Argumentative Mechanisms in the Generative Discourse of Artificial Intelligence: An Analytical Study of ChatGPT Responses to Controversial Issues

Abstract

This study aims to uncover the argumentative mechanisms used by the generative AI model ChatGPT-4 when dealing with controversial issues. It starts from a central problem: how does the programmed neutrality in large language models shift from being a technical feature to a systematic rhetorical strategy aimed at avoidance rather than persuasion? The research used a descriptive approach, based on a purposive sample of 80 texts generated by ChatGPT-4, equally divided across four areas: religious, political, social, and historical. The data was analyzed using Chaïm Perelman's theory of argumentation, leading to the proposal of a theoretical model called by the researcher the "Dual Argumentative Structure Model".

The study reached several key findings, including:

 1) that 92.5% of the texts rely on quantitative balancing as the main strategy.

 2) that the model's discourse is based on two layers: a solid core that is fixed at 70% and serves to protect, and a variable contextual mask at 30% which serves for camouflage.

3) that the historical domain is the most avoided, with an average of 4.60 strategies per text. The study also concluded that ChatGPT-4 does not engage in argumentation in the Peircean sense, but rather practices argumentative siege aimed at freezing the recipient's judgment. The study recommends training students in the argumentative deconstruction of AI outputs within critical thinking curricula.