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

Misgendering is a widespread problem in AI systems that extends beyond English: While English has established solutions like singular "they," the paper identifies that 42 different languages and dialects face unique grammatical and cultural barriers to preventing misgendering, suggesting this is a global AI fairness challenge that requires localized solutions rather than one-size-fits-all approaches.

Participatory design with native speakers is essential for culturally appropriate AI safeguards: The researchers used a human-in-the-loop approach involving speakers from different language communities to design guardrails, demonstrating that effective responsible AI requires input from people within those cultures rather than external experts alone.

Effective misgendering solutions don't require sacrificing AI quality: The paper shows that their proposed guardrails significantly reduced misgendering rates across all tested languages in a real-world application (meeting transcription summaries) without degrading the overall quality of the AI output, suggesting that inclusivity and performance are not competing priorities.

Scaling responsible AI across languages requires dedicated resources and transparency: By releasing their guardrails, synthetic dataset, and evaluation results across 42 languages publicly, the authors demonstrate a commitment to making multilingual responsible AI feasible for other organizations and researchers, rather than keeping solutions proprietary.