Beyond Compliance: Generative AI Safety Evaluation for Civil Society

This archive contains the presentation materials for “Beyond Compliance: Generative AI Safety Evaluation for Civil Society,” delivered by Ashley Khor at the IHS Summer Graduate Conference on May 31, 2025. The talk introduces a hybrid evaluation framework designed specifically for high-impact, community-driven AI use cases. Developed through both academic review and field practice, the framework incorporates five core dimensions of safety: Technical and content integrity Relational and ethical grounding (design justice, feminist, and trauma-informed principles) Usability and accessibility Organizational readiness Contextual appropriateness for sensitive, user-directed interactions. In addition to defining these dimensions, the project extends current GenAI safety practices through a participatory methodology that includes five test-a-thon methods: human-in-the-loop scoring, participatory red teaming, simulated user journeys, reflective feedback circles, and observer-based shadow scoring. Together, these methods offer a survivor-centered and emotionally attuned lens for evaluating AI systems — not only by their outputs, but by how they are experienced, trusted, and used.