Quantifying the Salience of Geo-Cultural Values for Pluralistic Safety Alignment

Arkadiy Saakyan, Charvi Rastogi, Lora Aroyo

Safe global deployment of AI models requires alignment with human values that vary across cultures. Yet rater pools in safety evaluation datasets remain largely geographically homo…