A Multilingual, Culture-First Approach to Addressing Misgendering in LLM Applications

Sunayana Sitaram, Adrian de Wynter, Isobel McCrum, Qilong Gu, Si-Qing Chen

Abstract

Misgendering is the act of referring to someone by a gender that does not match their chosen identity. It marginalizes and undermines a person's sense of self, causing significant harm. English-based approaches have clear-cut approaches to avoiding misgendering, such as the use of the pronoun "they". However, other languages pose unique challenges due to both grammatical and cultural constructs. In this work we develop methodologies to assess and mitigate misgendering across 42 languages and dialects using a participatory-design approach to design effective and appropriate guardrails across all languages. We test these guardrails in a standard LLM-based application (meeting transcript summarization), where both the data generation and the annotation steps followed a human-in-the-loop approach. We find that the proposed guardrails are very effective in reducing misgendering rates across all languages in the summaries generated, and without incurring loss of quality. Our human-in-the-loop approach demonstrates a method to feasibly scale inclusive and responsible AI-based solutions across multiple languages and cultures. We release the guardrails and synthetic dataset encompassing 42 languages, along with human and LLM-judge evaluations, to encourage further research on this subject.

Introduction

Misgendering—referring to an individual by a gender that does not align with their identity—is a significant source of harm in AI systems, contributing to exclusionary user experiences. While English-centric solutions often rely on the singular "they" to maintain neutrality, this approach is insufficient in a global context. Effective mitigation requires a deep understanding of both linguistic structures and the sociopolitical nuances of language use.

The Challenge of Linguistic Diversity

The primary obstacle to universal misgendering mitigation is the vast diversity in how languages handle gender. Linguistic systems range from those with no gender distinctions (e.g., Turkish) to those with highly complex systems (e.g., Polish, which can have five or more gender categories). Furthermore, languages like Spanish and French feature rigid masculine/feminine inflections for nearly every noun and adjective, making neutral phrasing grammatically complex.

Global distribution of gender systems according to the World Atlas of Language Structures (WALS), highlighting the diversity of gender categories across languages.

Beyond grammar, sociopolitical factors play a critical role. For instance, in German, the use of a "gender star" (e.g., Bürgerinnen*) to denote inclusivity is a subject of active debate. LLMs often default to cultural stereotypes—such as assuming a professional role is masculine—or fail to respect explicitly provided preferred pronouns. Addressing these issues requires a scalable, culture-first framework.

Methodology: Participatory Design and Guardrails

To address these challenges, researchers employed a Participatory Design approach across 42 languages and dialects. The process involved two main stages:

  1. Mining and Refinement: Researchers synthesized existing inclusive writing style guides and polled native speakers (10 per language) to refine these rules. This ensured that the resulting "guardrails" respected local cultural norms and grammatical realities.
  2. The Guardrail Pipeline: These refined rules were converted into natural language instructions and injected into the LLM's system prompt.

To evaluate these guardrails, a new multilingual synthetic dataset for meeting transcript summarization was created, covering 42 languages including low-resource ones like Welsh and Latvian. The pipeline used GPT-4o to generate transcripts, which were then verified by both an LLM and human editors.

Key Results

The implementation of language-specific guardrails yielded significant improvements in model behavior:

  • Reduction in Gender Assumptions (GA): The rate of unwarranted gender assumptions (e.g., assuming a participant is male without evidence) dropped from 70% to 24%.
  • Reduction in Gender Mistakes (GM): Actual errors (e.g., using incorrect pronouns for a person with a known identity) dropped from 26% to 14%.
  • Quality Preservation: Contrary to fears that constraints might degrade text quality, adding guardrails increased human-perceived quality scores in 65% of the languages tested.

Discussion and Future Work

The study highlighted a significant "LLM-as-a-Judge" gap: while LLMs are useful for generation, they often fail to detect subtle misgendering in low-resource languages compared to human evaluators. Additionally, feedback from native speakers emphasized that misgendering is a social error as much as a grammatical one. In languages like Hebrew, neutral alternatives can sometimes impact the natural flow of reading, requiring careful balance.

Future research will focus on expanding this framework to more diverse language families, such as the Bantu family (e.g., Kiswahili), which utilizes complex noun class systems rather than binary gender. As social norms evolve, continuous feedback loops with native communities will remain essential to maintaining effective and respectful AI systems.

Summary updated

Abstract

Misgendering is the act of referring to someone by a gender that does not match their chosen identity. It marginalizes and undermines a person's sense of self, causing significant harm. English-based approaches have clear-cut approaches to avoiding misgendering, such as the use of the pronoun ``they''. However, other languages pose unique challenges due to both grammatical and cultural constructs. In this work we develop methodologies to assess and mitigate misgendering across 42 languages and dialects using a participatory-design approach to design effective and appropriate guardrails across all languages. We test these guardrails in a standard LLM-based application (meeting transcript summarization), where both the data generation and the annotation steps followed a human-in-the-loop approach. We find that the proposed guardrails are very effective in reducing misgendering rates across all languages in the summaries generated, and without incurring loss of quality. Our human-in-the-loop approach demonstrates a method to feasibly scale inclusive and responsible AI-based solutions across multiple languages and cultures. We release the guardrails and synthetic dataset encompassing 42 languages, along with human and LLM-judge evaluations, to encourage further research on this subject.
BibTeX
@inproceedings{sitaram2025multilingual,
  author = {Sunayana Sitaram and Adrian de Wynter and Isobel McCrum and Qilong Gu and Si-Qing Chen},
  title = {{A Multilingual, Culture-First Approach to Addressing Misgendering in LLM Applications}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2025},
}
Figure