Toward Standards for Responsible AI in Health Communication: A Dual Method Pre-Deployment Risk Evaluation Framework and Case Study

Context: Generative AI tools are increasingly being considered for use in public health research and practice, yet standardized frameworks for pre-deployment evaluation of the risks AI tools pose - across dimensions of platform safety, audience appropriateness, and user readiness - have not been established for health communication contexts. Program or Policy: We developed and implemented a dual-method pre-deployment evaluation framework that combines adversarial red-teaming and structured user testing to assess the readiness of a health communication AI tool. The framework was applied to the Trusted Health Information Assistant (THIA), a retrieval-augmented generation chatbot designed by NORC at the University of Chicago for use in public health communication research. Implementation: The framework was implemented sequentially, beginning with red-teaming to identify safety, audience appropriateness, and reliability concerns before engaging end users. Five independent testers completed 28 structured red-team prompts, after which identified issues were remediated and verified. User testing was subsequently conducted with six participants (e.g., caregivers of older adults) and included direct interaction with THIA, pre- and post-survey assessments, semi-structured interviews, and a review of backend analytics. Evaluation: The sequential, dual-method design successfully identified distinct categories of risk. Red-teaming identified platform integrity issues, including personally identifiable information (PII) handling and citation reliability. User testing subsequently assessed platform readiness under real-world conditions, surfacing trust dynamics and usability patterns. The convergence of findings across both methods validated the complementary design: each method identified distinct dimensions, together producing a more thorough and actionable assessment. Discussion: As generative AI tools increasingly enter public health research and practice, ensuring that potential risks are systematically identified and addressed before they reach end users is paramount. Yet the field lacks shared standards for pre-deployment evaluation. AI developers rarely publish evaluation findings, leaving public health organizations with few replicable models to guide their own assessments. This report addresses that gap by documenting a structured, sequential framework including methods, instruments, scoring rubrics, and reporting structures for evaluating health communication AI tools prior to deployment.