(iPoster) Rethinking AI Risk to Democracy: From Catastrophic Harms to Epistemic Threats

• In Event: Poster Sessions: International Security II & Normative Political Theory
  In Poster Session: Poster Session: Normative Political Theory

Abstract

The potential risks of AI to democratic society have drawn the attention of legislators worldwide. For example, the EU’s recent Digital Services Act (DSA) and AI Act both target the regulation of systemic risks that AI or AI-enabled platforms may pose. Moreover, California’s SB 53 addresses concrete catastrophic risks, such as loss of life or substantial economic harm. While existing definitions of risk attempt to regulate AI by focusing either on discrete catastrophic harms or on systemic risks, they lack a coherent theoretical foundation for addressing a more fundamental question: what risks does AI pose to democracy itself.

I attempt to provide a theoretical foundation for the risk of AI by engaging with democratic theory. I argue that negative impacts on democracy can be categorized into two forms: one concerning short-term operational outcomes within democratic systems, and the other affecting the long-term foundational elements of democratic systems. Examples of the former include concerns about manipulating people for specific election results or undermining electoral processes, as well as the dissemination of specific illegal content. Disasters causing massive economic losses and loss of life also fall within this category. This type of risk means a one-time or short-term large-scale failure of the democratic system.

However, there exists a long-term risk, encompassing how individuals acquire knowledge and form attitudes toward democracy. Drawing from recent discussions in political epistemology, I argue that “knowledge about and trust in democratic systems” constitute a different systemic risk. For instance, individual AI hallucinations may only generate isolated instances of misinformation or manipulated action, but this differs substantially from the harm caused when AI systems consistently (whether deliberately or inadvertently) exclude certain topics or guide users toward ideological positions over extended periods, which might lead people to distrust the knowledge or institutions necessary for democratic functioning (such as distrusting existing media). Moreover, as people increasingly acquire knowledge through foundation models, these models will influence citizens’ capacity—whether people can deliberate and judge the veracity of information—which democracy requires. From this perspective, some AI models, which comprehensively assist humans across various domains and shape how they perceive and interact with others, may pose this type of risk.
This theoretical foundation can contribute to both political theory and legal frameworks addressing AI risk. First, it applies and critically reflects upon key concepts in political theory by examining their significance in the age of artificial intelligence: which democratic values are challenged by AI, and how should these challenges be addressed? Second, it provides a normative basis for developing more comprehensive regulatory approaches to AI risk. Focusing solely on isolated catastrophic harms or short-term systemic risks is insufficient; instead, we must more fully assess AI’s broader implications for democratic ideals and social structures.

By distinguishing these two potential risks, I provide recommendations for modifying existing legal frameworks. For instance, SB 53 emphasizes the concrete, event-specific negative impacts while only partially addressing the first category of risk identified in this study. Second, the Digital Services Act requires platforms to conduct self-analysis primarily focused on the first type of short-term failures that platforms may generate—such as illegal speech, electoral impacts, or violations of fundamental rights—while according less attention to the second category of risk. By contrast, the AI Act does not explicitly specify which type of risk it addresses; instead, it employs computational thresholds to determine the presence of systemic risk. I argue that refining the definition of risk and incorporating additional factors—such as market concentration and degree of platform openness—is necessary to adequately capture long-term risks to democratic governance.

Author

• Unregistered Participant