Dario Amodei just wrote America's AI homework. NZ's is different.

Anthropic's CEO has called the end of the "wait and see" AI policy era. But his essay comes from the seat of a country that builds frontier AI. New Zealand consumes it, and that changes which of Dario's recommendations apply to us, and how.

Released yesterday, Dario's Policy on the AI Exponential essay, which hit 4.6 million views and counting within a day, opens with a scene from The Lord of the Rings. Merry and Pippin try to warn Treebeard (the big ancient tree-like creature) that his forest is being destroyed. The problem is not that Treebeard does not care, it's that he moves so slowly that even saying hello can take him a full day.

That is the analogy Dario is using to illustrate his point. AI is moving at a ferocious pace, but governments and policy systems are built to move slowly, carefully, and years behind the thing they are trying to manage.

If you haven't read the essay, the next three paragraphs catch you up. If you have, skip ahead to "Where he's obviously right".

The essay is important because of who is saying it. Dario runs Anthropic, the most safety-conscious of the frontier labs (the handful of companies building the world's most advanced AI models), and for years he argued transparency was enough: publish your safety tests, report incidents, and build the evidence base. Now he's calling for binding regulation: mandatory third-party testing of frontier models, the way we certify aircraft before they fly, with the power for government to block a model's release if it fails (his point is made that regulating like aviation might not go far enough, and we should think of regulating AI more like regulating nuclear warheads).

The trigger was Claude Mythos Preview, a model so good at finding and exploiting software vulnerabilities that Anthropic declined to release it publicly, instead giving access to a defensive circle of Big Tech and critical-infrastructure firms (Fable 5, their latest public model, is the guardrail-heavy, safer version of Mythos).

When the people building the technology start asking to regulate the technology they themselves are building like aviation, it's time for more people to pay attention. The hook for me is thatthis is the firstdecentpiece of writing with real specifics on how AI regulation should work, rather than another sweeping opinion piece.

The key point for NZ is that four of Dario's five policy chapters assume you're the country building the model. NZ is not one of those. New Zealand will rent the "country of geniuses in a datacenter", as Dario calls it, by the hour through an API, like 99% of countries on Earth. That doesn't make the essay less relevant to us, but it changes which parts apply, and what the homework for our regulators and government should be.

Quick confession that I am not a policy person (shock... horror...). I've never set foot in the Beehive in a professional capacity, and if you asked me to explain how a select committee works I'd start confidently and finish badly.

But Dario's essay isn't just written for policy people either, it's written for anyone who can see the gap between how fast this technology moves and how slowly humans are responding. I think about that gap a fair amount (possibly too much).

Let's start with what I believe we should simply agree with.

The exponential is happening: the more computing power you throw at AI, the smarter it gets, which has been proven true for over a decade. The gap between "amusing toy" and "strategic technology", the kind that shapes national security and economic power, closed faster than almost anyone thought. The trap Dario describes is waiting until we know exactly where AI is going before acting. By the time that clarity arrives, the technology is already in everyday use, and policy becomes the ambulance at the bottom of the cliff.

Mythos made that risk concrete for him. Finding security holes in software used to take skilled specialists weeks; Mythos does it autonomously, in hours, for tens / hundreds of dollars. That exposes every country running modern software, and New Zealand runs its banking, health and energy systems on the same software as everyone else. We've had previews of this in the past: the NZX knocked offline for days in 2020, and the Waikato DHB crippled by ransomware in 2021. A small country is not a small target, but it is one with a thin security workforce, typically run on overseas-built (and sometimes hosted) infrastructure

Against that backdrop, look again at our July 2025 AI Strategy: no new AI laws, no AI regulator and lean on existing law (privacy, consumer protection, human rights), with the self-described positioning of ourselves as "smart adopters" of technology built elsewhere. Not very defensible at the time, and I can't say I supported it. Our 2025 problem was too little adoption (not runaway risk) and copying Europe's heavy AI rulebook would have buried our small nation in unnecessary compliance paperwork. But Dario's own journey is a potential pointer of where our (and my own) framework may fall short).

For years, Dario’s position was that AI companies should be more transparent about how they build, test, and release powerful models, which made sense when the risks were still hard to define.

But his position has shifted as the evidence has shifted. Once Anthropic started seeing stronger risks in its own frontier models, transparency was no longer enough, and he moved toward mandatory testing, outside audits, and the ability to stop a dangerous model from being released.

That is the useful lesson for governments like New Zealand. Transparency is probably the right place to start when we do not yet know exactly what we are regulating, but it should not become the final answer once the risks are clearer.

So what does action actually look like for a country New Zealand’s size?

We are not going to run our own testing lab for frontier AI models, and pretending otherwise is fantasy. We do not have the people, the budget, or the need to recreate what larger markets are already building.

The better answer is recognition, not replication.

Medsafe already works in a similar way with medicines, where New Zealand can rely on approvals from trusted overseas regulators rather than re-testing every drug from scratch. AI should follow the same logic.

If a frontier model passes serious safety testing under a trusted international regime, New Zealand should be able to recognise that. But if someone deploys that model into critical infrastructure like banks, hospitals, government services, or the power grid, they should have to report serious incidents to a New Zealand authority.

We do not need to run every test ourselves. We do need to know which systems are being used, where they are being used, and what happens when something goes wrong.

The other recommendation I would copy first is the least exciting one: measurement. Nobody gets good policy without good data. If AI started replacing New Zealand workers at scale today, we would probably find out too late, through lagging employment surveys and anecdotal evidence.

Stats NZ and MBIE should be building much faster tracking now, so we can see which jobs, sectors, and regions are being affected while there is still time to respond.

Dario’s economic worry is that AI could push the economy into what he calls a “hypergrowth, hyper-inequality” setting, where the economy grows extremely fast but most of the gains end up with the companies and investors who own the technology.

New Zealand’s more immediate problem is different. We are not yet dealing with too much AI being deployed across the economy, but with too little serious adoption, spread unevenly across too many businesses.

Plenty of New Zealand businesses have tried Copilot or ChatGPT, but very few have rebuilt the way work actually gets done. That is the difference between using AI for a few tasks and changing how sales, service, operations, finance, marketing, and product teams run day to day.

That is why a small advisory pilot for 51 businesses is a useful start, but nowhere near enough. New Zealand has more than half a million small businesses, and a small business in Te Awamutu can access roughly the same AI models as a company in San Francisco, on the same day, without building a lab or hiring a research team.

The gap is not access to the technology; rather it is helping business owners use it properly. That gap is high on the agenda of many NZ businesses because New Zealand’s productivity problem isn't a theoretical one. We work plenty of hours, but produce less per hour than many of the countries we compare ourselves with, and serious AI adoption across thousands of ordinary businesses is one of the more realistic ways we could change that. That needs funding and focus closer to a national infrastructure project, not a small investment pilot.

We also should not pretend job losses are impossible. A lot of New Zealand’s white-collar work sits in banks, insurers, professional services, and the public sector, so if AI starts changing office work quickly, we should not be designing the response after people have already been hit.

One of Dario’s suggestions is wage insurance, where someone who loses their job and moves into a lower-paid role gets topped up while they retrain or rebuild their career. New Zealand has already looked at a version of that idea through the Income Insurance Scheme, which was designed in 2022 and shelved in 2023.

I am not saying we bring that exact scheme back. But it'd be a worthwhile societal conversation to not start from a blank page if lasting disruption does arrive. We should do a version of that design work now, put the options on the shelf, and hope we never need to use them.

The coalition question is the one New Zealand cannot dodge. Dario argues that democracies should build an AI coalition around shared access to chips, compute, models, cyber defence, safety rules, and frontier capability. Countries inside the group get access and protection, and countries outside it fall behind.

New Zealand should want to be inside that tent. But we should also be truly honest about how hard that will be, and the reality of global agreement. We've seen this movie before with nuclear weapons and climate. Everyone agrees on the principle until national interest, commercial advantage, security, and politics start pulling in different directions. AI will be even harder, because the upside is broader, the concentration of power in a few countries is greater, and the incentives to get ahead are enormous.

That does not make the coalition idea wrong, not by any stretch. But it certainly makes it fragile, and, back to an earlier point, very difficult to measure, manage and maintain by virtue of society, as a collective, has very little in the way of answers as to how powerful AI will become.

The Project Glasswing example is a useful warning of this inaction. When Anthropic had a powerful cybersecurity model it did not release it to everyone, it gave early access to a trusted group of major technology, security, and infrastructure organisations. That is probably how frontier AI access will work more often from here, creating a haves and have-nots environment, with powerful capability moving through trusted circles first.

So the question for New Zealand is not whether we like that world. I don't think we can actually ignore it. The question is whether we are inside enough of those circles to get access to the compute, models, cyber defences, and safety tools we will need.

The good thing is that we do have cards to play, because we are a trusted democracy. We have mostly renewable electricity in a world that needs cleaner power for data centres. We also have a track record of moving quickly when we choose to, from the Digital Economy Partnership Agreement with Singapore and Chile, to the space rules that helped make Rocket Lab possible, to the current gene technology reforms. None of these make New Zealand an AI superpower, but they do prove that when we pick a lane, work with trusted partners, and build rules that are practical rather than performative, we can move faster than our size suggests.

That should be our angle in playing to our strengths by being one of the easiest, safest, and most trusted small countries to build with.

Dario's essay wasn't written for our economy and, importantly, our culture. Here are three things missing from Dario's essay.

Dario closes by saying Treebeard is waking up. That's fair enough, and might be true for Washington. New Zealand's risk was never sleeping through the chance to regulate frontier models (we were never going to build them).

Our risk is sleepwalking into three roles at once: rule-taker, technology-renter and price-taker, and not having a plan for the implications on our country. We'll be outside the room where access gets decided, with no safety net designed for displaced workers or our cultural heritage, while our productivity problem continues to compound.

My key takeaway and learning from Dario: although frustrated with the pace of NZ Government action in regulating this space, I figured we had more time on our side to understand where the technology was going first (classic mindset, according to Dario). I'm now convinced we need to act sooner. And with Dario's framework, take his recommendations, consider them in our context, and adapt our country's homework to be prepared for the tectonic shifts beginning to shake underneath our economy.