QueerGen: How LLMs Reflect Societal Norms on Gender and Sexuality in Sentence Completion Tasks

TL;DR

QueerGen presents a controlled prompting framework and dataset to study how LLMs reflect societal norms about gender and sexuality in sentence completions. By employing a tripartite identity setup (unmarked, queer-marked, non-queer-marked) and four evaluation metrics (sentiment, regard, toxicity, prediction diversity) across 14 models from 7 families, the study reveals that queer-marked prompts consistently lead to more negative sentiment and higher toxicity, with unmarked prompts often receiving more positive regard. The findings show that while autoregressive models can mitigate some biases, closed-access ARLMs can redistribute harms toward unmarked identities, indicating that alignment and model scale shift but do not eradicate representational harms. The work highlights the need for standardized, extensible evaluation frameworks to identify and mitigate normative biases in LLMs and informs responsible model development and policy considerations for LGBTQIA+ content. The methodology and dataset are openly shared to enable reproducibility and broader application beyond LGBTQIA+ contexts.

Abstract

This paper examines how Large Language Models (LLMs) reproduce societal norms, particularly heterocisnormativity, and how these norms translate into measurable biases in their text generations. We investigate whether explicit information about a subject's gender or sexuality influences LLM responses across three subject categories: queer-marked, non-queer-marked, and the normalized "unmarked" category. Representational imbalances are operationalized as measurable differences in English sentence completions across four dimensions: sentiment, regard, toxicity, and prediction diversity. Our findings show that Masked Language Models (MLMs) produce the least favorable sentiment, higher toxicity, and more negative regard for queer-marked subjects. Autoregressive Language Models (ARLMs) partially mitigate these patterns, while closed-access ARLMs tend to produce more harmful outputs for unmarked subjects. Results suggest that LLMs reproduce normative social assumptions, though the form and degree of bias depend strongly on specific model characteristics, which may redistribute, but not eliminate, representational harms.

Figures (11)

• Figure 1: The QueerGen dataset and evaluation framework: (1) generate a dataset using subjects, identity markers and template sentences; (2) input the dataset into several LLMs to generate completions; (3) evaluate the outputs for social biases.
• Figure 2: Comparison of average sentiment scores (measured with VADER) across subject categories and model groups.
• Figure 3: Comparison of average Regard scores across subject categories and model groups.
• Figure 4: Comparison of average toxicity scores (measured with Perspective API) across subject categories and model groups.
• Figure 5: Comparison of average prediction diversity percentages across subject categories and model groups.