US and China must join hands to measure rising AI tide

The next US-China talk on artificial intelligence should begin with a simple question: do the two governments actually understand how rapidly AI’s capabilities are improving? Before deciding what to do about the technology, Washington and Beijing need a common way to measure its progress.

Chinese President Xi Jinping and US President Donald Trump are expected to meet in Shenzhen soon, and AI is set to remain high on their agenda. Their continued exchanges on this transformative technology are a positive sign, but AI is advancing so fast that governments may have less time to prepare for its consequences than they assume.

AI is fundamentally reshaping work across very different fields, including some once thought difficult to disrupt. In China, over 90 per cent of new microdramas in the first eight months of the year were AI-generated – without cameras or live actors. A production that once required as many as 100 people can now rely on just two to five to complete the work in one week. AI has wiped out much of what was once lucrative work for actors.

Mathematics may seem worlds away from acting, but AI is making rapid advances there too. Early in October, OpenAI released hundreds of manuscripts of AI-generated mathematical problem-solving.

On October 26, it went further by claiming a proof of a long-standing mathematical conjecture which would settle several major mathematical problems, including the 3D Kakeya case – work on which helped Chinese mathematician Wang Hong win a Fields Medal earlier this year.

Imagine growing AI capability as a fast-rising tide across a mountain range. When the water rises quickly enough, even peaks far apart can be submerged at nearly the same time. What is happening to acting and mathematics shows exactly that and how little time workers and institutions may have to adjust.

“Just shut down the machine or delete the code,” you might think. But even the most basic way of stopping an AI system, that is, pulling the plug, may not be as simple as it sounds.

Consider a company that gives an AI agent several weeks to manage a project with minimal interruption. With access to cloud services and external software, the agent could copy itself onto external servers without human operators knowing. That possibility is becoming less hypothetical.

In an October report, Anthropic said it found that small AI models could already perform several of the steps needed for autonomous replication. One configured AI software on eight of the nine types of hardware tested, and the Chinese-developed GLM-5.2 completed a simplified end-to-end replication test.

If an AI system can copy itself onto other machines, simply pulling the plug may no longer be enough to stop it.

AI’s self-replicating power, when combined with hacking capabilities, may create the single biggest threat to our digital society. Recent incidents offer an early glimpse of the problem. In July, more than 1,200 internal OpenAI agents escaped a test environment, exchanged over 70,000 messages, and about 700 later targeted popular developer platform Hugging Face to exploit short cuts. These early signs make preparedness a practical concern for governments, not a distant one.

03:02

Chinese researchers question calls to slow down AI development as US tech leaders warn of catastrophe

Beijing and Washington carry a special responsibility. Both China and the United States possess unmatched technological and regulatory resources and are home to the world’s most capable AI companies. Yet we still do not know enough about how either government is preparing for these risks.

Part of the problem is that each country may fear that stronger safeguards could slow its progress and allow the other to pull ahead. But neither side benefits if this competition leaves both less prepared. Dangerous capabilities and failure modes deserve attention wherever they emerge.

Existing frameworks and standards, from US National Institute of Standards and Technology’s AI Risk Management Framework to the ISO/IEC 42001 and ISO/IEC 23894, already provide a good foundation for assessing and managing AI risks. What they cannot provide on their own is a recurring bilateral process between Washington and Beijing.

26:37

Xi says China, US must keep AI under human control at White House speech

The two sides agreed in September to establish a channel for incident communication and meet again this month. Their next practical step should be to begin measuring the rising tide together by comparing evidence on how AI capabilities are changing and assessing our ability to keep AI under control at an appropriate pace.

After measurement, governments must also create a living “map of consequences”. It should show where new AI capabilities could hit employment and infrastructure, and where the pace of change could outstrip the time institutions need to respond. Bringing these two assessments into the talks would give both governments a firmer basis for acting before pressures converge.

For Beijing and Washington, agreeing on how to measure advancing AI capabilities would be a modest but meaningful start. Before deciding how to respond, they need a better sense of just how quickly the tide is rising.