Taxonomy of AI Risks in the Spanish Press

This document describes a taxonomy of AI risks developed as part of a research on risk discourse in the Spanish press. To identify the main discourses on artificial intelligence risks, a corpus of 2,705 articles was compiled from six Spanish digital media outlets, covering the twelve months following the launch of ChatGPT (December 2022 to November 2023). From these texts, 878 fragments containing the term “risk” in the specific context of AI were identified. Of these, 547 explicitly referred to a risk object—that is, someone or something potentially affected by AI—which enabled a systematic classification. Drawing on risk discourse literature (Boholm, 2017; Steimers & Schneider, 2022), the analysis adopted the criterion of the risk object to structure the taxonomy. This approach categorises AI-related risks into seven main domains: risks to humanity, to individuals, to the economy, to social groups, to institutions, to technological systems, and to the environment. This classification helps to capture the multidimensional and often indirect nature of AI risks—unlike more linear threats such as nuclear energy or tobacco use. It also supports the design of targeted mitigation strategies based on the context and the groups affected. In parallel, the study identified the key voices involved in public debates—such as academics, tech companies, and political actors—and tracked their presence throughout the study period. The taxonomy of AI-related risks is structured around the concept of the risk object, that is, the entity potentially affected by artificial intelligence. This approach allows for a clearer understanding of the scope and diversity of concerns raised in public discourse. The risks identified are grouped into seven categories, ranging from existential threats to humanity to environmental impacts, encompassing personal, social, institutional, and systemic dimensions. This classification highlights the multifaceted nature of AI risks and the need for differentiated responses depending on the affected domain.