When rivals share the same fear: why the U.S. and China need common rules for AI

Some reports are worth less for what they describe than for what they hint at, and Will Knight's dispatch from Beijing's Zhongguancun tech district belongs to that category. The snapshot is sharp: at a conference organized by the Beijing Academy of Artificial Intelligence, featuring figures of the caliber of Whitfield Diffie and recent Turing Award winner Andrew Barto, the topic that ultimately took over was neither recursive self-improvement nor humanoid robots, but a concern shared on both sides of the Pacific. Chinese experts, the journalist reports, voiced a genuine alarm that mirrored that of their American colleagues. That convergence is, in itself, the story.

The underlying argument is elegantly summed up by MIT's Stephen Casper when he recalls that AI is 'a global technology with global benefits, global harms and a consistent tendency for new capabilities to end up proliferating.' The observation carries a practical consequence that is hard to dodge: open-weight models—Moonshot's Kimi, Alibaba's Qwen, Z.ai's GLM—cannot be recalled once released. Unilateral containment, in that scenario, is a comforting illusion. If a dangerous capability exists and circulates, the wall a country builds protects that country only from its own prudence, not from the real risk.

The analogy that structures the report is the nuclear one, and it is worth taking seriously without overdramatizing it. Casper notes that, even at the height of the Cold War, Washington and Moscow cooperated on nuclear safety while competing on weapons. His phrase—'AI does not need a Chernobyl moment'—works because Chernobyl was not just a technical accident, but the collapse of public trust in an entire technology. Applied to AI, such a moment could take the form of a large-scale automated cyberattack or a cascading failure of agentic systems over critical infrastructure. What matters is that both Chinese and Western researchers seem to understand that no one would come out ahead from an episode like that.

The article's most useful nuance comes from Lin Yun, of Shanghai Jiao Tong University, who rejects easy optimism: in the short term, he argues, attackers will hold the advantage because offensive tools will mature before defensive ones, and only over time could defensive AI rebalance the scales. Even so, he advocates seeking 'areas where sharing can reduce systemic risk without exposing sensitive operational details.' That is the sentence every negotiator should underline, because it sketches a realistic middle ground between naive cooperation and total secrecy.

From a constructive standpoint, the hopeful part is not that the risks have vanished—they have not—but that the technical communities of the two powers already share a diagnosis. The history of technology regulation teaches that standards are born first among engineers who understand each other, and only later reach, with luck, the political offices. If researchers speak the same language about vulnerabilities, agentic tool use and automated social engineering, there is a foundation to build on. The question, then, is not whether cooperation is desirable, but whether political will can seize the window science is opening before an incident closes it.