WAIC 2026 roundtable: Life-and-death decisions must never be left to AI; humans can authorize actions but cannot authorize responsibility

According to Beating monitoring, at the 2026 World Artificial Intelligence Conference, Xue Lan, Director of the Institute for International AI Governance at Tsinghua University, Nicholas B. Dirks, President of the New York Academy of Sciences and CEO, and Mark Nitzberg, Executive Director of the Center for Human-Compatible AI at the AI Research Laboratory of the University of California, Berkeley, held a roundtable discussion on AI agent governance. The guests pointed out that AI is moving from “decision support” to “autonomous action,” becoming an intelligent agent that substitutes for humans in getting things done—an entirely new manifestation of the traditional principal-agent problem in the digital age. Unlike human agents, AI faces misaligned goals and, as a black box, cannot bear legal responsibility, so accountability must shift to the entire execution chain, including developers, deployers, regulators, and others.

The guests reached strong consensus: For decisions involving life-or-death consequences, scenarios where once errors occur they cannot be repaired, and all issues involving ethical and value judgments, they must never be led by AI. Humans can authorize AI to act, but cannot authorize AI to be responsible; every authorization should be revocable, and every action should be traceable for accountability. The speed of delegating agency should never exceed the speed at which humans verify AI capabilities.

At the level of safety mechanisms, the guests proposed that building trustworthy AI must satisfy three major engineering characteristics: a solid foundation, operational transparency, and controllability during use. In terms of institutional development, it is necessary to establish globally unified AI safety assessment standards, a mutually recognized testing system, and an incident-data sharing mechanism; to define red lines for AI development and build an early-warning monitoring mechanism; and to make AI safety a global public good. The boundary of AI governance is not the technical endpoint, but the starting point for humanity to re-recognize its own values, responsibilities, and the direction of civilization.

Click the original link below to join the Beating · Feishu AI News channel and get 24/7 nonstop monitoring of global AI hotspots and news.