An Analysis of Gender Bias in Text-to-Image Models Using Neutral Prompts in Spanish

• Autores: Victoria Muñoz García, María Villalba Oses, Juan Pablo Consuegra Ayala
• Localización: Procesamiento del lenguaje natural, ISSN 1135-5948, Nº. 75, 2025 (Ejemplar dedicado a: Procesamiento del Lenguaje Natural, Revista nº 75, septiembre de 2025), págs. 147-160
• Idioma: inglés
• Títulos paralelos:
  • Análisis de Sesgo de Género en Modelos de Texto a Imagen mediante Prompts Neutros en Español
• Enlaces
  • Texto completo
• Resumen
  • español

    Los modelos generativos de texto a imagen pueden crear contenido visual a partir de descripciones textuales, pero suelen perpetuar sesgos de sus datos de entrenamiento. Este estudio analiza el sesgo de género en tres modelos—ChatGPT (DALL-E), Copilot y Gemini—usando indicaciones neutras en español, un idioma poco explorado en esta área. Se generaron 300 imágenes a partir de 50 prompts sobre salud y bienestar, evaluadas manualmente para identificar sesgos. Los resultados muestran que ChatGPT presentó más salidas estereotípicas y menor neutralidad, Copilot mantuvo estricta neutralidad y Gemini tuvo un comportamiento intermedio. En todos los modelos, la neutralidad fue menor en el sujeto principal de la imagen que en los elementos contextuales. Estos hallazgos resaltan la persistencia del sesgo de género y la necesidad de sistemas de IA más equitativos mediante evaluaciones sistemáticas.

  • English

    Text-to-image generative models can create visual content from text but often reflect biases in their training data. This study examines gender bias in three widely used models—ChatGPT (DALL-E), Copilot, and Gemini—using gender-neutral prompts in Spanish, an underexplored language in bias research. A dataset of 300 images from 50 neutral prompts on health and well-being was manually analyzed for gender representation biases. ChatGPT showed the highest stereotyping and lowest neutrality, Copilot maintained strict neutrality, and Gemini exhibited intermediate behavior. Across models, neutrality dropped when analyzing the main subject (gender-target annotations) versus contextual elements (genderrelated annotations). These findings underscore persistent gender bias, even with neutral prompts, and highlight the need for fairer AI systems through systematic evaluation.

• Referencias bibliográficas
  • Bansal, H., D. Yin, M. Monajatipoor, and K.-W. Chang. 2022. How well can text-to-image generative models understand ethical natural language...
  • Bianchi, F., P. Kalluri, E. Durmus, F. Ladhak, M. Cheng, D. Nozza, T. Hashimoto, D. Jurafsky, J. Zou, and A. Caliskan. 2023. Easily accessible...
  • Birhane, A., V. U. Prabhu, and E. Kahembwe. 2021. Multimodal datasets: misogyny, pornography, and malignant stereotypes.
  • Burns, K., L. A. Hendricks, K. Saenko, T. Darrell, and A. Rohrbach. 2019. Women also snowboard: Overcoming bias in captioning models. arXiv...
  • Cho, J., A. Zala, and M. Bansal. 2023. Dalleval: Probing the reasoning skills and social biases of text-to-image generation models. Proceedings...
  • Consuegra-Ayala, J. P., Y. Gutiérrez, Y. Almeida-Cruz, and M. Palomar. 2025. Bias mitigation for fair automation of classification tasks....
  • Consuegra-Ayala, J. P., I. Martínez-Murillo, E. Lloret, P. Moreda, and M. Palomar. 2024. A multifaceted approach to detect gender biases in...
  • Friedrich, F., K. Hämmerl, P. Schramowski, M. Brack, J. Libovicky, K. Kersting, and A. Fraser. 2024. Multilingual text-to-image generation...
  • Garg, N., L. Schiebinger, D. Jurafsky, and J. Zou. 2018. Word embeddings quantify 100 years of gender and ethnic stereotypes. Proceedings...
  • Mannering, H. 2023. Analysing gender bias in text-to-image models using object detection. arXiv preprint arXiv:2307.08025.
  • Muñoz-García, V. 2024a. Bias mitigation in corpora for llms training applied to text simplification.
  • Muñoz-García, V. 2024b. Open-source terminology: A gender-based perspective in menopause studies.
  • Naik, R. and B. Nushi. 2023. Social biases through the text-to-image generation lens. In Proceedings of the 2023 AAAI/ACM Conference on AI,...
  • Navigli, R., S. Conia, and B. Ross. 2023. Biases in large language models: origins, inventory, and discussion. ACM Journal of Data and Information...
  • Ramesh, A., P. Dhariwal, A. Nichol, C. Chu, and M. Chen. 2022. Hierarchical textconditional image generation with clip latents. arXiv preprint...
  • Rombach, R., A. Blattmann, D. Lorenz, P. Esser, and B. Ommer. 2022. Highresolution image synthesis with latent diffusion models. Proceedings...
  • Ross, C., B. Katz, and A. Barbu. 2023. Measuring social biases in grounded vision and language embeddings. arXiv preprint arXiv:2002.08911.
  • Ventura, M., E. Ben-David, A. Korhonen, and R. Reichart. 2025. Navigating cultural chasms: Exploring and unlocking the cultural pov of text-to-image...
  • Villalba, M. 2024. Artificial intelligence and natural language processing applied to design.
  • Villalba, M. and M. Palomar. 2024. A review of ai application trends in industrial design.
  • Wang, J., Y. Liu, and X. E. Wang. 2021. Are gender-neutral queries really genderneutral? mitigating gender bias in image search. arXiv preprint...
  • Weidinger, L., J. Mellor, M. Rauh, C. Griffin, J. Uesato, P.-S. Huang, M. Cheng, M. Glaese, B. Balle, A. Kasirzadeh, Z. Kenton, S. Brown,...
  • Wolfe, R. and A. Caliskan. 2022. American == white in multimodal language-and-image ai. In Proceedings of the 2022 AAAI/ACM Conference...
  • Wu, Y., Y. Nakashima, and N. Garcia. 2023. Stable diffusion exposed: Gender bias from prompt to image. arXiv preprint arXiv:2312.03027.
  • Zhang, Y., L. Jiang, G. Turk, and D. Yang. 2024. Auditing gender presentation differences in text-to-image models. In Proceedings of the 4th...