The AI Safety Theater: Why Washington and Beijing Will Never Brake
Both superpowers are publicly calling for AI safety while privately racing to accelerate. The decoupling between rhetoric and action isn't a failure of diplomacy — it's the strategy.
The Two-Speed Race
Before Donald Trump and Xi Jinping meet Wednesday in Washington, the expected headlines will be thin on AI substance. Expect a dialogue. Maybe a notification mechanism for national security incidents. What will not appear on any agenda is the coordinated slowdown that scores of researchers and industry leaders are now publicly demanding.
That absence is not an oversight. It is the point.
The United States and China, which together host nearly all frontier AI labs and most downloadable models in the world, have each firmly rejected calls to decelerate. The gap between their safety rhetoric and their competitive actions has opened into something wider than a policy disagreement — it is a structural feature of the current moment.
America’s Dismissal
Trump has been blunt. He called AI safety fears a “conspiracy” at the United Nations General Assembly and warned Washington would oppose any “globalist scheme” to control the technology. His framing is straightforward: AI dominance is inseparable from American power, and any limit on development is a limit on US leadership. He told reporters on Tuesday that guardrails were unnecessary under his administration.
That position places him at odds with some of his own country’s most prominent AI builders. Anthropic CEO Dario Amodei and OpenAI CEO Sam Altman have both urged slower development and stronger regulation. Nvidia’s Jensen Huang and Meta’s Mark Zuckerberg have pushed back, emphasizing AI’s benefits over its risks. The sector is split, and Washington has picked a side by doing almost nothing to regulate either.
China’s Calculus
China’s approach is more textured but no less committed to acceleration. This month, Beijing released the third iteration of its AI security framework — a move that signals attention to governance without any appetite for slowing output. Chinese officials have acknowledged risks around loss of control, misuse, and cyber threats. But they have also dismissed Silicon Valley’s existential-risk alarms as “fearmongering.”
The Chinese Ministry of State Security issued a rare essay this month warning that “hostile forces” could weaponize AI to spread political rumors and incite unrest. The threat China feels is not runaway machines but rival powers exploiting the technology. That distinction matters. It means Beijing sees a slowdown not as shared protection but as strategic containment.
Sun Chenghao at Tsinghua University put it directly: China worries that “AI safety” language will become the justification for restricting its access to chips, computers, and markets. If every safety discussion is tied to export controls, cooperation loses its appeal.
Who Wins, Who Loses
The immediate winners are the companies building these systems. Frontier labs in both countries continue to pour resources into increasingly capable models with fewer external constraints. Investors benefit. Workers in the AI sector see rising demand. Consumers get faster feature rollouts.
The losers are harder to name but easier to predict. They are the people who will inherit systems they did not help design and cannot fully control. They are the societies where AI-fabricated disinformation, autonomous weapons, or misaligned systems cause harm before any safety framework catches up. And they are the global population that bears existential risk — however unlikely some officials call it — while the superpowers bet that the risk is someone else’s problem.
Alvin Wang Graylin at the Asia Society Policy Institute captured the asymmetry plainly: a safety regime covering only one side’s models covers half the problem. The other half is built in a rival capital with rival priorities.
The Verification Trap
Even if both governments wanted a deal, the path is blocked by a technical impossibility they cannot easily admit. Wang Graylin noted that verification is hard for software you cannot see. Models can be copied, fine-tuned in secret, and distributed across jurisdictions. A promised slowdown leaves no clear audit trail.
Otto Barten of the Existential Risk Observatory argued that the critical red line is recursive self-improvement — the point where AI systems begin designing better versions of themselves without human supervision. He called for an explicit agreement that this threshold will not be crossed. But he also acknowledged that any deal is unlikely given the incentive for a rule-breaking party to shift the global balance of power.
This is the core dilemma. The very capability that makes AI useful — self-improvement, rapid learning, adaptability — is the same capability that makes it dangerous without oversight. And the very competition that drives innovation is what makes coordinated oversight nearly impossible.
A Mismatched Threat Landscape
Henry Gao at Singapore Management University identified the deeper issue: Washington and Beijing are not even discussing the same risk. American safety debates center on highly capable AI escaping control and threatening humanity. Chinese concerns are geopolitical — that the US will gain a decisive advantage and use it.
These are not complementary frameworks. They are opposing ones. From Beijing’s perspective, slowing development cedes ground to Washington. From Washington’s perspective, slowing development invites Beijing to catch up unchecked. Neither side can trust the other to stop first. So both accelerate.
Rui Ma, a tech consultant and founder of Tech Buzz China, argued that the focus should be on mitigating existing risks rather than pausing development outright. He called for evolving safeguards, open communication channels, and constant monitoring. It is a reasonable position — if you believe the technology can be managed faster than it can outrun management.
What Happens Next
The Trump-Xi meeting will produce a dialogue. That is something. Jiang Tianjiao at Fudan University called it a positive signal and expressed hope that the two sides can find common ground despite geopolitical rivalry. Major powers do bear responsibility for preventing失控 — out of control — systems, as Jiang put it.
But responsibility without enforcement is performance. A notification mechanism is useful until the first crisis arrives and neither side trusts the other to use it honestly. A dialogue is meaningful only if both sides believe the alternative is worse.
Right now, neither side believes that. The US sees acceleration as essential to maintaining global leadership. China sees it as essential to sovereignty and security. Both dismiss the other’s safety concerns as strategic cover. Both answer expert warnings with more compute.
The dangerous new dynamic is not merely competition. It is coordinated acceleration — two superpowers racing in opposite directions while pretending to share the same destination. That is the pattern worth watching. Not what Trump and Xi say on Wednesday, but what they build next year.