Three weeks ago, a foreign export-control decision briefly cut foreign users off from Anthropic's most advanced AI models. Those are the same models now being woven into public services around the world. The block was lifted after diplomatic negotiation, but the incident has put a concrete edge on a proposition New Zealand's government has yet to formally test: any country that puts overseas AI into public services without owning the models is one policy decision away from the same blackout.
Francisco Hernandez, the Green Party's public-service spokesperson, used a parliamentary intervention last week to argue that New Zealand's rollout of AI across government has not answered the most basic sovereignty question. Who controls the model when the lights go out?
The Anthropic suspension, which Interest.co.nz reports was triggered by U.S. export-control and national-security restrictions and resolved through negotiation, is the kind of dependency risk New Zealand should treat as a planning assumption rather than an edge case. Anthropic's most advanced models were unavailable to foreign nationals for the duration. The lesson is not about Anthropic specifically. Any model operated by a foreign vendor sits inside that vendor's home government's jurisdiction. That is the mechanism the rest of the world should plan around.
Hernandez's argument is not a call to ban AI in government. It is a call to put safeguards in place before deployment, and to spell out, in advance, what the country will do when supply is cut. His specific list is short enough to be useful: auditability, mandatory human review, an opt-out that lets any member of the public request a human decision-maker, and a clear line of human responsibility for every output the model produces. Each item is a leverage tool. Used together, they give the public service the ability to refuse, redirect, or override a model that is no longer available or no longer trusted.
The data question is sharper. A review of the government's digital delivery agency raised data-sovereignty concerns after finding that a large agency was running AI tools hosted in Australian data centres, not New Zealand ones. For most of the public, that distinction is invisible. For Māori, it is not. Māori data sovereignty, the principle that data describing Māori communities should be governed by Māori, is one of the unresolved issues Hernandez named, and it does not survive a foreign-hosted model. The same point applies to any group whose information a public service has a special duty to protect.
The counter-argument is the one New Zealand's government is currently making: that guidelines written after deployment is already underway can keep pace with the technology. That is the position every government takes when it is behind on procurement. New Zealand's Public Service Commissioner Sir Brian Roche has publicly described the country as falling behind on AI and has tied that gap to broader questions of public-service accountability. The 272 active AI use cases already documented across the public service suggest deployment is well ahead of the framework meant to govern it.
The portable version of the question, and the one worth carrying past New Zealand, is whether the safeguards are written into the contract or promised in a press release. If a model can be cut off by a foreign government with no notice, then any procurement contract that does not give the buyer ownership, audit access, and a fallback plan is not a sovereignty plan. It is a subscription.
New Zealand's public-service AI framework is under active revision. The next draft is where the safeguards either land in the contract or stay in the guidelines. Hernandez's parliamentary profile does not suggest he expects to win that fight inside the current term. The framework is a public document, though. The next draft is where readers, in New Zealand and outside it, can read the answer.