Link: http://arxiv.org/abs/2501.08814v1
PDF Link: http://arxiv.org/pdf/2501.08814v1
Summary: The rapid adoption of generative AI in the public sector, encompassingdiverse applications ranging from automated public assistance to welfareservices and immigration processes, highlights its transformative potentialwhile underscoring the pressing need for thorough risk assessments.
Despite itsgrowing presence, evaluations of risks associated with AI-driven systems in thepublic sector remain insufficiently explored.
Building upon an establishedtaxonomy of AI risks derived from diverse government policies and corporateguidelines, we investigate the critical risks posed by generative AI in thepublic sector while extending the scope to account for its multimodalcapabilities.
In addition, we propose a Systematic dAta generatIon Frameworkfor evaluating the risks of generative AI (SAIF).
SAIF involves four keystages: breaking down risks, designing scenarios, applying jailbreak methods,and exploring prompt types.
It ensures the systematic and consistent generationof prompt data, facilitating a comprehensive evaluation while providing a solidfoundation for mitigating the risks.
Furthermore, SAIF is designed toaccommodate emerging jailbreak methods and evolving prompt types, therebyenabling effective responses to unforeseen risk scenarios.
We believe that thisstudy can play a crucial role in fostering the safe and responsible integrationof generative AI into the public sector.
Published on arXiv on: 2025-01-15T14:12:38Z