Your Mileage May Vary: How Empathy and Demographics Shape Human Preferences in LLM Responses

Yishan Wang, Amanda Cercas Curry, Flor Miriam Plaza-del-ArcoEMNLP 2025

TL;DR

This paper reveals that user demographics and empathy levels significantly influence preferences for Large Language Model (LLM) responses to emotional dilemmas. It highlights the need to move beyond a "one-size-fits-all" approach to LLM empathy, suggesting that tailoring responses based on individual user characteristics can lead to more effective and inclusive human-AI interactions.

What Problem Does This Paper Address?

As LLMs become increasingly integrated into subjective decision-making processes, such as providing advice or engaging in moral reasoning, it's crucial to understand whose preferences these models align with and why. While previous research has often relied on aggregate human judgments, this paper addresses the underexplored issue of demographic variation and its linguistic drivers in shaping preferences for LLM-generated responses, particularly in emotionally charged scenarios. The authors aim to bridge the gap between understanding general LLM capabilities and tailoring them to diverse user needs and expectations.

Methods

The study employs a multi-dimensional analysis approach. First, they integrate demographic and self-reported empathy data (cognitive and affective) from human participants with their ratings of LLM responses to emotional dilemmas. Second, they connect these human preferences to the linguistic features of the LLM responses using tools like the Linguistic Inquiry and Word Count (LIWC) dictionary and the Connotation Frames of Power and Agency. This allows for a systematic investigation of how user characteristics and response styles influence perceived empathy and overall preference.

Datasets

The study utilizes data collected for the Language Model Council (LMC) framework, which includes emotionally charged dilemmas and LLM responses. This dataset extends the EmoBench benchmark and consists of 100 dilemmas, with 20 LLMs ranking each other's open-ended responses, resulting in 2,000 evaluated responses. Human participants then evaluated pairs of these responses. Demographic data (age, gender, country of origin, education, AI use) and empathy scores (Basic Empathy Scale) were also collected from participants.

Key Insights and Ideas

  • Empathy is Not Monolithic: The study strongly suggests that "empathy" as a general trait is insufficient. Distinguishing between cognitive empathy (understanding others' perspectives) and affective empathy (sharing others' emotional experiences) reveals different preferences.
  • Demographics Matter: User demographics, such as gender and country of origin, significantly correlate with preferences for LLM response styles. This challenges the idea of a universal "best" response.
  • Linguistic Cues are Key: Specific linguistic features, like the use of tentative phrasing, power verbs, and emotional tone, are strong predictors of group-level preference differences.
  • Context is Crucial: The findings highlight that cultural or national contexts might shape expectations about what constitutes a "good" response, influencing whether users prioritize actionability, emotional attunement, or pragmatic usefulness.
  • Beyond "One-Size-Fits-All": The research directly questions the assumption that a single, universally empathetic LLM output is optimal, advocating for more nuanced and personalized approaches.

Key Results

  • Demographic Divides: Significant demographic divides were observed in LLM response preferences. For instance, female participants exhibited higher empathy scores than males.
  • Empathy Level Impact: Participants with higher affective empathy tended to prefer responses that included emotional validation and compassionate language, while those with lower empathy levels favored more straightforward, practical answers.
  • LLM Performance Variation: Models like Claude-3-Opus, GPT-40, and Qwen1.5-110B showed broader cross-demographic appeal, consistently ranking in the top three. Conversely, smaller models like GPT-4 and LLaMA-3-8B ranked lower, suggesting a more limited alignment with diverse user expectations.
  • Linguistic Feature Correlation: The study identified markers like power verbs and tentative phrasing that predict group-level differences. For example, Qwen models showed a tendency towards thoughtful, reflective styles emphasizing reasoning, while Claude-3-Opus and GPT-40 scored high in affective and positive emotion language.
  • Model Ranking Variance: Qwen1.5-110B exhibited the highest rank variance across demographic groups, indicating its perceived quality varied widely.

Limitations and Future Work

The authors acknowledge several limitations:

  • Demographic Imbalance: The user preference data was not evenly distributed across all demographic subgroups, potentially biasing observed patterns and limiting generalizability.
  • Indirect Empathy Assessment: The study analyzed linguistic markers but did not directly assess the models' ability to express or understand empathy through specific empathy evaluations.
  • Lexical Category Reliance: The stylistic analyses relied on predefined lexical categories (LIWC, connotation frames), which might overlook context-dependent language use or dynamic emergent patterns.

Future work could address these limitations by:

  • Collecting more demographically balanced feedback.
  • Incorporating empathy-specific evaluations and established benchmarks.
  • Exploring complementary analytical tools to capture a richer understanding of language use.
  • Developing LLMs that can be tailored to individual user traits (cognitive/affective empathy) for more satisfying and appropriate responses in sensitive contexts.
  • Investigating how to dynamically adjust LLM tone and style based on lightweight user profiling or behavioral cues.

Summary updated

BibTeX
@inproceedings{wang2025your,
  author = {Yishan Wang and Amanda Cercas Curry and Flor Miriam Plaza-del-Arco},
  title = {{Your Mileage May Vary: How Empathy and Demographics Shape Human Preferences in LLM Responses}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2025},
}

Loading related content...

Figure