Safety is a Process, not a Score: A Symbol-Aware Safety Evaluation Methodology for GenAI for Social Good Tools in High-Emotion Contexts

Abstract

Generative AI tools are increasingly being deployed in sensitive social contexts – from mental health to justice systems – yet current safety metrics remain largely quantitative, decontextualized, and technically narrow. This paper introduces a novel, survivor-informed framework for evaluating GenAI systems in high-emotion, high-risk, or public-facing use cases. Rooted in trauma-informed design and symbolic resonance theory, the “Safety is a Process, Not a Score” framework prioritizes co-regulation, narrative fidelity, and epistemic alignment over one-size-fits-all benchmarks. We describe a collaborative methodology developed with survivors of gender-based violence, including a safety rubric, qualitative risk-mapping protocol, and structured, participant-led test-a-thons. Drawing from a recent field test involving a public-facing GenAI tool, we reflect on what it means to build safety relationally, not just statistically. This approach expands both the evaluative vocabulary and participatory possibilities for AI ethics in real-world deployment.