国会山演讲实录|薛澜:AI 治理需全球携手，摒弃零和博弈

点击蓝字

关注我们

薛澜

国务院参事、清华大学苏世民书院院长、人工智能国际治理研究院院长、中国科技政策研究中心主任

I-

当地时间 4 月 30 日，美国参议员伯尼·桑德斯（Bernie Sanders）于国会山主持了一场名为“人工智能生存性风险与国际合作”的公开研讨会。会议由桑德斯领衔，出席嘉宾囊括：清华大学文科资深教授、人工智能国际治理研究院院长薛澜，麻省理工学院（MIT）物理学教授、未来生命研究所创始人迈克斯·泰格马克（Max Tegmark），北京前瞻人工智能安全与治理研究院院长曾毅，以及蒙特利尔大学助理教授戴维·克鲁格（David Krueger）。四位专家就超级智能前景、AI 是否构成生存威胁、治理路径及国际合作等核心议题进行了深度对话。

2026 年 4 月 30 日听证会上，薛澜教授的发言主题为《人工智能治理需要全球合作，而非零和竞争》，以下为演讲全文（整理版）：

各位同仁、各位朋友：

大家好！

今日，我们聚焦人工智能生存性风险与国际合作展开研讨，该议题日趋紧迫，其全球化属性也愈发凸显。

当下，从人工智能安全峰会到联合国框架下的多边对话，再到各类区域性与双边协作机制，国际社会已做出诸多有益尝试。但总体而言，现有举措仍显分散，尚未构建出一套真正具备协调力与执行力的全球治理架构。

依我之见，当前国际合作主要面临三大挑战：

其一，人工智能风险本身充满不确定性。无论是技术演进路线，还是潜在社会影响，诸多关键问题仍有待厘清。

其二，治理节奏滞后于技术发展。AI 迭代迅猛，而传统政府治理机制相对迟缓。这种“速度差”已成为全球共同面临的治理难题。

其三，地缘政治博弈加剧了合作阻力。主要 AI 国家间互信匮乏，导致诸多必要的协调机制难以有效落地。

面对上述挑战，中国近年来持续探索 AI 治理的有效路径。在此，我想重点阐述中国正在实践的“敏捷治理 + 适应性治理”双轮驱动模式。

所谓敏捷治理，即通过快速响应、持续迭代的治理机制，摒弃传统监管中“猫鼠游戏”式的对立思维，推动政企形成协同共治的良性互动格局，共同识别并回应风险。

适应性治理则侧重“边干边学”策略。先确立原则性指引，再逐步完善基础性法律法规，并依据技术与应用的演进不断推出专项治理措施，力求在促进创新与保障安全之间寻求动态平衡，从而构建更完整、综合的治理体系。

去年世界人工智能大会期间，多家中国 AI 企业联合签署了《中国人工智能安全承诺框架》，这也是企业参与全球 AI 安全治理的积极实践。

在此，我需特别强调一点：

中美两国在 AI 领域的竞争并非零和博弈，而应致力于推出性能最优、最为安全可靠、真正造福社会的人工智能模型。

因此，在更为宏大的地缘政治竞争背景下，可为 AI 安全合作划定“安全区”（safe zones）。在这些区域内，中美及全球科学家可率先开展务实合作，具体涵盖：AI 安全技术研究、安全标准互认、技术互操作协议，以及风险预警共享机制等。

与此同时，中美还应加强与发展中国家的协作，共同助力其 AI 能力建设，推动缩小全球“智能鸿沟”。

此外，我们还需持续关注未成年人保护、虚假信息生成等问题，并随技术发展不断完善相应治理框架。

AI 作为一项全球性技术，其风险与机遇均具跨国属性，唯有通过持续对话与真诚沟通，我们才能真正实现创新与安全之间的平衡。

谢谢大家！

Strengthening AI Governance and International Cooperation to Address Global Challenges

Distinguished Senator Sanders, fellow panelists, ladies and gentlemen,

First of all, let me thank you, Senator Sanders, for organizing this important discussion. As someone who has studied science and technology policy for most of my life, I see AI development as a transformative change that humanity must learn how to cope with responsibly. So I am grateful for this opportunity to learn from all of you and to share some thoughts from the perspective of governance and international cooperation.

Getting back to your question, I think the international community has certainly been trying, but so far the efforts are still not enough and not very effective.

Today, there are various multilateral mechanisms addressing AI governance and safety. We have seen the AI Safety Summit process that began in the UK in 2023 and continued through Seoul, Paris, and this year in Delhi, with future discussions possibly taking place in Geneva. There are also various United Nations mechanisms, including scientific advisory bodies and multilateral dialogues scheduled later this year. In addition, many regional and bilateral initiatives are emerging around the world.

These are all important efforts. However, overall, the current landscape remains fragmented and has not yet produced the level of coordination or effectiveness that the world needs.

There are several reasons for this situation.

First, there remains great uncertainty surrounding AI risks themselves. People may not yet fully understand the full range of risks ahead, and therefore may not always understand the implications of their own actions and decisions.

Second, there is the so-called “pacing problem.” Technological change in AI is moving much faster than governments are able to react. This gap between innovation and governance is becoming a common challenge across countries.

Third, the broader geopolitical environment makes it difficult for major AI powers to come together and design effective mechanisms and guardrails against AI-related risks.

Against this backdrop, China also recognizes the risks associated with AI development and is trying to balance innovation and safety through what we describe as an agile and adaptive governance approach.

First, on agile governance: because regulations and policies are almost always slower than technological change, governments may need to give up the assumption that regulation can always be perfectly accurate and comprehensive from the beginning. Instead, governance needs to act quickly, even if some gaps remain initially, and then continue to update and improve over time.

Another important aspect is that governments and companies should move beyond a purely adversarial “cat-and-mouse” relationship and instead work together to identify risks and address them collaboratively.

In addition, governance should avoid relying excessively on punishment when guidance and incentives may achieve better outcomes, except in cases involving clear violations of stakeholder interests.

On adaptive governance, China did not seek to build a complete overarching governance framework in a single step. Instead, China has adopted a “learning by doing” approach.

The process began with the development of governance principles and general guidance. Later, China gradually established foundational legal infrastructure to support AI governance, including laws related to personal information protection, data security, and cybersecurity. These laws created the broader framework within which AI systems operate.

As AI technologies continued to evolve, China also introduced more targeted regulations responding to specific technological developments. For example, following the emergence of large language models, China introduced the Interim Measures for the Management of Generative AI Services. These regulations continue to be updated and adjusted as technologies evolve.

At the same time, Chinese companies have also developed voluntary commitments related to AI safety practices. Last year, during the World Artificial Intelligence Conference in Shanghai, several Chinese AI companies signed an updated version of those commitments.

Putting these elements together, China has gradually developed a multi-layered governance system for addressing AI-related risks. Of course, the system still has weaknesses and areas that need improvement. However, it has helped support China's AI development while trying to maintain a balance between innovation and safety.

Looking ahead, I believe the real competition in AI is not simply a race between China and the United States. Rather, it is a global effort to see who can develop AI systems and services that are more capable, safer, more reliable, and more beneficial to society.

In this area, countries actually share many common interests. Therefore, even amid broader geopolitical competition, I believe there is still room to establish “safe zones” for cooperation in AI safety. Scientists and researchers from China, the United States, and other countries can work together on issues such as AI safety technologies, mutual recognition of standards, interoperability protocols, and risk early-warning mechanisms.

At the same time, major AI countries should also work together to support developing countries in strengthening AI capacity-building and narrowing the global AI divide.

Artificial intelligence is a global technology. Its opportunities and risks transcend national borders. Only through continued dialogue, mutual learning, and international cooperation can we effectively balance innovation and safety for the benefit of humanity.

Thank you very much!