The perception of risk associated with artificial intelligence (AI) has become a topic of increasing importance as AI technologies proliferate across different sectors of life. This thesis addresses a gap in understanding how AI risk perceptions are formed and whether they can be modelled computationally through four studies that examine AI risk perceptions from conceptual, cognitive, and computational perspectives. Together, they form a bridge: starting with an assessment of the current state of research as part of a systematic literature review, extending to (qualitative) modelling of a mental model to better understand the cognitive mechanisms underlying risk perception of AI in health, to the development of embedding-based computational models for predicting risk perception based on large text corpora, as well as the exploration of their limitations in the context of different languages.
The thesis reports four main findings:
First, it provides a systematic literature review of 64 peer-reviewed studies on risk perception of Artificial Narrow Intelligence (ANI), which shows that research is geographically concentrated in the United States and China and thematically clustered around health, consumer behaviour, and finance. In this regard, a critical finding is that while many studies examine how AI risk perceptions manifest, little attention has been paid to how these perceptions are formed.
The identified research gap was addressed by introducing a new hybrid workflow that extends the 'classical' expert panel approach by utilising a large language model (here, ChatGPT), thereby addressing the original methods' organisational and practical limitations. By doing so, an expert model for AI in health has also been developed, which can be used for further research.
Building on this conceptual understanding, the subsequent studies evaluate embedding-based computational models for predicting risk perception, showing that augmenting word vectors with additional context words generally degrades rather than improves predictive accuracy, whilst operations on geographical vectors result in only minor shifts in predictive outcomes.
Embedding-based computational models were subsequently evaluated and shown to predict risk perception with high accuracy. It was further demonstrated that targeted manipulations of the embedding space alter semantic associations between risks and related concepts, illustrating the potential of embedding models to represent context-dependent patterns in risk perception. These findings served as a basis for a more rigorous cross-linguistic test, which revealed that predictive accuracy substantially decreases when the models are applied to a German-language dataset. This gap is only partially reduced by using language-matched embeddings. These results suggest that previous predictive successes may be partly explained by the alignment between the training corpus and the survey population, rather than reflecting universal relationships between language and risk.
Overall, the four studies suggest that risk perceptions (in relation to AI) should be understood as context-dependent cognitive phenomena rather than merely quantifiable outcomes. Expert models provide an important conceptual basis for the development and interpretation of embedding-based computational models for predicting risk perception, while the observed cross-linguistic limitations indicate that embedding-based models capture language- and context-specific semantic structures rather than universal representations of risk. Consequently, computational predictions cannot be assumed to generalise across linguistic and cultural settings.
Given that contemporary large language models such as ChatGPT and Claude are predominantly trained on English-language corpora, these findings highlight the importance of validating embedding-based computational models for predicting risk perception before their application to practical risk communication or decision support in multilingual contexts.
