An Empirical Exploration of Socioeconomic Linguistic Bias in Text-to-Image Generation
Published in Student Research Workshop (SRW), EACL 2026, 2026
Co-authored “An Empirical Exploration of Socioeconomic Linguistic Bias in Text-to-Image Generation,” presented as a long paper and poster at the Student Research Workshop, EACL 2026.
The study quantitatively analyzes whether income-related wording in prompts shifts the skin tone of faces generated by text-to-image models. Results show that GPT-5 Image-mini, Gemini 2.5 Flash-Image, and Grok-2 Image tend to produce lighter skin tones for high-income occupations, evidencing measurable socioeconomic bias in multimodal generative AI.
Recommended citation: R. G. Maurya, V. Shukla, S. Panat (2026). "An Empirical Exploration of Socioeconomic Linguistic Bias in Text-to-Image Generation." Student Research Workshop, EACL 2026.
Download Paper