Abstract While discourse around the risks of artificial intelligence (AI) is gaining importance, conceptual reflections on AI risk remain limited. Since both AI and risk have been defined differently across disciplines, there is in principle a plurality of ways in which AI risk could be understood. This makes it possible that the deceivingly simple notion of AI risk covers a variety of conceptual commitments and research objects. Especially in light of the interdisciplinarity of the field, a lack of clarity and reflective use of the concept of AI risk could muddle the field’s theoretical development. This paper therefore investigates the conceptualization and study of AI risk to structure the current understandings of AI risk. I do this by conducting a systematic literature review that specifically focuses on the way risk is defined, characterized, and researched in relation to AI. Among the 81 papers included in the study, I identify four distinct conceptualizations of AI risk: (1) risk as a regulatory tool to govern AI; (2) AI risk as a given reality that needs to be assessed and managed; (3) AI risk as a perception people hold and as a framing by the media; (4) AI risk as existential risk of artificial general intelligence. Based on a deep reading of the texts, I provide an overview of what research on AI risk from 2012 to 2024 looks like, contrast different approaches to AI risk, and identify of pressing lacunas within each cluster. Lastly, I highlight that the related notion of acceptable AI risk plays a key role yet is also underdeveloped.
Lieke Fröberg (Mon,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: