PhD defence Jonas Benjamin Krieger

Keywords: Risk perception, Artificial intelligence, Mental models, Word embeddings

"Understanding and Predicting Risk Perception of Artificial Intelligence - Conceptual, cognitive, and computational perspectives"

As artificial intelligence becomes part of everyday life, understanding how people perceive its risks is becoming increasingly important. To this end, this thesis conducted a review of 64 studies on AI risk perceptions, finding that most studies focused on measuring AI risk perceptions without addressing their development. The second study of the thesis suggested a novel approach for developing mental models, using AI risk perception in health as a case study. To this end, the study compared and expanded the traditional expert panel approach with results from ChatGPT, to determine whether large language models (LLMs) could streamline the model development process. Two further studies examined whether patterns in large text corpora could predict risk ratings. While this approach performed well with U.S. data, its performance declined substantially when applied to data from German-speaking participants. Using German text as training data did not eliminate this gap. These findings demonstrate that risk prediction models reflect their linguistic context and should not be transferred to other settings without thorough validation.

Click here for the live stream.