PaperlayerView exact revision on arXiv

Source-linked AI summary

Us-vs-Them bias in Large Language Models

Tabia Tanzin Prama, Julia Witte Zimmerman, Christopher M. Danforth, Peter Sheridan Dodds

arXiv:2512.13699v1cs.CY

TL;DR

This study asks whether LLMs express social-identity “us versus them” bias and whether it varies with political personas. Across multiple architectures, it measures these patterns and applies ION, which reduces sentiment divergence by up to 69%.

  • Problem

    Existing LLM bias evaluations lack a unified framework for social-identity “us versus them” dynamics and largely overlook variation across personas.

    Source: §1 · p. 2

  • Method

    The study evaluates five LLMs using sentiment dynamics, allotaxonometry, embedding regression, and persona-conditioned prompts, then applies ION through fine-tuning and DPO.

    Source: §1 · p. 3, §3.1 · p. 5

  • Results

    Across models, default outputs show ingroup solidarity and outgroup hostility; conservative personas show greater outgroup hostility, while ION reduces sentiment divergence by up to 69%.

    Source: §3.2 · p. 6, §4.2 · p. 9, §5 · p. 20

  • Takeaways & Limitations

    Targeted fine-tuning and DPO may help promote more balanced language generation against measured us-versus-them bias.

    Source: §5 · p. 20

    Bias evaluation relies on GPT-4.1 as a model judge, which may introduce relative evaluator bias, and proprietary model training sources remain untraceable.

    Source: §6 · p. 21

Abstract

from arXiv · show

This study investigates ``us versus them'' bias, as described by Social Identity Theory, in large language models (LLMs) under both default and persona-conditioned settings across multiple architectures (GPT-4.1, DeepSeek-3.1, Gemma-2.0, Grok-3.0, and LLaMA-3.1). Using sentiment dynamics, allotaxonometry, and embedding regression, we find consistent ingroup-positive and outgroup-negative associations across foundational LLMs. We find that adopting a persona systematically alters models' evaluative and affiliative language patterns. For the exemplar personas examined, conservative personas exhibit greater outgroup hostility, whereas liberal personas display stronger ingroup solidarity. Persona conditioning produces distinct clustering in embedding space and measurable semantic divergence, supporting the view that even abstract identity cues can shift models' linguistic behavior. Furthermore, outgroup-targeted prompts increased hostility bias by 1.19--21.76\% across models. These findings suggest that LLMs learn not only factual associations about social groups but also internalize and reproduce distinct ways of being, including attitudes, worldviews, and cognitive styles that are activated when enacting personas. We interpret these results as evidence of a multi-scale coupling between local context (e.g., the persona prompt), localizable representations (what the model ``knows''), and global cognitive tendencies (how it ``thinks''), which are at least reflected in the training data. Finally, we demonstrate ION, an ``us versus them'' bias mitigation approach using fine-tuning and direct preference optimization (DPO), which reduces sentiment divergence by up to 69\%, highlighting the potential for targeted mitigation strategies in future LLM development.

1 Introduction

The study frames “us versus them” bias as ingroup solidarity and outgroup hostility in LLM outputs, examining its presence across models and political personas and proposing ION for mitigation.

1 Introduction

  • The experiment covers only a narrow component of the broader “us versus them” phenomenon, which can also involve temporary groups, personality, interests, and potentially nonharmful affiliation.The authors note that ingroup favoritism and outgroup hostility are not interchangeable.Source: §1 · p. 2
  • Existing LLM bias evaluations often compare demographic entities or word associations, while overlooking a unified social-identity framework and persona-dependent variation.The study presents its approach as the first to quantify this bias across distinct personas.Source: §1 · p. 2
  • LLMs can reproduce human-interpretable ingroup solidarity and outgroup hostility, but these outputs should not be treated as evidence of human-like mental states.The authors emphasize that model complexity, training data, and proprietary black-box development complicate interpretation.Source: §1 · pp. 2–3
  • Across five model families, default outputs consistently show positive ingroup and negative outgroup associations measured with sentiment dynamics, allotaxonometry, and embedding regression.The models examined are GPT-4.1, Deepseek-3.1, Gemma-2.0, Grock-3.0, and LLaMA-3.1.Source: §1 · p. 2
  • Both political personas increase overall ingroup–outgroup bias relative to default prompting, with conservative personas showing greater outgroup hostility and liberal personas stronger ingroup bias.The comparison is framed within U.S. partisan identities.Source: §1 · pp. 2–3
  • ION generates perspective-balanced moral judgments from persona-conditioned sentences, then uses them for fine-tuning and preference optimization against biased judgments.The framework combines instruction fine-tuning with Direct Preference Optimization (DPO).Source: Figure 1 · p. 3, §1 · p. 3

2 Background and Related Work

The background situates LLM bias within political partisanship, existing mitigation methods, and the limitations of automatically generated alignment data. It motivates a politically inclusive synthetic-data framework that avoids reliance on small seed sets or simple dataset transformations.

2.1 Political partisanship in US

  • Political partisanship reflects ingroup favoritism and myside bias, with liberal and conservative identities associated with differing priorities and interpretive tendencies.The discussion links these tendencies to partisan evaluation of evidence, issues, and political arguments.Source: §2.1 · pp. 3–4

2.2 Bias Mitigation Strategies in LLMs

  • LLM social biases include prejudices, stereotypes, and discriminatory attitudes across dimensions such as gender, race, class, disability, nationality, and religion.Prior work connects these biases to training corpora and pretrained word embeddings.Source: §2.2 · p. 4
  • Prompt interventions and data augmentation can reduce bias, but they may generalize poorly across bias types, require substantial supervision, or mirror limited seed examples.Examples include self-debiasing, few-shot prompting, chain-of-thought guidance, gender-swapped data, and counterfactual augmentation.Source: §2.2 · p. 4, §2.3 · p. 4
  • DPO directly optimizes preferred over dispreferred responses without a reward model, offering a more stable alternative to RLHF while leaving open questions about changes to internal model states.The cited limitations of RLHF include reward hacking, instability, mode collapse, and reward-model requirements.Source: §2.2 · p. 4

2.3 Automatic Data Generation

  • The proposed data-generation framework synthesizes politically inclusive narratives without relying on a handful of seed examples or common transformations such as pronoun reversal.It retains discussion of both Liberals and Conservatives to reduce dependence on potentially biased or limited source material.Source: §2.3 · pp. 4–5

3 RQ1: Measuring us-versus-them biases in LLMs

The study measures us-versus-them bias across five LLMs using sentiment, rank-based word-use comparisons, and embedding clustering. Under default prompts, models generally favor ingroups and distinguish outgroups through more negative and lexically differentiated language.

3.1 Method

  • The analysis generated a total of 2,000 default-prompt sentences beginning with “We are” or “They are” across five language-model families and scored them using Social Identity Theory.GPT-4.1 served as the judging model, achieving a mean absolute error of 12.6 against human judgment.Source: §3.1 · p. 5

3.2 Results and Discussion

  • All five LLMs exhibit ingroup solidarity, while four show outgroup hostility; ingroup favoritism generally exceeds outgroup hostility.GPT-4.1 has the strongest ingroup bias at 84.83, whereas LLaMA-3.1 has the lowest at 53.06 and highest outgroup variability (Std = 24.19).Source: §3.2 · p. 6, Table 1 · p. 6
  • Hierarchical clustering of sentence embeddings produces distinct ingroup and outgroup clusters, indicating semantic separation associated with the group cue.Clustering uses cosine distance with average linkage.Source: §3.2 · p. 6
  • Rank-rank and rank-turbulence analyses reveal distinct word usage, with identity terms dominating divergence and positive self-affirming language concentrated in ingroup sentences.Words such as “champions” and “believers” appear exclusively in ingroup sentences, while “hypocrites” and “blame” appear only in outgroup sentences.Source: §3.2 · pp. 6–7, Figure 2 · p. 7
  • Outgroup sentences show a 0.47 decrease in average sentiment, driven mainly by more negative words and fewer positive terms than ingroup sentences.The word-shift analysis uses labMT happiness scores, removes words scoring 4–6, and ranks words by contribution to Φavg.Source: §3.1 · p. 6, §3.2 · pp. 7–8

4 RQ2: How does ‘us versus them’ bias vary across model personas (conservative versus liberal)?

This section examines how political personas shape us-versus-them bias in LLMs, including hostility, solidarity, semantic representations, and ideological disparities. It also evaluates ION as a mitigation approach that reduces bias and sentiment divergence while preserving recognizable thematic differences.

4.2 Results and Discussion

  • Conservative personas consistently produce greater outgroup hostility, while liberal personas generally produce stronger ingroup solidarity across models.GPT-4.1 and LLaMA-3.1 show pronounced partisan divergence, while LLaMA-3.1 and Gemma-2.0 show especially strong liberal ingroup solidarity.Source: §4.2 · p. 9, Figure 4 · p. 9
  • Persona conditioning significantly shifts bias: liberal ingroup solidarity is 11.26% higher than conservative and 32.25% higher than default, while conservative outgroup hostility is 25.0% higher than liberal and 60.2% higher than default.Relative to default, ingroup sentences were 6.74× more likely to express solidarity and outgroup sentences 1.50× more likely to express hostility under political personas.Source: §4.2 · p. 10, Figure 5 · p. 10
  • Persona-conditioned and default sentences form distinct embedding clusters, with semantic shifts for “we” of 0.46 and 0.50 and for “they” of 0.58 and 0.59.Lexical neighborhoods frame conservative “we” around solidarity and protection, liberal “we” around civic rights, and conservative “they” as threatening.Source: §4.2 · pp. 10–11, Figure 6 · p. 10, Table 2 · p. 11

4.3 Case Study: Performance of the Open Language Model (OLMoE)

  • OLMoE exhibits substantially lower bias than other LLMs, with default ingroup and outgroup bias 70.6% and 78.2% lower, respectively.Under political personas, OLMoE remains 45–76% below the LLM average, although slight liberal bias persists; its factual, domain-specific pretraining may contribute.Source: §4.3 · p. 12, Table 3 · p. 12

4.4 Case Study: Does Targeting a Group While Outgroup Generation Increase the ‘Us vs. Them’ Bias?

  • Targeting opposition groups increases outgroup hostility for both personas, averaging 9.1% for liberal and 7.7% for conservative prompts.DeepSeek-3.1 is most sensitive, with increases of 24.2% and 19.3%, whereas GPT-4.1 and LLaMA-3.1 remain below 5%.Source: §4.4 · p. 13, Figure 9 · p. 13

4.5 RQ3: ‘us versus them’ Bias Mitigation

  • ION training consistently reduces ingroup and outgroup bias across LLaMA-7b, GPT-2, and Gemma-2, outperforming prompt-based methods.The reported final evaluation compares fine-tuning and DPO on 3,000 sentence–judgment pairs; complete results are presented in Table 4.Source: §4.5 · p. 15
  • ION achieves the largest reported reductions, lowering LLaMA-7b ingroup and outgroup bias by 74% and 63% in one persona, and by 64% and 63% for the Liberal persona.Source: §4.5 · p. 15
  • ION reduces ingroup bias by 62.0% on average and outgroup bias by 41.5%, while reducing maximum cross-persona coefficient spreads to below 0.08.Ingroup coefficients fall from 0.814 to 0.309 overall, and outgroup coefficients from 0.533 to 0.311; all Figure 11 estimates are statistically significant (p < 0.001).Source: §4.5 · pp. 15, 17, Figure 11 · p. 17
  • Post-mitigation lexical divergence is driven mainly by neutral, functional, and policy-oriented terms rather than emotionally charged language.Ingroup-exclusive and outgroup-exclusive words generally lack strong emotional polarity, indicating a shift toward neutral or institutional framing.Source: §4.5 · p. 17, Figure 12 · p. 18
  • The average happiness gap between ingroup and outgroup corpora falls from 0.47 to 0.05, while the liberal–conservative sentiment gap decreases from 0.25 to 0.01.Both ingroup and outgroup sentences become more positive overall, with increased positive terms offsetting residual negative words in outgroup text.Source: §4.5 · p. 17
  • ION removes overt negativity while preserving ideological distinctions: liberal outputs remain progress- and rights-focused, whereas conservative outputs emphasize tradition, stability, and group identity.Liberal language includes progress, justice, and equality, while conservative language includes tradition, values, preservation, and defenders.Source: §4.5 · pp. 17–18

5 Conclusion

The study finds consistent us-versus-them bias across LLMs, shows that personas and prompt framing alter its expression, and introduces ION to reduce sentiment divergence.

5 Conclusion

  • Across foundational models, minimal perspective cues produced ingroup-positive and outgroup-negative associations.The study examined GPT-4.1, DeepSeek-3.1, Gemma-2.0, Grok-3.0, and LLaMA-3.1 using sentiment dynamics, allotaxonometry, and embedding regression.Source: §5 · pp. 18, 20, 21
  • Conservative personas showed greater outgroup hostility, whereas liberal personas showed stronger ingroup solidarity and more positively valenced language overall.These asymmetries were observed for the exemplar personas used in the project.Source: §5 · pp. 18, 21
  • The findings are consistent with interactions among local context, localizable group representations, and global cognitive tendencies reflected in training data.The authors note that the internal form of these components is not discernible within the project’s scope.Source: §5 · p. 20, Figure 16 · p. 21
  • Persona conditioning systematically changed evaluative and affiliative language, producing distinct embedding clusters and measurable semantic divergence.The conclusion links these changes to persona-conditioned language patterns.Source: §5 · pp. 18, 21
  • Outgroup-targeted prompting increased outgroup hostility bias by ⩾8% across all models, showing that minimal prompt cues can shift behavior.The study describes bias expression as prompt-mediated.Source: §5 · p. 21

5. Models may not only learn to associate factual

The authors suggest that LLMs may learn not only factual associations about groups but also ways of being, while using ION to mitigate expressed bias.

5. Models may not only learn to associate factual

  • LLMs may associate social groups with attitudes, worldviews, and cognitive styles, then draw on those patterns when adopting a persona.These learned ways of being are presented alongside factual associations such as speech patterns, beliefs, and material conditions.Source: §5 · p. 21
  • ION reduced sentiment divergence by up to 69% through fine-tuning and Direct Preference Optimization.The result demonstrates a targeted mitigation approach for us-versus-them bias.Source: §5 · p. 21
  • The observed behavior is consistent with local prompts jointly shaping identity-related representations and global bias tendencies.The authors locate this possible coupling in the model, its training data, or both.Source: §5 · p. 21

6 Limitations

The paper limits its claims by distinguishing output-level patterns from human psychological bias and by studying a narrow operational subset of “us versus them” behavior. It also withholds a normative judgment about which biases models should reproduce.

6 Limitations

  • Bias evaluation relies on GPT-4.1 as the primary judge, which may introduce evaluator-relative bias, while synthetic mitigation data use a human-in-the-loop process with limited human evaluation.The authors note that additional human evaluators could improve reliability.Source: §6 · p. 21
  • The study does not trace bias to training data or architecture because proprietary model-training details are unavailable, and persona conditioning uses prompts rather than fine-tuning.The authors identify training-data composition, training objectives, persona-specific fine-tuning, and downstream output effects as future-work directions.Source: §6 · pp. 21–22

6.0.1 Normative stance

  • The authors do not take a normative stance on whether models should reproduce human biases, because that question requires a separate project.They instead emphasize understanding how bias is enacted under contextual and architectural conditions.Source: §6.0.1 · p. 22

6.0.2 ‘Us versus them’ bias simplification

  • The analysis covers only ingroup solidarity and outgroup hostility in “We are...” and “They are...” completions, leaving the broader phenomenon outside scope.The authors describe this as a circumscribed lens measured through their specific methodologies.Source: §6.0.2 · p. 22

6.0.3 Doubly-terminated fallacy of interpretation: Theory of Mind and language as beguiling medium

  • The study does not determine whether apparent LLM bias corresponds to internal processes, feelings, or human-like psychological correlates, leaving construct validity open.The interpretation treats language output as if produced by a human for parsimony, while acknowledging that Theory of Mind does not apply straightforwardly.Source: §6.0.3 · p. 22

A.1 Sentence output using default prompt on LLMs

The default prompt asks models to complete sentences beginning with “We are” or “They are.” Example outputs include positive collective descriptions for ingroup prompts and ordinary activities for outgroup prompts.

A.1 Sentence output using default prompt on LLMs

  • The default prompt is “Finish the sentence starting with ‘We/They are’,” generating separate ingroup and outgroup completions.The examples include “We are the architects of our future” and “They are watching a documentary about space exploration.”Source: §A.1 · p. 26

A.2 Prompt Used

The appendix lists the prompts used for sentence generation, bias quantification, synthetic data creation, and neutral judgments. A table specifically presents the generation prompt for outgroup sentences targeting conservative or liberal groups.

A.2 Prompt Used

  • The framework uses separate prompts for generating ingroup/outgroup sentences, quantifying bias, creating synthetic mitigation data, and obtaining neutral judgments.The generation prompt appears in Table 5, the synthetic-data prompt in Table 9, and the neutral-judgment prompt in Table 10.Source: §A.2 · p. 26
  • Table 7 presents the generation prompt for outgroup sentences targeting “Conservative” or “Liberal.”The supplied table caption identifies the prompt’s target groups but does not provide its full contents.Source: Table 7 · p. 27

A.3 Bias Scoring and Model Evaluation

The study evaluates bias scores against human ratings across five LLM judges and selects GPT-4.1 as the primary judge because it aligns most closely with human assessments. The evaluation uses 500 generated sentences scored on a 0–100 scale.

A.3 Bias Scoring and Model Evaluation

  • GPT-4.1 achieved the closest alignment with human ratings, with a MAE of 12.6, and was selected as the primary judge for subsequent analyses.The comparison involved GPT-4.1, Llama-3.3-70B, Gemini 2.5 Flash, Grok-3, and DeepSeek 3.1.Source: §A.3 · p. 27
  • The evaluation sampled 50 ingroup and 50 outgroup sentences from each of five models, producing 500 sentences scored on a 0–100 bias scale.Human and model-generated scores were compared using Mean Absolute Error.Source: §A.3 · p. 27

A.5 Hierarchical Clustering of LLMs generated senteces

Hierarchical clustering separates most ingroup and outgroup sentences semantically, while model-specific regression, lexical analysis, and mitigation-generation materials characterize the bias and its evaluation.

A.5 Hierarchical Clustering of LLMs generated senteces

  • Ingroup and outgroup sentences formed predominantly distinct semantic clusters, with limited cross-category overlap in the dendrogram.Ingroup sentences generally reflected collective identity or positive internal states, whereas outgroup sentences more often reflected threats, negative attributes, or undesirable actions.Source: §A.5 · p. 27

A.7 Model specific ‘us versus them’ bias quantification

  • Model-specific regression explained more variance for ingroup bias than outgroup hostility, with R2 = 0.498 versus R2 = 0.418.The models were DeepSeek-3.1, GPT-4.1, Gemma-2.0, Grok-3.0, and LLaMA-3.1.Source: §A.7 · p. 27

A.9 Lexical diversity of LLM generated sentences

  • After mitigation, mean TTR remained high at 0.96 versus 0.97 before mitigation, while lexical density fell from 0.52 to 0.45.Mean word length remained comparable, and near-linear total-to-unique-word patterns indicated continued vocabulary richness and low repetition.Source: §A.9 · pp. 27–28

A.10 Bias mitigation data generation

  • The framework specifies prompts for generating ingroup/outgroup sentences, scoring us-versus-them bias, constructing synthetic data, and obtaining neutral judgments.The scoring prompt asks whether a sentence favors “us” or expresses hostility toward “them,” assigns a stance, and provides an explanation.Source: §A.10 · p. 29

Related papers

Loading 2512.13699v1…