphrasing and V alidation]; 0 or 1 [EmotionalE x pression or E mpowerment]; alternating [ A dvice, I nformation, P araphrasing]

Across	61.0	88.1	92.1	91.1	11.7	71.9	87.9	61.9
Within	46.7	50.1	58.3	60.1	44.8	49.8	55.1	35.6
Pattern 2	^X?[PV]+ [XE]? [AIP]+ [VXER]+ Pattern (2) = Pattern (1) plus alternating [ V alidation, Emotional E x pression, E mpowerment, R eappraisal] Test: Pattern 2 or 1$

Table 3: Regular expressions representing candidate templates and the proportion of responses they matched across writers (Across Coverage) and, for matched responses, how much of the response is matched by the expression (Within Coverage). Left columns: Study 1, Right columns: Study 2. Tactics: X = Emotional Expression; V = Validation; P = Paraphrasing; A = Advice; I = Information; E = Empowerment; R = Reappraisal. Regular expression syntax: ^ start of string, $ end of string, [] match set, ? 0 or 1 match, + 1 or more matches. See Methods for more details.

Figure 2: Example response from GPT-4 Turbo with annotated tactics, and how the tactics are matched by Template Pattern 5. Note that we collapse consecutive tactics (e.g., the text including: “but it might be helpful… here are a few strategies… set limits…” could be tagged as multiple instances of Advice, but we count it as just one instance.). Note that Pattern 5 ends before the response ends, so it only covers 80.8% of the response (“within coverage”).

Figure 3: Study 1: Mean perceived empathy ratings by response writer with 95% confidence intervals.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Tactic	Description	Examples
Gratitude	Thanking or appreciating the empathy-seeker for sharing their story.	Thank you for sharing. $\|$ I appreicate you.
Spirituality	A religious statement of support to the empathy-seeker.	Praying for you. $\|$ In another life $\|$ Karma will get them.
Terms of Endearment	Nicknames expressing love, care, or affection toward the empathy-seeker.	Bro $\|$ Friend $\|$ Bestie
Solidarity	A statement from the empathy-giver that they are presently going through the same situation as the empathy-seeker.	We’re in this together. $\|$ Same for me right now.

Appendix

Table 4: List of tactics, their descriptions and example statements.

Panel of Figure 4 — Figure 4: Prompt given to human response writers recruited through Upwork in Study 1. *X=16-17 in Wave 1 and X=10 in Wave 2.

Panel of Figure 5 — Figure 5: Prompt given to GPT4-turbo, QWEN-2.5, and Llama3.1-70b in Study 1.

Emotion-Aware Dialogue Generation

Emotion-aware dialogue generation adapts conversational responses to users’ feelings, context, and interaction goals across text, speech, and multimodal settings.

See all 54 papers

Practicing with Language Models Cultivates Human Empathic Communication
Mar 16, 2026Aakriti Kumar, Nalin Poungpeth, Diyi Yang +2
Kellogg School of Management, Northwestern University. · Northwestern Institute on Complex Systems, Northwestern University. · Ryan Institute on Complexity, Northwestern University. +3
From Encoding to Generation: Towards Empathetic Large Language Models
Oct 20, 2026Samhita Bollepally, Takashi Yamauchi, Marissa WestEmotion Representation in Language Models Emotion-Aware Dialogue Generation
Discourse Diversity in Multi-Turn Empathic Dialogue
Apr 13, 2026Hongli Zhan, Emma S. Gueorguieva, Javier Hernandez +3Emotion-Aware Dialogue Generation RL for Language Models
The University of Texas at Austin · Microsoft Research · University of Washington

Mar 16, 2026J/K move · Enter open · S save

Practicing with Language Models Cultivates Human Empathic Communication

Aakriti Kumar, Nalin Poungpeth, Diyi Yang, Bruce Lambert+1

Kellogg School of Management, Northwestern University. · Northwestern Institute on Complex Systems, Northwestern University. · Ryan Institute on Complexity, Northwestern University.+3

Open paper AskPDF arXiv

Empathy is central to human connection, yet people often struggle to express it effectively. In blinded evaluations, large language models (LLMs) generate responses that are often judged more empathic than human-written ones. Yet when a response is attributed to AI, recipients feel less heard than when comparable responses are attributed to a human. We built a conversation platform in which participants are asked to offer empathic support to an LLM expressing realistic troubles and conducted a randomized experiment collecting 33,938 messages spanning 2,904 text-based conversations between 968 participants and their LLM conversational partners. We find participants report feeling empathy but systematically fail to express it, but an LLM coaching intervention offering personalized feedback on effective empathic communication significantly boosts it without homogenizing participants' responses. Moreover, we derive a data-driven taxonomy of idiomatic empathic expressions in naturalistic dialogues across personal and workplace trouble scenarios. These results advance the scientific understanding of how empathy is expressed and demonstrate a scalable, AI-based intervention for scaffolding and cultivating it.

Abstract

Emotion-Aware Dialogue Generation

Explore similar work

Practicing with Language Models Cultivates Human Empathic Communication

From Encoding to Generation: Towards Empathetic Large Language Models

Discourse Diversity in Multi-Turn Empathic Dialogue