Mitigating Bias in Queer Representation within Large Language Models: A Collaborative Agent Approach

LLMs systematically misuse gendered pronouns in ways that exclude queer individuals. The paper identifies that large language models frequently default to traditionally gendered pronouns ("he," "she") even when inclusive language would be more appropriate, effectively misrepresenting or erasing people with non-binary and other queer identities.

A multi-agent system can significantly improve inclusive pronoun usage. The researchers' collaborative agent pipeline—which uses specialized AI agents to detect and correct biased pronoun choices—achieved a 32.6 percentage point improvement over GPT-4o in correctly identifying when traditionally gendered pronouns should not be used (with statistical significance at p < 0.0001).

Current state-of-the-art models like GPT-4o still have substantial room for improvement on this issue. The baseline performance of GPT-4o on inclusive pronoun classification was low enough that a 32.6 percentage point improvement was possible, suggesting even leading commercial LLMs struggle with representing queer identities accurately.

Technical solutions can make AI systems more socially responsible. The paper demonstrates that thoughtfully designed frameworks—in this case, using multiple specialized agents working together—can be practical tools for reducing specific types of bias and promoting fairness in AI-generated content.