China Has Built a Mirror for Every Layer of Washington's AI Strategy

The country Washington frames as the safety risk is now out-safety-ing Washington. In March, China's defense ministry warned of a "Terminator"-style dystopia if the United States keeps pushing military AI without human control ^[1].

using AI as a tool to violate the sovereignty of other nations, allowing AI to excessively affect war decisions, and giving algorithms the power to determine life and death not only erode ethical restraints and accountability in wars but also risk technological runaway. — Jiang Bin

A month later, the Trump administration exempted open-weight models — including Chinese rivals — from government safety testing, over the objections of the labs that build them ^[2].

A regulatory regime would not only strangle growth, development and innovation, and be enormously harmful to the industry, but it would be obsolete 48 hours after it was going through whatever process it had gone through. — Sean Cairncross

The timing is the point. Beijing has spent the year turning Washington's own safety language into leverage, and it is doing so at the precise moment Washington is walking away from that language. The safety layer is where the inversion is cleanest. When the two sides sat down to talk, Beijing set preconditions: American companies would have to prove they meet the same safety, disclosure, and audit rules China imposes on its own firms, and it accused Washington of using "AI safety boundaries to contain China's technological rise" ^[3]. The country that spent years being lectured about safety now writes the syllabus, and the lecturer has exempted its own models from the test. The software layer is where the mirror turns a physical wall into a sieve. Chinese open-weight models have passed American counterparts in downloads on Hugging Face, driven by DeepSeek's releases ^[4]. And there is no legal way to stop it: the First Amendment protects the right to receive foreign information, and an open-weight model can be downloaded and run on a private server, beyond any export control ^[5]. Chip restrictions can throttle the hardware; they cannot touch free software, which moves at the speed of a download. The diplomatic layer is where the mirror is most visible. China co-chairs a UN group on AI capacity-building with Zambia, explicitly framing the American approach as making AI "the preserve of rich countries or wealthy elites" ^[6].

It is essential that all countries, especially those from the Global South, can effectively participate in global governance, and that AI does not become the preserve of rich countries or wealthy elites. — Fu Cong

BRICS+ is building a sovereign AI stack meant to "decouple from Western cloud infrastructure," accusing the US of using safety guardrails to enforce "a shallow regime of artificial homogeneity" ^[7]. And Beijing has partnered with the African Union on a governance plan pitched as reducing Western dominance ^[8]. Each of these is a different door into the same room: a multilateral framework that lets the rest of the world adopt AI without choosing Washington's binary. The rival coalition is not seamless. When the US-Israeli air campaign against Iran tested BRICS+, the bloc split — India's chair stood with Israel while Russia, China, and South Africa condemned the strikes ^[9].

India stands with Israel, firmly, with full conviction, in this moment and beyond. — Narendra Modi

And some governments have barred DeepSeek from official systems over Chinese state censorship ^[10]. But the fracture runs along security lines, not AI governance. The same members who cannot agree on Iran still signed the sovereign AI stack. The coalition behind the alternative framework does not need to agree on Iran to agree on AI — and that is exactly what makes the diplomatic layer the one Washington cannot seal.