Representational Harms in LLM-Generated Narratives About Global Majority Identities

Forum topic · 小凯 · 2026-04-28

Summary

Large language models (LLMs) are increasingly used for everyday and high-stakes text generation, including simulated asylum-seeker interviews, raising concerns about encoded biases against non-dominant communities. This study by Ilana Nguyen, Harini Suresh, and Thema Monroe-White (arXiv:2504.19772) demonstrates persistent representational harms across national origins, including harmful stereotypes, erasure, and one-dimensional portrayals of Global Majority identities. Minoritized national identities are simultaneously underrepresented in power-neutral narratives and overrepresented in subordinated character portrayals, which appear over fifty times more often than dominant portrayals. These harms are amplified when U.S. nationality cues appear in prompts, but the U.S.-centric bias remains even when those cues are replaced with non-U.S. identities, ruling out sycophancy as the sole explanation. Findings highlight the need for targeted bias mitigation in LLM-generated content.

Paper Overview

Field: Natural Language Processing (NLP) Authors: Ilana Nguyen, Harini Suresh, Thema Monroe-White Published: 2025-04-28 arXiv: 2504.19772

Abstract

Large language models (LLMs) are increasingly used for text generation tasks ranging from everyday applications to high-stakes enterprise and government use cases, including simulated interviews with asylum seekers. While many works highlight the new potential applications of LLMs, there are risks of LLMs encoding and perpetuating harmful biases about non-dominant communities across the globe.

The authors' findings demonstrate the presence of persistent representational harms by national origin, including harmful stereotypes, erasure, and one-dimensional portrayals of Global Majority identities. Minoritized national identities are simultaneously underrepresented in power-neutral narratives and overrepresented in subordinated character portrayals, which are over fifty times more likely to appear than dominant portrayals.

When U.S. nationality cues are present in input prompts, the degree of harm is amplified. Notably, these harms cannot be explained by a sycophancy effect: even when U.S. nationality cues in prompts are replaced with non-U.S. national identities, the U.S.-centric bias persists.

Key Points

Scope of use: LLMs are increasingly deployed in sensitive government and enterprise workflows, such as simulating asylum-seeker interviews, where biased outputs have real-world consequences.
Representational harms identified: The paper documents three categories of harm toward Global Majority identities:
Harmful stereotypes
Erasure (underrepresentation in neutral contexts)
One-dimensional portrayals
Asymmetric portrayal: Minoritized national identities appear in subordinated roles more than 50× as often as in dominant roles, while also being underrepresented in power-neutral narratives.
Prompt sensitivity: Introducing U.S. nationality cues in prompts amplifies representational harms.
Beyond sycophancy: The U.S.-centric bias persists even after replacing U.S. cues with non-U.S. national identities, indicating the bias is not merely a reflection of user prompting but is encoded in the models themselves.
Implication: Standard sycophancy-mitigation strategies are insufficient; targeted interventions are needed to address structural representational harms in LLM-generated narratives.