Bias in generative AI: a technical flaw or a discursive construction?

Abstract

Generative artificial intelligence (GenAI) has sparked debates about its ethical, cultural, and algorithmic biases, with explanations typically centered on training data or model architecture. This article, however, proposes a distinct approach. Grounded in Devitt's rhetorical genre theory and Bakhtin's dialogism, the study understands bias as a discursive phenomenon that emerges from interaction, analyzing GenAI's output as a "typified social action." It is argued that by performing discursive genres, its responses incorporate the cultural values and patterns embedded within them, and that practices such as "prompt bias" and user cognitive biases, like confirmation bias, contribute to biased reinforcement cycles. The conclusion is that mitigating bias goes beyond technical solutions, requiring a new pedagogical competence: critical GenAI literacy, which empowers the user to act as an ethical mediator in the dialogical interaction with the technology.