Skip to main navigation Skip to search Skip to main content

Systemic AI risks: from catalyst to cataclysm

Activity: Talk or presentationTalk or presentation at a conference

Description

“Systemic” risks have become key in recent EU digital regulations. In the AI Act, systemic risks are associated with the “high-impact capabilities” of general-purpose AI models (GPAIs, e.g., ChatGPT): the higher the capabilities, like the cumulative amount of computation used for the model’s training, the higher the risks. In the Digital Services Act, they concern specific online platform functionalities that rapidly amplify negative effects, e.g., on public health. In other fields, it is the complexity and cascade effects that either expand or narrow our understanding of risk, depending on the “system” considered, such as the collapse of health systems due to pandemics
or the vulnerability of digital networks and cybersecurity systems. However, what makes a risk distinctively “systemic”, and how AI capabilities or platform functionalities are defined, remain vague and underdiscussed.
The concept of “risk society” helps us understand not only specific instances of risk, but also how, as technological and health risks grow, our collective uncertainty about the future has made risk a political concern to be managed via institutions, processes, and regulations. By analyzing legal provisions and the literature on the sociology of risk, we question how systemic risks challenge traditional models of uncertainty and risk governance. We hypothesize that the systemic risks posed by GPAIs differ from the classical physical risks in the risk society, and not merely due to their more “immaterial harms”. Classical risk paradigms framed uncertainty about
future risks as a feature of the environment (e.g., the unpredictable spread of a disease) or as a result of human ignorance. In the case of (GP)AI, by contrast, uncertainty is no longer merely a matter of ignoring the environment’s responses or humans’ behavior, but seems to stem from the raw, domain-independent capabilities of AI systems to catalyze possible consequences of events to systemic cataclysmic dimensions.
Period29 Oct 2025
Event titleCaring during crisis: navigating risk and uncertainty in health, care and beyond (Midterm Conference 2025 European Sociological Association, Research Network 22 Sociology of Risk and Uncertainty)
Event typeConference
LocationRotterdam, NetherlandsShow on map
Degree of RecognitionInternational