This talk highlights the effectiveness of semantic representations from free association data rather than text-based language models for understanding thought and behavior in the domain of risk. I will present two projects. The first project shows how word embeddings from free association data more accurately predict risk perception and enhance text-based models. The second project uses free associations to uncover how people understand risk and their individual differences in risk preferences. These projects reveal that incorporating human-generated data can offer a more accurate and insightful approach to understanding risk-related behaviors, suggesting potential applications beyond this domain