AI Policy Goes Global: Export Controls, Consumer Law, and State Promotion Reshape Who Controls AI

Taken together, these five stories show AI policy moving from abstract principles toward concrete instruments. China is drawing its export-control perimeter with unusual precision. MOFCOM's clarification says that most AI technologies remain freely exportable, and that what matters is whether restricted technology physically or technically leaves the country, not whether a model is served through an API or SaaS product. That is a significant interpretive choice: it treats the model as a delivered service rather than a shipped artifact, and it asks regulators to police the transfer of know-how and weights rather than the business model. The companion draft revision, which would add specific AI technologies to the restricted catalogue alongside biomedicine and photovoltaics and require pre-export registration, shows that the perimeter is being drawn in the same breath. The draft's silence on which AI technologies are covered is itself informative: the state is preserving discretion.

The U.S. export-control backdrop explains why this matters. The question of whether chip restrictions have slowed Chinese AI is contested, and the argument that they spurred domestic substitution in semiconductors suggests controls can produce second-order industrial effects that policymakers do not fully anticipate. Chinese models approaching top U.S. systems complicates any simple story of containment, and it pushes governments toward controlling deeper layers such as weights, training methods, and registration of capabilities.

At the same time, states are not only restricting AI but actively promoting it. India's working paper on the Consumer Protection Act, 2019, and China's Ministry of Culture and Tourism pilot programme point to a dual posture: encourage deployment in sectors such as culture and tourism while building accountability frameworks for harms. The India analysis is especially useful. Its technology-neutral definitions may already reach biased outputs, psychological harm, and loss of control, yet the statute assumes discrete manufactured products with identifiable defects. Generative and agentic systems are continuously updated, probabilistic, and shaped by many parties, so liability attribution becomes the hard problem.

The tension running through all five is a mismatch between where the law locates control and where AI actually resides. Export rules key on movement of files and know-how; consumer law keys on products and defects; promotion programmes key on deployment. Each framework captures part of the system and misses the rest.

Signals to watch include the final text and list of AI technologies in China's catalogue revision, whether registration requirements extend to model weights accessed remotely, how U.S. and allied controls respond to Chinese substitution, whether Indian regulators or courts accept that general-purpose AI fits the product-defect model, and whether pilot programmes publish outcome data. For technical readers, the practical lesson is that compliance will increasingly depend on documenting provenance, model access paths, and failure modes, not just on the product label.