Gendered representations of occupations in large language model generated images and texts

As large language models (LLMs) increasingly mediate everyday communication, their potential to reproduce gender inequality requires rigorous examination. The study argues that gender inequality in LLM outputs is not merely an algorithmic bias, but a socially sedimented representational phenomenon shaped by training data, model design, user interaction, and broader cultural assumptions about gender and work. Grounded in Sociocognitive Linguistics and Social Semiotics, this study examines how widely used LLM-based generative systems, ChatGPT-4o, Ernie Bot 4.0 Turbo, Gemini Advanced, and Qwen 2.5, construct occupational gender through images and texts. The findings reveal that gender inequality is expressed through systematic asymmetries in visual composition and lexical choice. These patterns are repetitive across models and modalities, socially grounded in inherited occupational schemas, and often concealed beneath surface-level neutrality.