According to monitoring by Dongcha Beating, Xue Lan, dean of Tsinghua University’s Institute for AI International Governance, Nicholas B. Dirks, president and CEO of the New York Academy of Sciences, and Mark Nitzberg, executive director of the Center for Human-Compatible AI at the University of California, Berkeley’s AI research lab, joined a roundtable at the 2026 World Artificial Intelligence Conference to discuss governance for AI agents.
The speakers said AI is moving from assisted judgment toward autonomous action, turning into an agent that can act on behalf of people. They framed that shift as a new digital-age expression of the traditional principal-agent problem. Unlike human agents, AI faces goal-alignment issues and, as a black box, cannot assume legal liability. That means accountability has to be assigned across the full execution chain, including developers, deployers and regulators.
Panel draws a line around life-and-death decisions
The panelists said they shared a strong consensus on several boundaries. Decisions involving life-and-death consequences, situations where mistakes cannot be repaired, and all questions involving ethics and value judgment should not be led by AI.
They added that humans may authorize AI to take action, but cannot authorize AI to bear responsibility. Each authorization, they said, should be revocable, and every action should remain traceable for accountability. The speed at which agency is delegated should never exceed the speed at which humans can verify AI’s capabilities.
Three engineering properties for trustworthy AI
On safety mechanisms, the speakers said trustworthy AI should meet three engineering requirements: a solid foundation, transparent operation and controllability in use.
Calls for shared standards and early-warning systems
On institutional design, the discussion pointed to the need for globally unified AI safety assessment standards, mutually recognized testing systems and incident data-sharing mechanisms. The panel also called for clearly defined red lines for AI development and early-warning monitoring mechanisms, with the aim of making AI safety a global public good.
The speakers said the boundaries of AI governance are not the endpoint of technology, but the starting point for humanity to rethink value, responsibility and the direction of civilization.