Our Own Machines

A proposed AI-policy framework for Aotearoa New Zealand — produced by deliberation

• Offered for public discussion — not the policy of any party, nor the position of any organisation
• An AI-generated demonstration: 28 fictional composite members, 10 policy teams, arguing from real sources inside a working deliberative platform
• Every claim sourced [S-n]; every missing figure declared as a gap [G-n]; every position ratified by anonymous vote, cryptographically sealed

Speaker note: open by saying what this is not — not lobbying, not a manifesto. It is a demonstration that AI policy can be built in the open and checked.

The problem

• New Zealand has no AI policy — and, for the most part, neither do the parties asking to govern it
• Last in the OECD to publish a national AI strategy; the one it got is a strategy for adopting other people’s AI, not holding authority over it [Investing with Confidence, S-10]
• The technology is already in hospitals, classrooms, councils and workplaces — mostly on terms written somewhere else
• June 2026: a single overseas legal order disrupted a major AI tool worldwide; New Zealand was simply downstream [In Our Own Hands, S-1]

Speaker note: this is a foundations question, not a left-versus-right one.

The frame: custody, not competition

Clancy’s coupled-strands map — compute, energy, data, people, security move together — reframed for a country of five million:

• No sovereign frontier laboratory, and none proposed [G-3]
• No chip fabrication — none operating, none planned, none plausible at our scale [S-21][G-1]
• So the question is never “can New Zealand compete at the frontier?” It is: whose hands, whose terms, whose court order?
• Sovereign compute here means three things the country can actually own: custody, access, jurisdiction [S-1]

Speaker note: the strands stay coupled — every position in the framework binds to the others — but the ambition is sized to the country.

Policy you can audit

• Every position survived argument between participants who started apart
• One team’s plan was blocked by another team and rewritten; one conclusion was sent back by the full membership and strengthened
• Every decision point, vote and disagreement was cryptographically sealed at the time it was made
• The sealed records can be verified independently of the platform that produced them (§14)

Speaker note: no pamphlet, manifesto or ministry strategy in this country currently offers this. The collisions are the reason to trust the result.

The evidence picture: what New Zealand has

• Electricity most countries would envy: 43,879 GWh in 2024 at 85.5% renewable; a record 96.4% quarter in December 2025 [S-11]
• Technology-neutral privacy law that already reaches AI — the Privacy Act 2020 and its 13 principles, with AI-specific guidance issued twice [Privacy Commissioner, S-16]
• The Algorithm Charter (2020), 20+ signatory agencies [S-17]; the Data and Statistics Act 2022 [S-18]
• On Māori data, the hardest design work is done and Māori-led: the Māori Data Governance Model, co-designed with 16 Crown agencies in 2023 [Te Kāhui Raraunga, S-22]

Speaker note: good policy starts with an inventory, not an ambition.

What it depends on — and what it does not know

• The first hyperscale cloud region (Dec 2024) is US-owned through an Irish subsidiary — the racks are here, the control is not [S-15]
• Every GPU is imported; 45% of the tech workforce is on work visas; domestic digital enrolments down 33% since 2010 [S-21][S-19]
• 48% of workers have put sensitive information into public AI tools; 56% have used AI at work unsure it was allowed [S-1]
• Eight declared gaps [G-1 to G-8]: no AI procurement rule, no data-centre load in Transpower’s forecasts, no water-draw disclosure, no displacement data — stated plainly, never papered over with invented numbers

Speaker note: what the country does not know is itself a finding — and it shapes every instrument that follows.

Authority you can hold

Six years of voluntary instruments have delivered signatures, not checkability [Algorithm Charter, S-17]. The answer is a hybrid:

• Binding procurement tests, now — jurisdiction of inference + data residency, inserted into the Procurement Rules by executive act, this term [S-23][S-1]
• A public register of significant government AI — councils included from day one; they buy algorithmic systems outside every current instrument’s reach [S-23]
• An audit standard with a usability test — writable by a council with three IT staff; every automated decision ends in proceed, refuse, or escalate to a human [S-1]
• A statutory authority decided by Parliament, via select-committee inquiry — not asserted in this framework

Speaker note: rules first, Parliament decides the body. Procurement is the enforcement lever for everything else in the document.

Te Tiriti and Māori data sovereignty

Proviso. This chapter draws on Te Kāhui Raraunga’s own published Māori Data Governance Model [S-22]. It is offered in that spirit — for their review and for Māori-data leadership to shape, not as a settled position. Māori data sovereignty is for Māori to determine; the framework’s role is to make space for it in the architecture, not to define it.

• The foundation already exists and is Māori-led: eight Pou, co-designed with 16 Crown agencies, signed in 2023 [S-22]
• The Crown’s favourite delay — “we must first develop a framework” — is unavailable; the framework is developed and signed
• It is also, across most signatory agencies, unimplemented — the controlled experiment this chapter learns from [S-22]

Speaker note: endorsement at the top without appropriated implementation underneath moved almost nothing.

The non-severable pair

• A Māori Chief Data Steward with statutory basis — the office the Model itself proposes — in the room as of right [S-22]
• Appropriated multi-year funding for eight-Pou implementation, including data infrastructure iwi themselves control — not contestable pilot pools [S-22]
• Non-severable in drafting: no future budget delivers the office and defers the money, or funds the Pou and leaves the Steward on memorandum legs
• Classification before residency: Māori-defined data classification binds the national handling standard as input, not advisory context — upstream of every public AI purchase [S-22]

Speaker note: the ask is not for something new; it is for the Crown to inhabit what it already signed.

The physical ledger: electricity

• Transpower’s load forecasts include no hyperscale data centres; ~32 GW of prospective load queues for connection; north-Auckland exit points overload by 2028 without upgrades [Transpower, S-12][G-8]
• One hyperscale AI facility draws 300–600 MW continuously — 3–6% of the installed fleet, committed around the clock including winter peak [S-12][S-11]
• The position is a gateway, not a wall: disclose (into Transpower’s forecast), flex (interruptibility as a connection condition), locate (away from the 2028 constraint), pay true cost [S-12]
• The deliberation changed its own policy here: the electricity team blocked the compute team’s unconditioned staging; the staging was rewritten and re-passed — sealed in the record

Speaker note: firm, creditworthy demand is what gets new renewable generation financed. The constraint is visibility and delivery, not energy.

The physical ledger: water

• The entire public evidence base is one consent: Datagrid Makarewa, Southland — 220 million litres of groundwater a year for cooling [S-14][G-7]
• 220 million litres is trivial nationally and potentially significant against one aquifer — catchment is the unit of analysis; national milestones cannot answer a catchment question [S-14]
• Instruments: per-facility disclosure standard; national assessment guidance with water-free cooling as the reference technology; in over-allocated catchments, new cooling takes are simply unavailable [S-14][S-15]
• Microsoft’s Auckland region runs entirely water-free — proof the choice is real [S-15]
• The water team’s disagreement with national-milestone framing was kept on the record, not smoothed into false consensus

Speaker note: a deliberation that can hold a disagreement is a deliberation whose agreements mean something.

Compute without illusions

• The national research ceiling is Māui: ~1.4 petaFLOPS, commissioned ~2018, ageing [REANNZ, S-20]; the frontier sits three to four orders of magnitude beyond [G-2]
• No unsourced FLOPS targets anywhere in the framework; sovereignty means custody, access, jurisdiction [S-1]
• The floor: an owned research baseline (Māui’s replacement, due on ordinary research grounds) — where custody is load-bearing, the workload runs here, full stop [S-20]
• The ladder: subsidised access on the Canada/Singapore pattern, every cloud component passing the jurisdiction tests [S-1]
• Every owned stage grid-milestone-tied, with a symmetry clause watching dependency in both directions

Speaker note: owning tin is the expensive way to feel sovereign; a jurisdiction clause in a services contract does not survive the supplier’s home-state law.

Hardware: a permanent dependency, managed

New Zealand makes no chips and never will [S-21][G-1]. Accelerators are 3–5-year assets — the exposure re-runs at every refresh:

• Buy as one — pooled all-of-government procurement, multi-vendor, two refresh generations, vendor share caps [S-21][S-23]
• Stagger the fleet — no more than a third refreshes in any window [S-21]
• Buy alongside partners — trans-Tasman refresh-synchronised supply, allocation protocols agreed before anyone is desperate [S-1][S-21]
• Keep it running longer — funded lifecycle and second-life tier; longer useful life per imported unit is the only hardware sovereignty available [S-21]

Speaker note: the membership sent the first version of this chapter back for treating a recurrence as a one-off purchase. The return and the fix are both in the sealed record.

Data under New Zealand law

• The law is stronger than the debate credits — but 48% of workers put sensitive data into public AI tools anyway [S-16][S-1]: the law failed to reach the keyboard on a Tuesday afternoon
• A class-based handling standard: classification first (Māori classification binding, from the Model [S-22]), handling rules a human can hold in their head, enforcement upstream through procurement [S-23]
• A protected research pathway — ethics-approved, de-identified collaboration continues; the deliberation put a blanket “everything onshore” mandate to a vote and rejected it [S-16][S-18]
• A right to domestic processing, not a wall — and a one-page worker standard for every workplace that adopts AI [S-1]

Speaker note: whatever is proposed must be usable by a school office manager and a parish treasurer, not just by counsel.

People and work

Argued only from data New Zealand has; refused to invent what it lacks [G-5]:

• ~98,290 ICT professionals, growing 3.4% a year; 45% of the tech workforce on visas; enrolments down 33% since 2010; shortage acute at the senior end [S-19]
• A workforce-transition disclosure standard — creates the displacement evidence the next policy cycle needs [G-5]
• A redeployment fund on the cross-party pattern — moving people into better work, not compensating exits [S-1]
• A conditioned employer levy-credit — pays out only against portable, recognised micro-credentials; poaching stops paying [S-19]
• The income floor is named but not settled here — a whole-of-welfare design choice, not an AI line-item

Speaker note: the levy-credit existed in neither side’s opening position; the collision produced it.

Security as resilience

• No NZ-specific quantified AI threat metric exists — and the framework will not manufacture one [G-4]
• The evidenced risk is jurisdictional and dependency-shaped, not military: the June 2026 disruption, US-owned hyperscale, imported silicon, and 48% of workers as the widest attack surface [S-1][S-15][S-21]
• A register of critical AI dependencies — top 20 critical-service dependencies mapped in a quarter: which supplier, which jurisdiction, which refresh falls due when
• Synthetic-media provenance for official communications — the shortest fuse in an election year [S-1]
• The jurisdiction-of-inference test as the permanent ratchet; a sovereign fallback on the owned baseline where classification requires

Speaker note: a dependency you have mapped is a risk; one you have not mapped is a surprise.

Against the cross-party consensus

Tested against all seven commitments of In Our Own Hands (June 2026) [S-1]: aligns with all seven, extends five, differs in three places — with recorded reasons.

• The physical ledger is a precondition, not an afterthought — no owned compute stage without grid disclosure, interruptibility and constraint-consistent siting [S-12]
• Water cannot be governed by national milestones at all — catchment is the unit; national frameworks lack the instrument class [S-14]
• Residency absolutism rejected in terms — data should not have to leave; classes and safeguards, not walls; the research pathway protected [S-16]

Speaker note: where a consensus is sound, join it and say so; where the evidence points elsewhere, the difference goes on the record with reasons.

Sequence and reversibility

Three phases, adopted from the cross-party skeleton and extended [S-1] — every step reversible until the evidence says otherwise:

• Phase 1 — ground rules, now, low cost: procurement tests, the register, audit and handling standards, water disclosure, the non-severable Tiriti pair as legislation-plus-appropriation together
• Phase 2 — capability, years 1–3: the fund inside a sourced envelope; two or three domains, each with a named stop condition; every owned compute stage grid-gated
• Phase 3 — govern, federate, sustain: Parliament decides the authority; open standards, not a central registry; scheduled review of what the new disclosure data shows

Speaker note: sequence is the argument — each step creates the information the next one needs.

What New Zealand does not know

Eight gaps, declared and designed around — none filled with an invented number:

• Permanent: no chip fabrication [G-1]; no frontier compute [G-2]; no frontier programme [G-3] — design constants, not problems to solve
• Closable: no AI procurement rule [G-6]; no data-centre load in grid forecasts [G-8]; no water-draw disclosure [G-7]; no AI-specialist headcount or displacement data [G-5]; no quantified threat metric [G-4]
• Five of the eight are answered by disclosure standards — instruments whose first job is to create the missing evidence

Speaker note: where you cannot see, the first policy is sight. A policy that hides its ignorance ages badly.

The invitation

• Read the framework. Every claim carries a source; every gap is declared
• Check the record: every argument, block, revision, dissent and vote is cryptographically sealed and independently verifiable
• Not “trust us, we consulted” — a sealed record and the means to check it without trusting anyone
• Offered to every party, and to everyone outside the parties, as a demonstration that this standard can be met — and asked of anyone who claims a policy was built on evidence

Do not take this document’s word for it. Check the record.

Speaker note: close on the standard, not the content — the process is the demonstration.