Algorithmic Semiosis and Racial Bias: A Study of Images Created by Generative AI

Abstract

Objective: To investigate the trirelational relationship among object, sign, and interpretant in the functioning of generative Artificial Intelligence (AI) tools for image production, with an emphasis on racial bias. Method: Exploratory and quali-quantitative research, which employed a semiotic and critical content approach. Data collection was carried out in four stages: 1) selection of 10 tools; 2) formulation of eight textual prompts in English; 3) generation and storage of 155 images; 4) categorization, analysis of these images, and selection of 47 of them to demonstrate the observed patterns and markers. Results: Predominance of a specific ethnic and social group in the generated images, with an absence of diversity markers. When using the generic prompts: 'a man' and 'a woman,' 90.9% of the images of men and 92% of the images of women portrayed white, upper-middle-class individuals. When using more specific prompts: 'a black man' and 'a black woman,' the images often replicated stereotypes and characteristics that reinforce racial and class prejudices. Conclusions: The generative AI tools analyzed are part of a new cycle of visual reality production that reflects, reproduces, and amplifies existing raciality devices. The technical images generated by AI reflect power relations, as well as markers of whiteness and racism, highlighting how assistive technology intertwines with social and cultural representations in its semiotic action. The study helped denaturalize algorithmic semioses by demonstrating how the functioning of generative AIs reveals ethical and social implications that are guided by perceptions of race and otherness, shaped by hierarchies that contribute to the creation of control images.