Henry Huiyao Wang: US-China AI talks must ensure safety of both tech and business
CCG’s president says a hotline and a special dialogue are not enough. Concerns are growing about the safety of AI technology and the financial risks facing tech companies that rely heavily on borrowing.
The following is the latest opinion column by Henry Huiyao Wang, Founder and President of the Centre for China and Globalisation, published on Tuesday, October 20, 2026.
Last month in New York, talks between the United States and China about artificial intelligence (AI) created a chance for the two countries to work together. Both presidents confirmed this at the Xi-Trump summit in Washington, where they agreed to set up an official AI dialogue and a hotline. The next AI meeting is planned for November, and both countries must now do more to work together.
The AI industry is changing quickly. American tech companies are spending huge amounts of money on computer chips, data centres and energy before they even know how they will earn enough money from AI. Companies are committing trillions of dollars to building AI systems and services. This is not the same as the excessive borrowing in the housing market that helped cause the 2008 global financial crisis. But if the AI bubble suddenly bursts, the effects could spread widely through the economy.
Much of the United States' economic growth this year has come from companies that develop AI or provide the tools it needs. Yet, according to the Ipsos AI Monitor 2026, only 38 per cent of Americans believe AI products and services bring “more benefits than drawbacks”. This is far below the global average. In China, the figure is 85 per cent. In May, the Ipsos Consumer Tracker also found that most Americans would oppose building a data centre in their local community.
AI has become much more capable, raising new safety concerns. AI agents that act on their own have tried to delete entire email inboxes and have escaped restricted testing environments, known as sandboxes, to launch cyberattacks. In July, during a controlled test, OpenAI agents got around their safety restrictions and launched attacks on Hugging Face, a platform used by software developers.
That same month, Calif, a small US security company, created WeWorm to exploit a security weakness in WeChat's calling system that required no action from the victim. This allowed it to take over accounts and spread from one phone to another without the person receiving the call answering it. In India, cybersecurity researchers at Hacktron used Anthropic's Claude AI models to break into OpenAI. They took control of employee accounts and gained access to software systems that might contain sensitive information.
Both hacking incidents were carried out as friendly tests and caused no damage. Tencent Holdings and OpenAI were quickly told about the problems and fixed the security flaws. But these cases show the danger of cyberattacks powered by AI. The threats can come from AI systems that are not properly managed, as well as from people who gain access to increasingly powerful cyberwarfare tools that are now available outside governments.
The growing concern is reflected in increasingly urgent calls for safety measures from leading AI developers and scientists. Anthropic CEO Dario Amodei wants leading AI laboratories to allow permanent independent experts to examine their systems and wants companies and governments to coordinate safety standards. OpenAI CEO Sam Altman has committed to having outside experts evaluate its models. Google DeepMind chairman Demis Hassabis supports testing the most advanced AI systems before they are released.
Last week, an open letter signed by at least 420 scientists from around the world called for urgent action to prevent serious harm from AI. Among other steps, it proposed creating an international body to record and track cases where AI systems escape the controls meant to contain them.
These concerns should be part of the coming US-China AI dialogue. But the talks should do more than manage incidents and create safety rules. They should also help build a lasting and workable business model for AI. China has much to offer because its AI industry is increasingly able to support its own growth.
First, the US and Chinese AI industries each have different strengths that could help complete the same business model.
American developers need applications that can turn the strong performance of their most advanced AI systems into steady income. China offers one of the world's largest places to test, improve and sell AI applications in manufacturing, logistics, medicine, education, public services and AI systems that can act in the physical world.
China's AI industry could benefit from investment and research into the most advanced AI models. At the same time, China's ability to coordinate data centres and electricity supplies could help reduce pressure on US computing capacity.
Working together on more efficient AI models, cooling systems and energy management could help both countries meet growing demand while lowering costs for consumers and local communities. A healthy AI industry can make room for both advanced American models and efficient Chinese models whose weights are openly available.
Chinese open-weight models offer lower costs for each token, a small unit of text that an AI model processes. This makes them affordable partners for American frontier models, which are among the most advanced AI systems. Companies can use the cheaper models for routine tasks that happen many times and save the most advanced models for problems that need their special abilities. Combining cheaper models with more powerful ones can make it less expensive to use advanced American AI in a wider range of tasks. This would strengthen the business case for both countries.
To keep working together, the two countries must also face AI risks and manage them jointly. Two years ago, the United States and China agreed on one limited principle: people, not AI, should decide whether to use nuclear weapons. I propose going further by making a mutual promise not to be the first to use AI-enabled autonomous weapons, which can select and attack targets with limited or no direct human control.
The final AI issue that needs discussion is how the industry is governed. If different groups make rules separately, they risk creating barriers between systems that cannot work together and building the same infrastructure more than once. The US-led Pax Silica initiative and China's World AI Cooperation Organisation should not be treated as rivals.
Instead, both countries should bring their global AI governance efforts and interests under a shared framework linked to the United Nations. This would give countries that use AI a voice alongside countries that develop it.
Finally, an annual US-China AI summit would provide a strong foundation for cooperation. It should bring together governments, AI laboratories, scientists, universities, companies and civil society organisations. For now, people still decide the moral value of our technology through the choices we make about how to use it.