About the proposed AI-policy framework for Aotearoa New Zealand, published on agenticgovernance.digital. Source references [S-n] point to the published Source Register; gap references [G-n] to the published GAP register. Both accompany the framework.
A proposed AI-policy framework for New Zealand: what the country should do about the AI already running in its hospitals, classrooms, councils and workplaces. It covers who holds authority over public-sector AI, te Tiriti and Māori data sovereignty, jobs, data protection, electricity and water, compute and hardware, security, and how to pay for it. It was produced by a deliberation whose full record is cryptographically sealed and independently checkable (§14 of the framework).
No. It is a proposed framework offered for public discussion — not the policy of any party, and not the position of any organisation. It says so on its face. It is also a demonstration: of what a policy document looks like when every claim is sourced, every unknown is declared, and the whole process of producing it can be audited. It is offered to every party, and to everyone outside the parties.
An AI system, and the document says so plainly. The framework was produced by a simulated deliberation run inside a working deliberative platform: twenty-eight fictional composite members, organised into ten policy teams, arguing over real published sources — MBIE, Transpower, NZTech, Te Kāhui Raraunga, the Privacy Commissioner and others [S-10 through S-24]. Every factual claim carries a source reference; every figure New Zealand does not have is declared as a gap rather than invented. No words were placed in the mouth of any real person.
The members are composites; the process is not. The teams ran genuine multi-turn arguments, took formal votes at decision points, cross-checked each other’s positions, and ratified every conclusion by an anonymous all-member vote — and those events changed the outcome. One team’s expansion plan was formally blocked by another team and rewritten. One team’s conclusion was sent back by the full membership as too shallow, rebuilt, and re-ratified. One disagreement was kept on the record rather than smoothed into false consensus. All of it is sealed in a verifiable record (§14). The framework is labelled as a demonstration everywhere it appears, so nobody is being passed a simulation as a citizens’ assembly. What it demonstrates is a method: policy you can audit.
The deliberation engine and the drafting tool are different things, and the distinction is the framework’s own argument. The deliberative platform — where the arguments ran, the votes were cast, and the records were sealed — runs on models hosted in New Zealand, under New Zealand jurisdiction. A frontier model was used as a writing tool for prose. That is exactly the class-based approach the framework proposes (§6): sort the work into classes, decide which class needs to stay under domestic authority, and apply the safeguards where they carry weight. The votes, the record and the deliberation stay home; the typesetting does not have to. A blanket “never touch offshore tools” rule is the residency absolutism the deliberation itself put to a vote and rejected.
Nobody, and the chapter says so. Section 4 opens with a proviso: it draws only on Te Kāhui Raraunga’s own published Māori Data Governance Model [S-22] — co-designed with sixteen Crown agencies in 2023 — and is offered for Māori-data leadership to review and shape, not as a settled position. No tikanga position was fabricated; the te Tiriti strand argued only from published Māori-led sources. The chapter’s substance is that the Crown should implement what it already signed: the Māori Chief Data Steward the Model itself proposes, and appropriated multi-year funding for the eight Pou, as a non-severable pair. Māori data sovereignty is for Māori to determine; the framework’s job is to make space for it in the architecture.
You can check it. Every deliberation record — team threads, decision points, the cross-check block and revision, the ratification return, the final tallies — was cryptographically sealed at the moment it was made. Sealing means the record cannot be quietly altered afterwards: change one character and the verification fails. The sealed records can be verified independently of the platform that produced them, using the walkthrough published with the record set. The claim is deliberately stronger than “trust us”: you can confirm the positions in the document are the positions the deliberation actually produced, without trusting anyone.
Too small for some things, and the framework says which. No domestic chip fabrication — none plausible at this scale [S-21][G-1]. No frontier-scale compute — the national research ceiling is about 1.4 petaFLOPS, three to four orders of magnitude off the frontier [S-20][G-2]. The framework builds nothing on those fantasies. But small countries are not too small to decide whose court order their public systems answer to, what happens to information about their people, or which loads their grid connects on what terms. Denmark — a nation of comparable size — runs sovereign public-sector AI on home-or-EU-jurisdiction compute without pretending to race the frontier [S-1]. Sovereignty at this scale means custody, access and jurisdiction, not owning the frontier.
The framework publishes no figure nobody has costed. Phase 1 — the procurement tests, the register, the audit and data-handling standards — is regulatory, carried by existing machinery, and cheap by design [S-23][S-24][S-16]. Phase 2 adopts the cross-party proposal’s sourced envelope: low tens of millions over three to four years for a national capability fund, benchmarked against Canada’s compute access fund (up to CA$300 million) and Singapore’s Enterprise Compute Initiative (up to S$150 million), with the precise figure reserved for Treasury and MBIE [S-1]. Two items are named as appropriations bids to be costed in government: eight-Pou implementation [S-22] and the Māui replacement, which is due on ordinary research-infrastructure grounds anyway [S-20]. A costing you cannot check is a hope, not a budget.
Not if they are visible and conditioned — and right now they are neither. Transpower’s load forecasts include no hyperscale data centres; roughly 32 GW of prospective load is queued for connection; and north-Auckland grid exit points are overloaded by 2028 without upgrades [S-12][G-8]. A single hyperscale AI facility draws 300–600 MW continuously — 3 to 6 per cent of the installed fleet, committed around the clock [S-12][S-11]. The framework’s answer is a gateway, not a wall: mandatory load disclosure into Transpower’s forecasting, interruptibility as a standard connection condition, siting steered away from the 2028 constraint, and true-cost connection terms (§7). Firm demand is also what gets new renewable generation financed — at 85.5 per cent renewable and rising [S-11], the marginal electron here is among the cleanest available.
The entire public evidence base is one consent: Datagrid’s Southland site, consented to draw 220 million litres of groundwater a year for cooling [S-14]. No disclosure standard exists for any facility [G-7]. The framework’s position is that water is a catchment question, not a national one — 220 million litres is trivial nationally and potentially significant against one aquifer. So: per-facility disclosure, water-free or closed-loop cooling as the reference technology for consent assessment, and in fully or over-allocated catchments, no new industrial cooling takes at all — build dry or build elsewhere (§7). Microsoft’s Auckland region already runs water-free cooling [S-15]; taking nothing is a proven engineering choice.
No — and the deliberation voted that reading down explicitly. The commitment is that sensitive information should not have to leave the country, or its legal protection, to be useful: a right to domestic processing, not a prohibition on movement [S-1]. The instrument is a class-based handling standard — sort information into classes, then set rules per class (§6) — with a protected pathway for ethics-approved, de-identified research data, because a blanket onshore mandate would relocate health research offshore along with the researchers [S-16][S-18]. Māori data takes its classification from the Māori Data Governance Model, as a binding input [S-22].
New Zealand does not measure AI-attributable displacement, and the framework refuses to invent a forecast [G-5]. What it does instead: a workforce-transition disclosure standard so the next policy cycle argues from evidence; a redeployment fund that moves people into better work; a training levy-credit that pays out only against portable micro-credentials, so employers stop solving skills shortages purely by importing — 45 per cent of the tech workforce is on work visas while domestic enrolments have fallen 33 per cent since 2010 [S-19]; and a retraining pipeline built for employed adults, not just school-leavers (§5). The income floor under a fast transition is named as a whole-of-welfare-system question, not smuggled in as an AI line-item.
On its evidence, not its authorship — which is the test any policy document should face. Every figure is sourced to a published register; every missing figure is declared rather than invented; every position survived argument, cross-checking and an anonymous ratification vote; and the whole record is sealed and checkable. No pamphlet, manifesto or ministry strategy in New Zealand currently offers that. If the document is wrong somewhere, the sourcing makes it checkable — which is more than can be said for most documents written by people.
Three things. Read it — the executive brief is one page if eight thousand words is too many. Check it — the sealed record, Source Register and GAP register are published alongside, with a verification walkthrough. And take what stands: the framework is offered to every party and to the select-committee process the cross-party proposal points at [S-1]. The deeper ask is about method — that New Zealand start expecting policy that shows its working, from anyone who claims their policy was built on evidence.
This Q&A accompanies “Our Own Machines”, an AI-generated demonstration framework produced by a simulated deliberation of fictional composite members, offered for public discussion. It is not the policy of any party, and not the position of any organisation.