Comment: When I arrived in New Zealand last July, the restaurants confused me. A server would clear a plate and ask, “you guys all good?” and seem to mean it. During the Wellington on a Plate food festival in August, asked how a new burger was, I gave a careful and honest opinion before I understood the question was mostly a courtesy. My misunderstanding encountering a new culture. It took me a while to read the warmth for what it was.
I have been thinking about these exchanges because the task buried inside them is exactly the kind of thing now being automated away. In Seoul, you order and pay from a tablet at the table. In Tokyo, robots carry plates across the floors of restaurants. The machines remove the task. What they cannot remove is the small but unmeasured surplus: a person who wants to know how the burger was or offering Kiwi hospitality.
This distinction, between a task and a job, sits at the centre of New Zealand’s bet on artificial intelligence. You would not know it from how the bet is described. In May, Finance Minister Nicola Willis announced 8700 public service job cuts alongside a drive to embed AI across government, promising $2.4 billion in savings over four years.
The framing was liberation: AI would free public servants from drudgery. But as US labour economist David Autor has shown, technology rarely replaces whole jobs. It replaces tasks. And the tasks it takes first are the legible ones: the parts of a role that can be written down, counted, and optimised.
Consider what policy analysts do. They search documents, summarise meetings, draft briefings. They also read a room, give free and frank advice to ministers, and weigh whose interests are at stake. AI may be capable of tasks such as searching documents and summarising meetings but it’s useless at reading a room. So the honest description is not that 8700 jobs will vanish. It is that thousands of roles will be hollowed of their measurable tasks, leaving a remainder nobody has thought to value.
But this isn’t only a story about AI hollowing jobs. It’s also a story about who controls the AI tools that will be used in government agencies.
A case in point: on June 12, the US Commerce Department ordered tech company Anthropic to disable access to its most advanced AI models (Fable 5 and Mythos 5) for all foreign nationals worldwide for “national security” reasons: no grace period, no exemption for allies, even Five Eyes members. New Zealand’s National Cyber Security Centre was also affected as it had previously been given privileged access to one of these systems. This agency, responsible for the country’s digital sovereignty, held little of its own in this instance.
Setting aside whether the US decision to restrict access was justified, the case highlights New Zealand’s reliance on AI tools that are controlled elsewhere.
We’ve built no general purpose foundation AI model for our own use. The government’s AI strategy is to be a “sophisticated adopter” of frontier technology, not a builder of it. The general purpose AI systems that ministries use are owned by a handful of US companies. Microsoft’s Copilot tool seems to be the default option across Wellington.
Relying on imported AI models for the sake of efficiency not only makes us dependent on this overseas tech but it also traps us in what German sociologist Max Weber called “the iron cage”. Rational systems built for efficiency come to constrain the people inside them, not by a tyrant’s design but by the way they’re assembled piece by piece, in the name of doing more, until those within can no longer ask why they do what they do.
When Minister for the Public Service and Digitising Government Paul Goldsmith was questioned by a reporter about whether he used AI, he said he asked a large language model for opinions on his ideas. Is this also how officials will be expected to use AI? Shifting from writing briefings to revising what a model has written. Starting from a blank page and amending a supplied one require different kinds of thought. The draft arrives already carrying assumptions.
The future of a world with AI is genuinely uncertain in ways that bear directly on the Government’s bet. Professor Alexandra Andhov, chair in law and technology at Auckland University, has warned that AI prices are still “heavily subsidised” with firms selling access below cost to capture the market, so today’s prices are not the real ones. The regulatory vacuum continues, leaving no framework for when something goes wrong. And geopolitics can shift overnight.
This uncertainty is the reason to scrutinise the Government’s wager on AI now rather than in hindsight.
Much of what people know how to do is intuitive – as the philosopher Michael Polanyi put it, “we can know more than we can tell”. And what cannot be told cannot be costed or outsourced. The measure of the Government’s strategy will not be what it saves by 2030, but whether a nation that has learned to reason through borrowed intelligence still knows how to reason, to work, and to judge without it.
