AI risk perceptions and the effect of AI Act Safeguards. Results from a representative population sample

The EU AI Act introduces risk assessments for AI use cases alongside mandatory mitigation measures, aiming to ensure trustworthy AI systems. However, the ultimate success of establishing trusted AI systems depends on whether public perceptions of risk and mitigation align with these legal classifications. Currently, little is known about whether legal risk classifications and their corresponding safeguards match public risk perceptions. We address this gap by assessing how the public perceives the risks of different AI applications and evaluating how various mitigation measures, such as human oversight or independent certifications, are perceived to reduce those risks. Specifically, participants from an Austrian quota-representative population sample ( N = 1557 ) evaluated the perceived risk of three vignettes representing different regulatory risk tiers (workplace emotion detection, driver fatigue detection, and personalized media recommendations). We find that socio-demographic characteristics, such as age and gender, have only small, context-dependent effects on risk perception. Furthermore, mitigation measures decrease risk perception in two of our three use cases. We conclude that public risk perceptions and the efficacy of mitigation measures are highly context-dependent, suggesting that one-size-fits-all safeguards are not a universal solution to perceived AI risk.