The People’s Republic of AI

01 / FORTY-EIGHT HOURS

A diplomatic signature became a market opening.

On July 17th, Indonesia’s coordinating minister for economic affairs, Airlangga Hartarto, met Huawei and ByteDance in Shanghai and invited them to deepen their cooperation with his country.

He was explicit about why. Indonesia had just become a founding member of a brand-new organization called the World AI Cooperation Organization, he said, and he wanted that membership to become a springboard for a new stage of tech partnership — pointing specifically to Huawei’s strength in AI, telecom infrastructure, and cloud computing as a strategic asset for Indonesia’s digital transformation. The following day, reporting from Jakarta confirmed that Indonesia intended to deepen its AI cooperation with China on top of the existing scale.

Two days. That’s how long it took for a diplomatic signature to turn into a market opening. Hartarto had signed the founding agreement of the new organization in Shanghai on July 16th. By July 18th, he was publicly steering Indonesian industrial policy toward tech giants heavily backed by the Chinese government. If you wanted a single data point for how China converts governance architecture into commercial reach, this is it — and it happened in less time than it takes to recover from jet lag.

02 / CONFERENCE AND CONTRACT

The deal was done before the cameras arrived.

The World Artificial Intelligence Conference (WAIC) ran in Shanghai from July 17th to July 20th this year — 140-plus forums, over 1,100 exhibiting companies, more than 3,000 exhibits, over 100,000 square meters of floor space, the biggest edition in the event’s nine-year history. Nine Turing and Nobel laureates showed up, including reinforcement-learning pioneer Richard Sutton and deep-learning godfather Yoshua Bengio, who used his slot to help unveil a UN framework for AI governance.

But the real business happened the day before the cameras started rolling. On July 16th, foreign minister Wang Yi signed, on China’s behalf, the agreement establishing the World AI Cooperation Organization (WAICO) — a fully independent intergovernmental body, headquartered in Shanghai, with 29 founding member states. The signing ceremony wasn’t livestreamed. That evening, Xi Jinping and his wife hosted a dinner for the assembled delegations. Only the next morning did Xi take the stage for the conference’s opening keynote. The order of operations tells its own story: the deal got done in private first; the conference was the unveiling, not the negotiation.

Look at who actually signed. Twelve to fourteen Asian states, mostly Central Asian and Southeast Asian partners, plus Russia. Ten African countries. Four in the Americas, including Venezuela, Cuba, and Nicaragua. Precisely one European country: Serbia. Not a single G7 economy. Not one EU member. WAICO’s own stated purpose, in the government’s own language, is to help Global South countries accelerate their AI development and close what Chinese officials call the “intelligence gap” — capacity-building, not frontier competition. This is an organization built, by name and by membership, for everyone outside the American-led AI order.

03 / WHAT XI ACTUALLY SAID

The sharpest security critique pointed outward.

Xi’s keynote carried the title “Working Together to Build a Fair and Just Global AI Governance System,” and it laid out four principles: open, mutually beneficial innovation; heightened risk-awareness with safety guardrails; inclusiveness that protects civilizational diversity; and a call to strengthen multilateral governance through the UN.

The second point contained the speech’s sharpest line, even if it never named a country. Xi called for a joint rejection of any attempt to stretch the concept of national security to cover AI, and specifically criticized placing one nation’s security above everyone else’s. Anyone tracking the last two years of chip export controls knows exactly which government that sentence was aimed at.

It’s a real critique. It’s also one that’s awkward to make from Beijing. China’s own AI governance runs on a security logic just as expansive as the one Xi was criticizing — it’s simply pointed inward instead of outward.

04 / TWO WAYS TO RAISE AN AI

Build first, litigate later. Or certify before birth.

Here’s where the comparison gets concrete instead of rhetorical. The United States regulates AI the way it regulates most new technology: build first, litigate later. There is still no comprehensive federal AI law. Enforcement runs through existing statutes and after-the-fact litigation — the FTC pursuing deceptive practices, the EEOC pursuing discriminatory hiring outcomes under civil rights law — and only after a harm has already occurred or a complaint has been filed. States have stepped into the vacuum with a patchwork: California’s transparency rules for frontier models, Colorado’s disclosure requirements, New York City’s mandate that hiring algorithms undergo a bias audit only after a year in production. And in December 2025, the White House moved in the opposite direction of “more oversight” — an executive order aimed at challenging state AI laws in court and tying federal funding to how burdensome a state’s rules are judged to be. The single biggest federal AI move of the past year was an attempt to weaken regulation, not build it.

China runs the opposite model: certify before birth. Any generative AI service with what regulators call “public-opinion properties or social-mobilization capacity” must clear a security assessment and file its algorithm with the Cyberspace Administration of China before it can legally launch — not after a scandal, before day one. Call it submitting your brain for inspection. As of this spring, more than 860 generative AI services and over 500 downstream applications had gone through this filing process. Separately, a new rule on content traceability requires AI-generated material to carry both a visible watermark and an invisible embedded marker — a double-layered paper trail that exists alongside, not instead of, the algorithm filing. And output has to align with what the rules call socialist core values — a compliance category with no real Western equivalent, best translated as: stay close to the wall of official values, or don’t get published at all. A brand-new rule regulating “anthropic-style interactive AI” — companion bots and persona-driven assistants — took effect on July 15th, two days before WAIC opened, requiring a fresh safety review once a service crosses a million registered users or 100,000 monthly actives, and banning any product designed to substitute for human relationships or induce dependency.

Both systems claim to protect the public. But only one of them treats every model as guilty until licensed.

05 / THE PRESUMPTION OF GUILT

The range of possible output is narrowed before the sentence exists.

That’s the deeper point buried under all the compliance paperwork. China’s censorship regime has always extended from what a person can say to what a platform can publish. What’s new is that it now extends one layer further: to what a person and an AI system can produce together. The logic isn’t “we’ll catch harmful output after it spreads” — it’s “we don’t trust what this pairing might generate, so we’re locking down the range of possible outputs before a single sentence exists.” That’s a presumption of guilt applied to a conversation that hasn’t happened yet.

Contrast that with Xi’s own words about not letting one country’s definition of security override everyone else’s. The principle isn’t wrong. It’s just being applied selectively — outward, toward chip controls and export bans, and never inward, toward the country’s own content architecture.