Technical Summary: Probing Social Identity Bias in Chinese LLMs with Gendered Pronouns and Social Groups

Problem Statement

Large Language Models (LLMs) are increasingly deployed in user-facing applications, raising concerns that they may reflect and amplify existing social biases. While prior research has extensively documented category-specific biases (e.g., gender, ethnicity) in English-centric LLMs using benchmarks like CrowS-Pairs or embedding-based analyses, there is a significant gap in understanding social identity biases within the Chinese linguistic context. Social identity theory posits that individuals exhibit favorable attitudes toward their "ingroup" and negative attitudes toward "outgroups." However, it remains unclear whether Chinese LLMs reproduce these dynamics and how specific linguistic features of Mandarin—particularly the distinction between the default plural pronoun 他们 (tāmen, used for mixed-gender or unspecified groups) and the explicitly feminine plural 她们 (tāmen)—modulate these biases. Current evaluation strategies often overlook the relational mechanism of intergroup sentiment and the unique morphological structures of Chinese.

Methodology

The study employs a controlled generation framework across ten representative Chinese LLMs, covering both base (pretrained) and instruction-tuned variants from major developers (e.g., Alibaba, Baidu, Tencent, DeepSeek).

1. Prompt Design and Data Collection

• Framing Conditions: The authors constructed sentence-completion prompts using two primary framings:
  • Ingroup: First-person plural ("我们", "We").
  • Outgroup: Third-person plural, further split into:
    • Default Plural: "他们" (unmarked/mixed-gender).
    • Feminine Plural: "她们" (explicitly marked feminine).
• Social Groups: Prompts were instantiated with 240 social groups salient in the Chinese context (e.g., age, disability, nationality, education level) using the formulation "As [Group], we/they are..."
• Mitigating Refusals: To ensure stable generation from instruction-tuned models, a neutral-context scaffolding strategy was used, prepending high-quality, neutral sentences from the ChineseWebText corpus to the prompts.
• Dataset: The study generated 297,600 responses from the ten models. Additionally, a supplementary exploratory analysis was conducted on 4,079 user-assistant interactions from the WildChat corpus to assess real-world validity.

2. Measurement Framework

A two-tiered measurement framework was adopted to assess both sentiment and toxicity:

• Sentiment Analysis: Three classifiers (Aliyun, Baidu NLP, and DeepSeek-V3) were used to label responses as positive, negative, or neutral. Consensus labels were derived via majority voting, with manual validation on a 200-sentence subset showing 83.5% agreement.
• Toxicity Analysis: The Perspective API was used to generate continuous toxicity scores (0–1), capturing safety-relevant signals (rude, disrespectful, or unreasonable comments) that sentiment analysis alone might miss.
• Statistical Modeling: Regression models (logistic for binary sentiment, linear for toxicity) were estimated to test for ingroup solidarity and outgroup hostility, controlling for sentence length and lexical diversity (Type-Token Ratio).

Key Results

1. Systematic Ingroup–Outgroup Asymmetries

Across nearly all models, the study found systematic asymmetries:

• Ingroup Solidarity: Responses under "We" framings were generally more positive than under "They" framings.
• Outgroup Hostility: Responses under "They" framings exhibited higher negative sentiment and toxicity.
• Model Variance: Pretrained models displayed particularly strong outgroup hostility. Instruction-tuned models showed more balanced patterns, with sentiment asymmetries often attenuated, though toxicity gaps remained more persistent.

2. The Impact of Gendered Linguistic Marking

The distinction between 他们 and 她们 revealed specific bias patterns:

• Feminine-Linked Toxicity: In several models, the explicitly feminine plural (她们) was associated with higher toxicity scores compared to the default plural (他们), even when sentiment asymmetries were modest.
• Instruction Tuning Effects: While instruction tuning reduced sentiment disparities, the toxicity gap associated with the feminine-marked form remained visible in some systems, suggesting that safety alignment does not fully eliminate gendered bias.

3. Variation Across Social Categories

Bias expression was not uniform across the 240 social groups:

• High Sensitivity: Groups related to Gender, Nationality, and Race showed the most pronounced ingroup–outgroup asymmetries (odds ratios > 1.4).
• Dampened Effects: Categories like Ethnicity and Level of Education exhibited dampened or reversed effects, suggesting selective safety alignment mechanisms that temper negative outputs for certain sensitive categories while leaving others susceptible.

4. Real-World Dialogue Evidence

The supplementary analysis of WildChat dialogues provided exploratory evidence that similar ingroup–outgroup patterns exist in naturalistic interactions, although the authors caution that these findings are not directly comparable to the controlled experiments due to the use of English-centric models in the WildChat corpus and unbalanced identity markers.

Key Contributions

1. Language-Aware Evaluation Framework: The study introduces a methodology specifically designed for Chinese LLMs that leverages the morphological distinction between default and feminine plural pronouns to probe bias, a dimension not directly observable in English-only settings.
2. Empirical Evidence in Chinese Context: It confirms that social identity biases (ingroup favoritism and outgroup hostility) previously documented in English also manifest in Chinese LLMs across a broad spectrum of 240 social groups.
3. Dual-Metric Insight: By combining sentiment and toxicity analysis, the paper reveals that instruction tuning may reduce positive/negative sentiment asymmetries but often fails to eliminate toxicity gaps, particularly those linked to gendered linguistic markers.
4. Supplementary Real-World Analysis: It provides preliminary evidence that these synthetic benchmark findings may extend to deployment-like settings, highlighting the need for culturally grounded alignment strategies.

Significance and Claims

The paper claims that social identity bias in Chinese LLMs is a systemic issue that extends beyond specific categories to a general relational dynamic between "us" and "them." Crucially, it argues that Mandarin-specific linguistic structures (specifically gendered pronouns) can reveal bias patterns that are invisible in English-centric evaluations.

The authors assert that while instruction tuning can partially mitigate these biases, it does not eliminate them, particularly regarding toxicity and gendered framing. This suggests that current alignment strategies may be insufficient for ensuring fairness in interactive, user-facing applications. The study calls for systematic monitoring of LLM behavior in high-stakes domains and the development of mitigation strategies that are sensitive to both the cultural context and the specific linguistic properties of the Chinese language.

The authors maintain a modest tone regarding their claims, acknowledging limitations such as the exclusion of some models due to computational constraints, the reliance on automated sentiment proxies rather than full human gold standards, and the exploratory nature of the real-world dialogue analysis. They emphasize that their findings document model behaviors rather than reflecting the views of the authors or accurate representations of the social groups involved.