AI & Algorithm Governance

China's AI policy in mid-September 2026 stays promotion-first, and the promotion layer just gained a new, quantified instrument: MIIT's dedicated action plan to embed AI across the software and IT-services industry ¹. The control layer is unchanged — no new content or security rules, only the familiar "AI for good"/anti-misuse framing. What moved is the international layer, now a two-front attribution fight: foreign firms and agencies are naming Chinese models and China-linked actors as threats, while Beijing answers with self-reliance claims and "innovation dividend" numbers ^2,3.

The MIIT software plan is the clearest new evidence of how China operationalizes "AI+". It sets a two-stage roadmap: by 2028, raise the intelligence level of the software industry by cultivating high-standard coding tools and development platforms and pushing them into more than 20,000 major software enterprises, plus 100 benchmark applications for agent-based software across key industries ¹. That is the same ministry-led, target-driven pattern used for AI+manufacturing, now extended from factory floors to the software stack — coding assistants, agent platforms and tooling that firms will be pushed to adopt.

The surrounding promotion architecture keeps compounding. MIIT Vice Minister Xie Yuansheng said the ministry will deepen its industrial-internet innovation project on a "unified layout, scenario-driven, open-source" basis for AI+manufacturing at the Shenyang conference ⁴; SASAC's Pang Xiaogang said the commission will draft central enterprises' 15th Five-Year Plan to push SOEs into emerging industries ⁵. Liaoning is the showcase — 18 national excellence-level smart factories, 1,312 basic and advanced smart factories, and a first-batch national computing-power interconnection node ⁶ — and the compute geography keeps widening, with Guizhou's unique power-grid-inclusive consortium and Heilongjiang's China Mobile 10,000-card cluster ^7,8.

Control remains quiet and framed as misuse-prevention. The only governance-adjacent instrument in this batch is the 15th Five-Year Plan for IP protection, which refines rules for algorithms and AI-generated content ⁹; the MFA's stock line is that AI should be guided "toward good" and abuse prevented — said when asked about a US firm's claim its AI hacker tool could hijack WeChat accounts ². Beijing's security reflex is to treat AI risk as someone else's misuse, not to tighten its own model rules.

The international layer is where the temperature is rising, and it has split in two. One is the standards/certification contest: the US Embassy in Panama is piloting a global AI supply-chain certification and traceability platform covering semiconductors, critical minerals and AI-infrastructure goods, and Guo Jiakun called closed "small circles" harmful to the Global South while offering to build a fair AI governance system with Latin America ¹⁰. The other is direct attribution against Chinese models: CISA accused DeepSeek and Moonshot of stealing US large-model technology ², and on Sept 11 Mao Ning was pressed on Anthropic's report that it banned accounts allegedly linked to the Chinese government using Claude to gather information on overseas religious groups, Uyghurs in Syria and diaspora communities ³. Her answer — "I don't know the specifics" plus "firmly oppose distorting facts and smearing China" — is the same defensive template, now covering both alleged state-linked intelligence use and corporate technology theft.

China's counter-narrative leans on cooperation and measurable output. Assistant Foreign Minister Liu Bin told the fourth AI Capacity Building Seminar (30-plus countries, Peking University) that capability building must give developing countries autonomous capability plus governance skills, not one-way grants ¹¹; firms pitch "token export," exemplified by Unisound serving Ethiopian cross-border e-commerce from domestic U2-model compute with 10 billion tokens in week one ¹²; and Mao Ning cited WIPO's Global Innovation Index, where China holds 25 of the top-100 innovation clusters for a fourth straight year, with "Shenzhen-Hong Kong-Guangzhou" first ³.

The read for outside observers: China keeps buying capacity, goodwill and now software-sector productivity — reaching for plans, IP law and standards rather than content bans wherever governance pressure appears. Three pivots to watch: whether the AI+software plan's 2028 targets get funded and enforced ¹; whether the Panama certification scheme hardens into bloc-wide standards ¹⁰; and whether the attribution fight (CISA, Anthropic) escalates into procurement or export restrictions targeting named Chinese model firms rather than rhetoric alone ^2,3. Compute-node expansion and factory and software targets keep demand for chips, power and industrial data climbing.