Generated by Codex with GPT-5

This summary covers The Economist’s May 9th, 2026 China article listed in the contents as US-China AI negotiations and published under the headline A MAD problem.

The article argues that artificial intelligence is becoming a strategic problem of the sort usually associated with nuclear weapons: both America and China fear the danger, but neither wants to slow down if the other might race ahead. That tension is what makes AI diplomacy urgent and difficult. The technology promises economic power, military leverage and scientific advantage. It also raises the prospect of cyberattacks, biological misuse and systems that become hard for humans to control.

The immediate setting is Donald Trump’s planned meeting with Xi Jinping in Beijing. Trade, Taiwan and the Middle East are already on the agenda, but the article suggests that AI belongs beside them. As frontier models grow more capable, they stop looking like ordinary commercial products and start looking like strategic assets. Whoever leads may gain an advantage that is hard to reverse; whoever mismanages the risk may endanger everyone.

A Shared Risk With Rival Incentives

The article’s central contrast is between common danger and competitive mistrust. American officials are no longer treating AI regulation as a matter best left entirely to the market. Anthropic’s decision to restrict access to Mythos, a model described as unusually capable at finding cyber vulnerabilities, has sharpened attention in Washington and abroad. If models can help break into software, design dangerous biological tools or pursue goals misaligned with human intent, then safety is not only a corporate governance problem. It becomes a national-security problem.

China has reasons to care, too. Chinese advisers and AI-safety experts have publicly called for global rules, even for slower development. Beijing has pushed ideas such as a UN body for AI governance, international standards to keep humans in control and broader access to AI tools for poorer countries. That gives China the language of shared safety and global fairness.

Yet neither side trusts the other’s motives. Some Americans suspect China is performing concern about safety to constrain American labs or score diplomatic points. Some Chinese strategists see American safety talk as a way to preserve technological dominance, especially while Washington restricts China’s access to advanced chips. The result is a classic arms-race trap: both sides may prefer some limits in theory, but each fears that accepting them would leave it weaker in practice.

Why Bilateral Talks Matter

The article treats a direct American-Chinese process as more plausible than a broad global bargain. America and China hold most of the world’s frontier computing power, so they are the countries with the most meaningful leverage over the strongest models. They also have the most to lose if AI becomes destabilising.

There is already a thin history of cooperation. Trump and Xi have previously agreed in general terms to work together on AI, and Joe Biden secured a narrower understanding that nuclear weapons should remain under human control. Researchers and former officials from both countries have also met through unofficial channels. Some Chinese labs have adopted open-source technical work from Anthropic for agent communication.

But the article is clear that this is not yet enough. Informal dialogue often lacks the technical experts needed to align standards. Official meetings can become hostage to unrelated disputes, such as export controls or Taiwan. Safety discussions are also hard because the information needed to evaluate models can be close to the information that would help improve them. A country may be reluctant to share test results if it thinks doing so reveals strategic capability.

The piece lays out three possible levels of cooperation. The simplest is dialogue: the two governments could discuss risks, expectations and technical assumptions so that parallel rulemaking does not drift into dangerous misunderstanding. A deeper version would involve common safety tests, even if each side kept its own results private. The most ambitious version would resemble arms-control verification, with shared evaluations and some way to check compliance, perhaps through inspections or reporting on data-centre activity to an international monitor.

The Verification Problem

The hardest question is whether either country could believe the other. Nuclear arms control has always depended not only on promises but on ways to verify them. AI makes that harder. Data centres, model weights, training data, safety evaluations and deployment practices are not as visible as missiles or enriched uranium. Some evidence of safety is also evidence of capability. Sharing too much could feel like handing over a map of one’s own labs.

China’s position is especially complicated. Its leading models have often been more open than American frontier systems, which worries Washington because open weights can be copied and adapted by bad actors. At the same time, Chinese officials are wary of arrangements that might freeze their firms below America’s level. The article connects that suspicion to China’s historical memory of unequal treaties and to its present frustration with chip controls.

America’s politics are not simple either. Some American officials and technologists talk about winning the AI race in existential terms. If policymakers see AI supremacy as a must-win contest, serious restraint will be hard to sell. A safety agreement that looks prudent to one audience may look like unilateral disarmament to another.

The Real Lesson

The article’s strongest point is that AI diplomacy cannot wait until disaster makes cooperation obvious. Past safety regimes often gained momentum only after catastrophe. That is a dangerous pattern for a technology that may move faster than governments can study it.

The practical takeaway is not that America and China are close to an AI treaty. They are not. The article instead presents a narrower but important claim: the two rivals may eventually need to treat frontier AI as a shared strategic hazard, not merely as an industry to dominate. Competition will continue, because both governments see AI as central to power and prosperity. But unmanaged competition could make both countries less secure.

In plain English: The Economist sees AI as reviving a cold-war dilemma in digital form. America and China have incentives to race, incentives to cooperate and powerful reasons to distrust each other. The hard task is to build enough common ground before the technology itself supplies the warning.