All editions · Policymaker Edition

Your Community, Your AI — CC BY 4.0

The Policy Opportunity — A Menu, Not a Manifesto

How to Read This Article

The first four articles were educational and non-partisan, and this one stays that way. What changes is that it becomes practical: it sets out the levers a government actually has. To do that, it draws on a specific worked example — In Our Own Hands, a draft AI-sovereignty policy proposal published for public discussion. That proposal is offered by its author as an independent contribution to public debate. It is not a party document and not a product; no party, official, or individual has endorsed it; and its author has declared a commercial interest in sovereign AI infrastructure. It is presented here as a menu, not a manifesto — a set of options any party could draw from, in whole or in part, without surrendering what makes it distinct. (Any unfamiliar term in this series is defined in plain language in the glossary.)

The organising idea is affordability of a particular kind. Owning the chips, the data centres, and the frontier models is out of reach for a small economy. Keeping authority, custody, and control over records and decisions is not — it is mostly a matter of rules and standards, and it is available to any government willing to write them. Most of what follows repurposes machinery that already exists.

The worked example is Aotearoa New Zealand's; the levers are not tied to it. Procurement rules, public registers, capability funding, and standards participation exist in some form in every system of government. Read each one below as a mechanism to translate into your own — the instrument's name will differ, the move it makes will not.

Seven Cross-Party Commitments

The proposal frames a set of foundations it argues any party could stand on together — common ground, stated as commitments rather than mechanisms:

1. Authority stays here. Public services and public data remain under domestic legal control.
2. People move up, not out. AI is used to lift workers into better roles, paired with redeployment support, rather than to displace them.
3. Sensitive information stays home. Commercial and personal data remain within domestic jurisdiction and legal reach.
4. Te Tiriti and Māori data sovereignty are respected — with the substance deferred to Māori-led frameworks (see below), not prescribed by the Crown.
5. People decide, machines assist. Consequential decisions stay with accountable humans.
6. What public AI does is checkable. Government AI decisions leave auditable records.
7. We build our own capability, with others — domestic capacity developed through partnership, not protectionism.

These are deliberately non-mechanical. Their value to a legislator is that they can be affirmed across the aisle before anyone argues about instruments — a starting point for cross-party durability rather than a policy anyone has to own alone.

Procurement: The Cheapest Lever

The single most powerful lever a government already holds is what it buys, and how. Two tests, applied through existing procurement rules, do most of the work:

• A jurisdiction-of-inference test. Where does a model actually run, and whose law reaches the company that controls it? (This is the sovereignty question from Article 2, converted into a procurement criterion.)
• A data-residency test. Do sensitive data — and anything derived from them — stay within domestic legal reach, not merely on domestic soil?

Agencies then buy from a pre-vetted list of suppliers that meet both tests. Note what this is not: it is not a tariff, a ban, or a preference for any country's vendors. It is supplier-agnostic — any supplier, foreign or domestic, is free to meet the tests. The decision stays with the country rather than the vendor, which is exactly the point Article 2 argued.

A Provenance-and-Audit Standard, and a Public Register

Article 3's argument — that a principle written as policy can drift, while a proof chain cannot be silently rewritten — becomes concrete here. A provenance-and-audit standard would require that significant government AI systems keep tamper-evident records of what they decided and why: append-only, independently verifiable, so that a decision can be reconstructed after the fact without having to trust the vendor's own account of it.

Paired with it, a public register of significant government AI systems records where AI is used in public administration, at what level of consequence, and under what oversight. This is the transparency commitment made checkable rather than promised. For a legislator it also answers a question that will otherwise be asked in the House sooner or later — where, exactly, is the government using AI, and who is accountable for each use? — with a document rather than a scramble.

A National AI Capability Fund

Every lever above is a rule and costs little. The one substantive spending item in the proposal is a National AI Capability Fund — modest, phased, and benchmarked against comparable small-economy programmes rather than pitched at frontier scale. The proposal describes it in the low tens of millions over three to four years; a legislator should treat that figure as the proposal's own estimate, to be tested, not as a settled number. Its three streams:

• Subsidised compute access for small businesses, councils, not-for-profits, and primary-sector operators that could not otherwise adopt AI safely.
• Advisory support for safe adoption.
• Community- and sector-scale models that keep sensitive work onshore.

It is explicitly paired with redeployment funding, so the "people move up, not out" commitment has money behind it rather than words alone. The sequencing matters: the procurement rules (which cost little) come first and create demand; the fund follows the demonstrated demand. Spending trails need rather than leading it.

A Reversibility Requirement

One provision deserves particular attention from anyone who has watched a public IT programme become impossible to unwind. The proposal would build reversibility into publicly funded AI at two levels:

• At the project level, every funded pilot has explicit stop conditions and can be unwound if the evidence turns negative. A pilot that cannot be stopped is not a pilot; it is a commitment wearing a pilot's clothing.
• At the system level, open standards (such as decentralised identifiers and the ISO/IEC standards track) are preferred over a single central registry, so that a failure in one authority does not collapse the whole.

For a policymaker, a reversibility requirement is close to free insurance: it costs little to write in and preserves the option to change course when — not if — some pilots disappoint.

Alignment with the EU AI Act — as Trade Policy

The proposal treats interoperability with EU frameworks as alignment as trade policy, not as deference. The reasoning is straightforward: aligning with the EU AI Act (Regulation 2024/1689) where it counts protects market access for exporters and spares firms a dual-compliance burden, without adopting the entire EU regime wholesale. A legislator can present this as economic self-interest rather than regulatory imitation — which is both more accurate and more durable politically.

Deeper Standards Participation

Related, and among the cheapest sovereignty available: deeper engagement with the international standards that set the global ground rules for AI — the ISO/IEC JTC 1/SC 42 track, covering terminology, the machine-learning lifecycle, risk management, and AI management systems (ISO/IEC 42001). A small country already participates on a one-country-one-vote basis; deepening that engagement buys a seat at the table where the rules everyone will eventually follow are written. The cost is a delegation's time; the return is influence disproportionate to size.

Te Tiriti and Māori Data Sovereignty — Deference, Not Prescription

This is the area where a policymaker most needs to be careful about who is speaking, and the right posture is deference. The proposal does not author new consultation or speak for Māori, and neither does this briefing. Instead it defers to frameworks that are already Māori-led — among them Te Kāhui Raraunga, the CARE Principles for Indigenous data governance, and the Kaitiakitanga Licence — and commits that indigenous data stays within domestic jurisdiction and is treated as taonga requiring consent-based governance.

Two cautions for a legislator. First, this is a matter to be developed with Māori-led bodies through their own instruments, not resolved for them in a Crown policy document; presenting the Crown as having settled it would be a mistake of both substance and process. Second, there is a genuine tension a policymaker should name rather than paper over: the Crown's interest in a single national AI framework can sit uneasily against Article 2's protection of rangatiratanga over taonga, including data. The right move is to flag that tension and leave its resolution to the appropriate te-Tiriti processes — not to declare it resolved.

What a Legislator Can Take From This

The value of reading the proposal as a menu is that no single item commits a party to the whole, and most items cost little because they reuse machinery a government already has. The procurement tests, the provenance standard and register, the reversibility requirement, and deeper standards participation are largely regulatory assembly. The capability fund is the one real appropriation, and it is modest and phased. And the whole set is supplier-agnostic and cross-party by design: it names no favoured vendor and asks no party to surrender its distinctiveness.

Whatever position you ultimately take, the four articles before this one were meant to let you take it clearly — to say what AI is and is not, why where it runs is a question of state, why principles need architecture to hold, and what the technology actually does today. That clarity is the point. The menu is here to be marked up, argued with, and improved. It is offered in exactly that spirit — as a draft for discussion, not a position to be adopted on anyone's say-so.

Want to use AI tools like these well, and safely? Our free courses — Working with Claude and Agents at Work — teach the practical skills, from getting trustworthy answers to deciding what to hand an agent. For the full technical architecture behind Village AI, see Village AI — Agentic Governance.