Apple's China AI licence, and per-country rules for how AI gets used
Updated August 15: re-pointed the source link from the Daring Fireball post to the South China Morning Post report it cited.
South China Morning Post · source ↗
Chinese regulators have granted Apple a licence to ship Apple Intelligence on iPhones in China, with Alibaba’s Qwen and Baidu as the technical partners — reported by the South China Morning Post (Alibaba-owned), alongside approvals for six other smartphone AI services. The same feature runs on Apple’s own models across most of the world and on state-approved domestic models inside China. But the licence isn’t really about which model runs. Approving the model is how the state governs what AI may be used for inside its borders — the allowed-model list is a control point for usage, not just a supplier choice.
That is the thing the telephone network has lived for a century. A telephone number was never only an address; it came with rules about what you may do with it and who may put traffic on the network. Numbers are assigned under conditions, permission to originate is licensed, and usage is governed and enforced — telemarketing limits, robocall prohibitions, do-not-call, know-your-customer obligations, lawful-use requirements — backed by a national apparatus that mandates call authentication, requires robocall mitigation, runs traceback, and blocks or penalizes the traffic that breaks the rules. The network is a governed space; permission to use it for a given purpose is granted, conditioned, and policed at the country level.
AI provisioning is arriving at the same place from the other direction. By controlling which models are allowed, a jurisdiction controls what usage is permitted — and gains the leverage to enforce policy at the layer everyone relies on. The interesting questions stop being about the feature and become the governance ones telecom already asks: what usage is permitted or prohibited here, who is accountable for enforcing it, how you prove that only sanctioned usage actually happened (the attestation problem STIR/SHAKEN answers for calls), and what becomes of a user who moves between regimes whose rules disagree.
The feature was never the story; the usage-governance regime behind it is — and telecom has been deciding what its network may be used for, and enforcing those decisions number by number, for a hundred years. Worth watching whether AI builds that governance-and-enforcement layer deliberately, or bolts it on the way the phone network bolted on everything that came after voice.