DeepMind researcher proposes Normative Alignment for principled agentic safety

verena_rieser · x · 2026-08-28

Verena Rieser's ICML keynote explores safe autonomous decision-making for agents in complex environments. She proposes 'Normative Alignment', moving beyond passive harm avoidance to equip agents with 'Agentic Integrity' to interpret and dynamically apply abstract principles. This path faces challenges in capability, metrics, and governance, emphasizing contextual reasoning over reward maximization.

Original post →

More from Safety

Safety channel →