ilana27/llm-nationality-bias-us-narratives

Representational Harms in US-Based LLM Narratives: Nationality Bias Dataset

Dataset Summary

This dataset contains 9,710 LLM-generated narratives that reference non-US national identities, extracted from a larger corpus of 500,000 stories generated by GPT-3.5, GPT-4, Claude 2.0, Llama 2, and PaLM 2. The stories were generated in response to open-ended prompts set in US contexts across three domains: Learning, Labor, and Love. This dataset was created to study representational harms and nationality bias in LLM-generated narratives, specifically examining how Global Majority and Global Minority identities are portrayed in power-neutral versus power-laden scenarios.

Dataset Details

Source

Extends the dataset from Shieh et al. (2024) "Laissez-Faire Harms: Algorithmic Biases in Generative Language Models"

Languages

• English (en)

Data Instances

Each instance contains:

• story_text: The generated narrative (≤100 words)
• prompt_template: The story prompt used
• power_condition: "neutral" or "laden" (dominant/subordinated character structure)
• domain: "Learning", "Labor", or "Love"
• character_role_subject: Role of the dominant/main character
• character_role_object: Role of the subordinated character (for power-laden prompts)
• model: LLM used for generation (gpt-3.5-turbo, gpt-4, claude-2.0, llama-2-70b, palm-2)
• nationality_subject: Extracted country/demonym references for dominant character
• nationality_object: Extracted country/demonym references for subordinated character
• global_majority_minority: Classification of mentioned countries as Global Majority (G77 member) or Global Minority

Dataset Statistics

• Total narratives: 9,710 stories with character-level nationality references
• Power-neutral stories: 210 (2.2%)
• Power-laden stories: 9,500 (97.8%)
  • Dominant character nationality references: 145 (1.5%)
  • Subordinated character nationality references: 8,919 (93.9%)
• Countries represented: 40+ nations
• Models: 5 (GPT-3.5, GPT-4, Claude 2.0, Llama 2, PaLM 2)
• Prompts: 50 scenarios (15 Learning, 15 Labor, 20 Love)

Intended Uses

Primary Uses

• Studying representational harms in LLM-generated narratives
• Analyzing nationality and national origin bias in AI systems
• Understanding power dynamics in character portrayal across Global Majority/Minority identities
• Evaluating intersectional biases in open-ended text generation
• Developing bias mitigation strategies for narrative generation systems

Out-of-Scope Uses

• Training generative models (contains biased content)
• Direct deployment in production systems without bias mitigation
• Drawing conclusions about real populations or nationalities

Dataset Construction

Source Data Collection

Stories were generated using temperature=1.0 with 1,000 samples per prompt across 50 scenarios. Prompts were co-developed with minoritized students with immigrant backgrounds in the United States.

Annotation Process

Nationality references were extracted using a fine-tuned GPT-4.1 Mini QA model to identify:

1. Country names and demonyms in narratives
2. Character associations (subject vs. object roles)
3. Distinction between character identity references and other mentions (e.g., food, locations)

Quality Control

• Manual validation on 611 hand-labeled examples
• Precision: 0.977 (subject), 0.982 (object)
• Recall: 1.000 (subject), 0.982 (object)

Considerations for Using the Data

Limitations

• US-centric prompt contexts only
• English language only
• Limited to UN-recognized nations (excludes territories, Indigenous nations, stateless populations)
• Excludes city-based geographic identity cues
• May undercount nationality bias (only explicit country/demonym references)
• Contains harmful stereotypes and representational harms

Ethical Considerations

This dataset documents harmful content including:

• Stereotyping of Global Majority nationalities
• Subordination and one-dimensional portrayals
• Representational erasure and omission
• Orientalist and colonial narrative patterns

Users should:

• Approach this data with critical awareness of documented harms
• Not use for training without explicit bias mitigation
• Consider psychosocial impacts of exposure to biased narratives
• Center Global Majority perspectives in analysis

Citation

BibTeX:

@inproceedings{nguyen2026representational,
  title={Representational Harms in LLM-Generated Narratives Against Global Majority Nationalities},
  author={Nguyen, Ilana and Suresh, Harini and Monroe-White, Thema and Shieh, Evan},
  booktitle={ACM Conference on Fairness, Accountability, and Transparency},
  year={2026}
}

APA:

Nguyen, I., Suresh, H., Monroe-White, T., & Shieh, E. (2026). Representational harms in LLM-generated narratives against global majority nationalities. In Proceedings of the 2026 ACM Conference on Fairness, Accountability, and Transparency (FAccT '26). Association for Computing Machinery. [Manuscript accepted for publication]

Dataset Card Contact

Ilana Nguyen (ilananguyen@brown.edu)

Additional Resources

• Study 2 Dataset: https://huggingface.co/datasets/ilana27/llm-nationality-bias-global-narratives
• Code: https://github.com/ilana27/representational-harms-nationality