Abstract:

Generative Artificial Intelligence has become a widely adopted tool for producing visual content in higher education environments, supporting communication strategies, student engagement initiatives, and institutional design workflows. However, concerns persist regarding bias in image generation, particularly in scenarios where diversity, inclusion, and accurate cultural representation are essential. This study investigates how different text-to-image systems portray university life when prompted with neutral and explicitly inclusive descriptions. Our goal is to evaluate whether these models provide visual outputs aligned with the representational needs of Latin American public universities and to identify design implications for educational communication teams. We conducted a study consisting of three phases. First, a design intervention established a structured prompt protocol, distinguishing neutral prompts from inclusion-oriented prompts that explicitly referenced visible dimensions of diversity such as race, gender expression, body type, and cultural appearance. Second, a generation and collection phase produced images across four widely used generative systems. Each output was stored with metadata including model name, prompt, seed, and configuration parameters. Third, we evaluated the resulting dataset using a mixed-methods approach. Quantitatively, two independent raters applied a diversity rubric assessing racial variation, gender representation, cultural attributes, body-diversity, and contextual adequacy for university communication. Qualitatively, a content analysis examined recurring patterns, stereotypical representations, omissions, and cross-model differences. Results indicate a persistent tension between aesthetic quality and representational diversity. While inclusive prompts improved demographic variation across all models, several systems still tended to idealize body types, homogenize cultural features, and reproduce gendered stereotypes. Models also differed in their sensitivity to Latin American context, with some generating imagery more aligned with North American or generic global aesthetics. The study provides evidence that, although generative AI can support communication teams, it cannot yet replace human-guided design workflows, especially in institutional contexts that require cultural accuracy and equitable representation. The main contribution of this work is a replicable evaluation framework that integrates prompt design, structured diversity assessment, and qualitative interpretation. A real university communication project served as a contextualized case to demonstrate practical applicability, but the framework is adaptable to other educational settings.

Keywords:

Generative AI, Bias, Diversity, Inclusive Design.