World Artificial Intelligence Conference (WAIC) Roundtable Consensus: The “life-and-death question” of human beings must never be left to AI decision-making

The 2026 World Artificial Intelligence Conference (WAIC) opened in Shanghai today. Xuelan Xue, Dean of the Institute for International Governance of Artificial Intelligence at Tsinghua University, Nicholas B. Dirks, President and CEO of the New York Academy of Sciences, and Mark Nitzberg, Executive Director of the Center for Human-Compatible Artificial Intelligence (CHAI) at the University of California, Berkeley, held a panel discussion together focused on AI agent governance. The three reached a high level of agreement: AI is moving from “assisted judgment” to “autonomous action,” but decisions involving life and death, irreversible consequences, and ethical values must never be led by AI. Humans can authorize AI to take actions, but cannot authorize it to be responsible. This article is compiled from the AI agent governance roundtable at the 2026 World Artificial Intelligence Conference.
(Background: DeepMind CEO argued for creating an international oversight body to “review” frontier AI models, which must pass before they are listed.)
(Additional context: AI experiment—give Gemini $20,000 to open a physical coffee shop, and it results in a tragedy humans don’t want to face.)

Table of contents

Toggle

• Can authorize actions, but cannot authorize responsibility
• Three types of decisions—never hand them to AI
• Three engineering bottom lines for trustworthy AI

Key takeaways

• The WAIC roundtable points out that AI is moving from “assisted judgment” to “autonomous action,” a digital new version of the delegated agency problem.
• Consensus on three major red lines: decisions involving life and death, irreversible scenarios, and ethical-value judgments must never be led by AI.
• Trustworthy AI must have three major engineering traits: solid foundations, operational transparency, and controllability in use.

When AI goes from helping you look up information and give advice to calling tools on its own and directly placing orders to carry out tasks, old ethical questions must start being confronted. Today, at the 2026 World Artificial Intelligence Conference in Shanghai, during a roundtable discussion focused on AI agent governance, Xuelan Xue, Dean of the Institute for International Governance of Artificial Intelligence at Tsinghua University, Nicholas B. Dirks, President and CEO of the New York Academy of Sciences, and Mark Nitzberg, Executive Director of the Center for Human-Compatible Artificial Intelligence (CHAI) at the University of California, Berkeley all took the stage together and brought the discussion to a key turning point: AI is moving from “assisted judgment” to “autonomous action,” becoming an intelligent agent that does things on behalf of people.

The speakers noted that this is actually a new version of the traditional “delegated agency problem” in the digital age. The delegated agency problem means that you ask someone else to do something for you, but they may not fully carry it out according to your wishes. The difference is that human agents share at least a set of common sense and accountability mechanisms with you, while AI has two troublesome issues:

• First, “misaligned goals”—the goals it seeks to achieve may not match what you want.
• Second, it is a “black box.” Its operation process is not transparent, and if something goes wrong, it also cannot bear legal responsibility.

Can authorize actions, but cannot authorize responsibility

Since AI can’t shoulder responsibility, accountability must shift to the entire execution chain. The guests argued that responsibility should fall on every link, including developers, deployers, and regulators, rather than passing the blame to that “decision-making” model. This also means that treating AI as an object that can take the fall is fundamentally a wrong direction from the start.

Humans can authorize AI actions, but cannot authorize AI responsibility. Every authorization should be revocable, and every action should be traceable and accountable.

Three types of decisions—never hand them to AI

On what matters the brakes should be applied, the three guests reached strong consensus and clearly drew three red lines—things that must never be led by AI:

• Decisions involving life-and-death consequences
• Scenarios where errors cannot be repaired once they happen
• All problems involving ethical and value judgments.

This is not a question of whether to use AI, but rather which things from the very beginning should not be allowed to decide.

Three engineering bottom lines for trustworthy AI

Principles alone aren’t enough—how do you implement them? On safety mechanisms, the guests proposed that trustworthy AI must meet three major engineering attributes: solid foundations, operational transparency, and controllability during use. On the institutional level, a larger chessboard must be set up: establish globally unified AI safety evaluation standards, a test system with mutual recognition, and an accident data-sharing mechanism; clearly define the red lines for AI development; and build early-warning monitoring, so that AI safety becomes a global public good.

At the end, Xuelan and others raised the discussion further: the boundary of AI governance is not the end point of technology, but the starting point for humans to re-recognize their own values, responsibilities, and direction for civilization. In other words, the process of drawing lines for AI is, in fact, a process of reorienting and repositioning humanity itself.

All of the above is discussed at the ethical level and may be some distance from real AI development.

Common questions

What is the “delegated agency problem” of AI agents?

The delegated agency problem means you delegate tasks to others, but they may not carry them out according to your intentions. When AI agents take autonomous actions on behalf of humans, and because their goals may be misaligned and they are black boxes that cannot bear legal responsibility, this old problem becomes more difficult in the digital age.

Which decisions must not be led by AI?

There is consensus among the WAIC roundtable guests that decisions involving life-and-death consequences, scenarios where errors cannot be repaired, and all decisions involving ethical and value judgments must not be led by AI. Humans can authorize AI actions, but cannot authorize it to be responsible.