Artificial Judgment: AI Models and the Evaluation of LGBTQ Candidates

This study examines whether issue domain moderates evaluative penalties faced by LGBTQ political candidates in AI systems. Drawing on theories of stereotype congruity, leadership schemas, and political viability heuristics, it examines candidate evaluations across four policy domains: economy, education, policing/crime, and national security. Using 240 AI-generated evaluations of fictional candidates from six large language models, regression models account for candidate identity, issue domain, and AI model differences. Results reveal an LGBTQ electability penalty, including a 0.48-point disadvantage in the economy reference condition. Interactions between LGBTQ identity and issue domain were not statistically significant, providing limited evidence that penalties varied systematically across policy contexts. LGBTQ identity was not significantly associated with competence or leadership evaluation penalties. These findings suggest that AI systems may distinguish between candidate capability and electoral viability, raising questions about how algorithmic evaluations may reinforce inequalities in political representation.