(iPoster) Essentially Contested Concepts in Generative AI: The Politics of Visibility

• In Event: Poster Sessions: Information Technology and Politics & Political Networks
  In Poster Session: Poster Session: Information Technology and Politics

Abstract

The emergence of generative artificial intelligence (AI) signifies a profound transformation in the interplay between information technology and political power. While traditional media technologies regulate information dissemination through gatekeeping, generative AI autonomously generates information via vector space computation, altering political communication dynamics. Gillespie (2024) conceptualizes this as the "politics of visibility," positing that generative AI's influence extends beyond determining what remains unseen to shaping what and how information becomes visible. This insight elucidates the shift of information technology from a mere instrument of politics to the arena in which political processes unfold. This study utilizes "essentially contested concepts" (Gallie, 1956) as a framework for analyzing the intersection of politics and technology. Concepts such as democracy, freedom, human rights, and justice are characterized by inherent contestability, with their meanings being the subject of political struggle and lacking objective definitions. When generative AI systems are tasked with explaining or visualizing these concepts, they cannot maintain neutrality and must present a specific version of political representation. This study poses the following research questions: How do generative AI systems from the United States and China represent essentially contested concepts through text and images? How do these differences reflect ideological encoding and constitute a site of confrontation in the politics of visibility?

This study employs Lukes's (2005) three-dimensional view of power as an analytical framework. The first dimension examines observable decision-making, analyzing whether AI systems respond to queries or refuse to generate content. Such refusals delineate the political boundary between what is permissible and impermissible within technological systems. The second dimension addresses agenda-setting and frame-shaping, scrutinizing definitional frameworks, narrative structures, and visual symbols adopted by AI systems, as well as which political aspects are emphasized or hidden. This dimension reveals how technological systems engage in selective political representation beneath their responses. The third dimension explores deep cognitive shaping, investigating whether contestation is transformed into objective knowledge after users repeatedly receive specific versions of political concepts and whether ideological encoding becomes internalized through technological interfaces.

The research design utilizes a cross-polity comparative experiment, selecting four models exemplifying US-China geopolitical opposition: ChatGPT and Gemini from the United States, and DeepSeek and ERNIE Bot from China. The design principle is "identical input, comparative output," wherein all models receive the same essentially contested concepts as prompts. The experiment encompasses two languages (English and Simplified Chinese) and two modalities (text and image generation) to investigate cross-linguistic and cross-modal differences. The procedure comprises five stages: concept selection according to Gallie's (1956) criteria covering political systems, values, actions, and national symbols; prompt design using uniform formats; data collection through API queries with multiple generations; coding analysis focusing on definitional frameworks and visual symbols; and matrix construction integrating results into a "representational confrontation matrix" revealing how AI systems translate identical political vocabulary into divergent cognitive schemas.

This study anticipates three contributions. First, it introduces essentially contested concepts into research on information technology and politics, offering a framework for analyzing ideological encoding in AI. Second, it incorporates multimodal analysis of text and images, addressing the gap in research focusing on single modalities. Third, it applies three-dimensional view of power to elucidate how the politics of visibility operates across refusal, shaping, and internalization. Practically, this study provides a basis for AI literacy education and algorithmic accountability, enabling users to discern that AI engages in selective political representation.

Keywords: generative artificial intelligence, essentially contested concepts, politics of visibility, three-dimensional view of power, ideological encoding

References
Gallie, W. B. (1956). Essentially contested concepts. Proceedings of the Aristotelian Society, 56, 167-198. http://www.jstor.org/stable/4544562
Gillespie, T. (2024). Generative AI and the politics of visibility. Big Data & Society, 11(2). https://doi.org/10.1177/20539517241252131
Lukes, S. (2005). Power: A radical view (2nd ed.). Palgrave Macmillan.

Author

• Hungyen Hsu, National Chengchi University