As AI technologies continue to spread across sectors such as healthcare, finance, and consumer behavior, the perception of risks associated with artificial intelligence (AI) has become an increasingly important topic. This dissertation aims to address this lack of understanding by examining perceptions of AI-related risks from complementary conceptual, cognitive, and computational perspectives through four empirical studies. Together, these studies form a coherent whole, progressing from a review of existing research to expert-based qualitative modeling and computational predictions, ultimately addressing the linguistic limitations of such predictions.
Chapter 2 provides a systematic literature review of 64 peer-reviewed studies on the perception of risk associated with Artificial Narrow Intelligence (ANI). The review reveals that the research is concentrated geographically in the United States and China, and is clustered thematically around healthcare, consumer behaviour and finance. A key finding is that, although many studies examine how AI risk perceptions manifest, virtually none investigate how these perceptions are formed. This observation motivates the subsequent chapters.
Chapter 3 addresses AI in health by examining risk perception for AI through a mental model approach. The study introduces a hybrid workflow combining structured expert panel methods with large language models (e.g. ChatGPT) to develop and compare cognitive risk models. It shows that these models can provide a structured basis for domain-specific risk communication.
Chapter 4 shifts focus from qualitative to computational methods, exploring word embedding models to predict risk perception. The results show that, rather than improving predictive accuracy, augmenting word vectors with additional context words generally degrades it, while geographic vector operations yield only modest shifts in predictive outcomes. These findings emphasise the potential and limitations of embedding-based approaches to risk perception.
Chapter 5 puts these computational approaches to a rigorous cross-linguistic test. Using a newly collected German-language dataset, the study examines whether models based on word embeddings trained using U.S. data can be applied to a German context. The analysis reveals a systematic language bias: predictive accuracy decreases substantially, and this gap is only partially closed by language-matched embeddings. These results imply that previous prediction successes may be partly due to the similarity between the training corpus and the survey population rather than reflecting universal language-risk associations.
Taken together, the four studies converge on a central insight: risk perception is shaped by psychological, societal, and linguistic factors that resist simple quantification. For researchers and policymakers alike, the implication is that effective risk communication about AI must be grounded not only in statistical models but also in an understanding of the linguistic and societal contexts in which risk perceptions arise.
