Green policy to build Kiwi AI 'isn't fanciful, it's necessary'

Imagine waking up to find that a foreign company had bought State Highway 1. Every drive to work, to school, or to hospital now comes with a toll. The owner sets the price, sends the profits offshore, while paying minimum tax here. The owner also decides who may drive and can close the road whenever its own government tells it to.

Would you accept such a regime? We doubt it.

Yet that is the deal we are drifting into with artificial intelligence (AI). Large language models (LLMs) are the AI systems behind tools like Co-pilot, Gemini, ChatGPT or Claude. They are fast becoming the roads, the digital infrastructure that businesses, and increasingly public services, rely on every day. And almost all these roads are owned by a few overseas companies.

We vote on November 7, and the parties have only recently started to present their AI ideas. The Green Party’s new AI policy contains one line every party should be asked about: “Government will encourage research towards the development of Aotearoa-based and focused LLMs to avoid dependence on foreign vendors.”

Building our own LLMs is just the start for our independence and sovereignty.

The toll: today’s price is not tomorrow’s

The price we pay for AI today is not its real price, and this needs to be our first recognition. AI price is held down by investors, not set by what these systems really cost to run. OpenAI, for example, even with projected revenue of US$50 billion in 2026, continues to operate with a net loss and does not expect to turn a profit until 2029. However, investors will eventually want their return, and prices will follow.

By then, though, it may be too late to change our course. Once our government agencies have built their systems, workflows, and records around one supplier, moving the data elsewhere becomes slow and very expensive. Economists call this a “lock-in”. And once that takes place, the AI companies will increase the fees, and a road that once cost one dollar will cost five tomorrow, and by then there is no other way to drive.

Not just the fee, but also our knowledge and secrets

Every trip on someone else’s road also reveals what we are carrying. When public agencies use foreign AI systems, they hand data to foreign companies: citizens’ questions, case files, policy drafts, health and welfare records. Much of it is valuable and sensitive, and some of it touches on national security.

Yes, the AI companies do promise to protect our data, but some protections are beyond what the technology allows. To use these systems, our data usually has to leave New Zealand to be processed. And even where it stays here, it does not necessarily stay protected. Under the United States Cloud Act of 2018, American providers must hand over data in their possession, custody, or control when US law enforcement orders it, even if that data is stored abroad.

A promise is only as good as our ability to verify it. These AI systems are complex and opaque systems and in most cases, we cannot even audit them. That leaves us unable to know where our data goes, who can reach it, or how it is used. 

The foreign state that can close the road

This is not a far-fetched risk. On 12 June this year, the US government ordered Anthropic, one of the world’s biggest AI companies, to block its two newest, and arguably the most capable models, for every foreign national on the planet. The company switched them off for all its customers, New Zealand included.

The lesson is not about one company. The most capable AI systems are concentrated in a handful of firms, in a handful of countries, and those firms answer to their own governments. Our dependence becomes their leverage. They can press us on trade deals, sanctions, and other decisions we might not otherwise take.

New Zealand can build its own LLM

Building our own LLM upon which we can build a variety of tools does not mean competing with Google or OpenAI. It means building a smaller model designed to support specific activities that will support our public servants.

Other small countries have built what they need. Switzerland released its own model, Apertus, in 2025. It was built by Swiss public universities and the national supercomputing center, while anyone can inspect how it was made, and it understands all official Swiss languages. Singapore put S$70 million into its own model, SEA-LION, deliberately smaller and faster than commercial giants and built for Southeast Asian languages.

Building is also getting cheaper. Capability that only a few years ago could only be achieved at a great cost is now within a small country’s research to own.

Can we shape the AI and keep the value in New Zealand?

AI is a powerful tool, but also highly experimental and risky. Not just from a cybersecurity perspective, but also in its potential impact on the way we think, and whether we think in the first place. In an experiment with more than 1,500 participants, Cornell researchers found that writing with an AI assistant built to favour one side of an argument shifted not only what people wrote but the opinions they reported afterwards. Another research by Carnegie Mellon University found that the more people trusted generative AI, the less critical thinking they applied to its output.

A model built overseas carries the values of those who made it. A model built here, under our laws, can be designed from the start to reflect our values, Te Tiriti o Waitangi, kaitiakitanga and respect for the Māori and for our sky, land and water.

Building the LLM ourselves is not only safer, but it is also a better use of public money. Every dollar spent renting overseas AI leaves the country for good. Every dollar invested here stays here, building local skills, jobs, and intellectual property that New Zealand owns, and giving our graduates a reason to build their careers at home rather than abroad.