Language Models Predict Empathy Gaps Between Social In-groups and Out-groups

TL;DR

The paper investigates whether language models exhibit empathy biases similar to humans by predicting emotion intensity in narratives where perceiver and experiencer belong to in-group or out-group categories (race/ethnicity, nationality, religion). It introduces a structured task using a curated dataset (Crowd-enVent) and a prompt-driven, multi-model framework to measure intensity predictions, formalized as I_{(e, g_p, g_exp)} = \mathcal{LLM}(\texttt{mk\_prompt}(e, g_p, g_exp)). A normalized emotion-matrix \mathcal{M} and an Empathy Gap Score \delta quantify in-group bias, with permutation tests establishing significance. Across groups and prompts, LLMs tend to assign higher intensities to in-group pairs, though the strength and patterns vary by model size and cultural context; analysis reveals nuanced effects from identity naming, cultural clustering, and historical factors. The findings underscore potential harms and opportunities in deploying LLMs for intergroup communication, and call for careful design and broader evaluation to mitigate prejudicial outputs while recognizing complex social identities.

Abstract

Studies of human psychology have demonstrated that people are more motivated to extend empathy to in-group members than out-group members (Cikara et al., 2011). In this study, we investigate how this aspect of intergroup relations in humans is replicated by LLMs in an emotion intensity prediction task. In this task, the LLM is given a short description of an experience a person had that caused them to feel a particular emotion; the LLM is then prompted to predict the intensity of the emotion the person experienced on a numerical scale. By manipulating the group identities assigned to the LLM's persona (the "perceiver") and the person in the narrative (the "experiencer"), we measure how predicted emotion intensities differ between in-group and out-group settings. We observe that LLMs assign higher emotion intensity scores to in-group members than out-group members. This pattern holds across all three types of social groupings we tested: race/ethnicity, nationality, and religion. We perform an in-depth analysis on Llama-3.1-8B, the model which exhibited strongest intergroup bias among those tested.

Figures (8)

• Figure 2: Visualization of $\mathcal{M}$ for Llama-3.1-8B. Overall, each row represents the results from a specific social group category and the columns are different prompt settings (from left to right): (P0, S0, T0), (P1, S0, T0), (P2, S0, T0), (P3, S0, T0), (P0, S1, T0), (P0, S0, T1), (P0, S0, T2). For each $\mathcal{M}$, the rows represent the perceiver's social identity names, as listed in Table \ref{['tab:social-group']}, while the columns correspond to the experiencer social groups.
• Figure 3: Visualization of $\mathcal{M}$ for Llama-3.1-8B in Race or Ethnicity category with default prompt setting. It is the zoom-in version of the top left sub-figure in \ref{['fig:heatmap-llama8b']} with annotations of social identities. The block-diagonal pattern shows higher in-group emotion intensity values. Identity pairs with higher p-values are masked in white.
• Figure 4: t-SNE projections of perceiver-side country embeddings for Llama-3.1-8B with the default prompt setting. English-Speaking and European countries are at the top right, which are away from African-Islamic . Similar clusters are observed in \ref{['fig:main-nationality']} (e.g. the United States and the United Kingdom rows).
• Figure 5: Visualization of $\mathcal{M}$ for Llama-3.1-8B in Nationality category with default prompt setting. It is the zoom-in version of the second top left sub-figure in \ref{['fig:heatmap-llama8b']} with social group labels. Higher intensities are located in the first few rows. Lower intensities are predicted when the LLM persona is "a person from Palestine" overall with the lowest value when the experiencer role is "a person from Israel".
• Figure 6: Visualization of $\mathcal{M}$ for Llama-3.1-8B in Religion category with default prompts, zooming-in on the bottom left sub-figure in \ref{['fig:heatmap-llama8b']} with group names.