New Zealand’s Government AI Drive Collides With Global Copyright Lawsuits

New Zealand’s public service is about to lose 8,700 positions. Finance Minister Nicola Willis tied that figure to a savings target of NZD$2.4 billion by 2029, naming artificial intelligence as the mechanism to fill the productivity gap those departures leave behind. Wellington’s plan sounds clean until you look at the legal dossier building against every major AI vendor the government already relies on.

A Munich court ordered OpenAI to pay licensing fees after ChatGPT reproduced copyrighted song lyrics without permission. Five of the world’s biggest publishing houses, joined by crime novelist Scott Turow, filed a class-action suit against Meta and its chief executive on May 5, 2026, alleging the company pirated millions of books to train its Llama model. Anthropic, already embedded in several New Zealand agencies, settled its own copyright lawsuit in September 2025 for USD$1.5 billion. The government is committing to tools that courts, across three separate jurisdictions, are treating as copyright infringers.

Wellington’s 8,700-Job Calculation

Minister for Digitising Government Paul Goldsmith put the sequencing plainly when pressed on regulation timing. “I wouldn’t commit to regulating AI before we get on with transforming the public service,” he told Newsroom. “We’re getting on with that now.” The cuts are scheduled across approximately 40 core agencies by 2029, with AI positioned to absorb the missing capacity rather than those roles being redistributed elsewhere in the system.

The deployment data confirms the sprint already has momentum. A 2025 cross-agency survey of New Zealand government organisations recorded 272 AI use cases across 70 agencies, more than double the 108 use cases across 37 agencies logged the year before. Microsoft Copilot and Microsoft Fabric have broad coverage inside public agencies and sit inside key government contract frameworks. Some agencies have already shifted portions of their workload to Anthropic, signalling that procurement flexibility exists even before formal policy catches up with deployment.

This week a government delegation travelled to Microsoft’s US headquarters for a briefing covering, among other subjects, the future of work. The agencies represented include the Financial Markets Authority, the NZ Super Fund, MBIE (the Ministry of Business, Innovation and Employment, which leads New Zealand’s AI strategy), Massey University, NZ Trade and Enterprise, and the Environmental Protection Authority. Brandon Hutcheson, director of quantum at AI data firm HSO and part of the delegation, is presenting on quantum computing while having publicly outlined 160 proposed AI initiatives for the public sector. The New Zealand Public Service AI Framework, launched within the July 2025 National AI Strategy, encourages rather than compels alignment with responsible-use principles. No AI-specific legislation is planned.

Labour’s science spokesperson Reuben Davidson has questioned publicly whether the government has worked out what the AI transformation will actually cost. Unlike Australia, where every federal department employs a chief AI officer and New South Wales agencies must complete a formal risk assessment before deploying AI, with high-risk systems listed on a public register, New Zealand has none of those requirements in place.

The Copyright Cases Mounting Against AI

At least 50 copyright lawsuits are pending in US federal courts against AI developers, including OpenAI, Anthropic, and Google, filed by publishers, authors, and entertainment companies. Courts have reached divided conclusions on whether training on copyrighted material qualifies as fair use. Wellington is moving at speed into the middle of that unsettled territory.

• November 11, 2025: Munich Regional Court ruled, in a case brought by Germany’s music rights collecting society GEMA, that ChatGPT memorized and reproduced copyrighted German song lyrics without a licence, ordering OpenAI to pay damages, cease reproduction, and disclose training data details. Bird and Bird’s legal analysis of the GEMA v OpenAI judgment describes it as the first European ruling to directly address AI training liability. OpenAI is appealing.
• May 5, 2026: Publishers Hachette, Macmillan, McGraw Hill, Elsevier, and Cengage, joined by Scott Turow, sued Meta and CEO Mark Zuckerberg in New York federal court, alleging Meta downloaded over 267 terabytes of copyrighted material from piracy sites including LibGen and Anna’s Archive to train its Llama model. The complaint states Zuckerberg personally authorized halting licensing negotiations after an internal message warned that licensing even one book would undermine a fair-use defense strategy.
• September 2025: Anthropic settled its copyright lawsuit for USD$1.5 billion, after a federal judge who initially found the company’s use of copyrighted books “exceedingly transformative” later ruled that training on millions of pirated books without consent or compensation was not acceptable.

Meta said it would fight “aggressively,” citing courts that have “rightly found that training AI on copyrighted material can qualify as fair use.” OpenAI is appealing the Munich ruling. The legal split between jurisdictions is intact, and no binding global standard is in sight.

Three NZ Voices on the Legal Exposure

Jenny Nagle, head of the New Zealand Society of Authors, framed the core procurement risk directly in comments to RNZ. The government would be paying subscriptions to international companies selling products that are part of active international lawsuits. The subscription charges keep accruing while the underlying model’s legal status remains unresolved in multiple courts simultaneously.

It will come down to an increase in licensing costs, and that’s where we’ll get stung.

Brandon Hutcheson, director of quantum at HSO, made that forecast when asked about the litigation risk for the public sector. He thinks outright model shutdowns are the less likely outcome of the legal wave. More probable, in his analysis, is a series of licensing regimes that AI companies build to settle cases or pre-empt new ones, with those costs passed directly into subscription prices at renewal.

His timeline for clarity runs to roughly three years. By then, some providers may have been shut down after losing cases, others will have built content compensation mechanisms into their pricing, and a handful will be operating from jurisdictions outside New Zealand’s legal reach. Procurement decisions made now will be renewing into whatever that landscape looks like.

Graeme Austin, professor of law at Victoria University of Wellington, questioned the ethics of the pace itself. Litigation is simultaneously testing the legality of AI products and services across the world, he told RNZ, and that test should produce verdicts before workers are replaced at scale. He pointed to the Dutch childcare benefit scandal, in which an automated fraud-detection system falsely accused tens of thousands of families of cheating, causing some to lose their homes, as evidence that the downside risk of moving faster than the regulatory landscape is not hypothetical.

His concern extends beyond copyright to what governments owe to the creators whose work trained the tools being purchased. An AI system built on material it may not have had legal permission to use is, he argued, not a stable foundation for public service delivery regardless of how fair-use questions ultimately resolve in court.

How Britain, Australia, and the EU Shifted Course

New Zealand’s voluntary, light-touch approach is an outlier among the three major jurisdictions whose AI vendor decisions will most directly shape the market Wellington is buying from. All three have moved since 2025 to restrict or more tightly monitor text and data mining (TDM, the programmatic scraping of copyrighted material for AI training), while New Zealand has moved in neither direction.

Britain’s House of Lords Communications and Digital Committee backed a licensing-first framework for AI and copyright in March 2026, and the UK government simultaneously confirmed it was dropping its earlier preference for a broad training exception, citing overwhelming opposition from creative industries. In Australia, 18 media organisations issued a joint statement this month urging Canberra to “hold the line” after reports emerged that the government was considering reopening copyright law as part of negotiations with AI companies seeking investment commitments. Sam Irvine, head of Copyright Licensing NZ, told RNZ he hoped New Zealand would follow Australia’s lead on data-protection resourcing at minimum.

Japan and Singapore amended their copyright laws to attract AI development activity. New Zealand has not followed that path, but has not moved in the other direction either. The practical result is that the government is purchasing from vendors whose legal obligations around training data are being determined by courts in Munich, New York, and London rather than by policy Wellington is helping shape.

The Marketplace NZ Does Not Have Yet

Irvine’s organisation, owned by publishers and writers, is building a licensing structure designed to trigger recognition and payment for creators when a copyrighted work is referenced inside an AI prompt. The mechanics operate at the point of use, not only at the point of training. “What I’d be recommending to the New Zealand public service,” he told RNZ, “is that they are looking to develop a marketplace for New Zealand creators so that we can have an internet where creativity is recognised and paid for.” His group is working on a licence that bundles consent, credit, and compensation into a single framework. The July 2025 NZ AI Strategy from MBIE acknowledged that fairly compensating creators “can support continued creation” and the availability of new works for ongoing AI training; that acknowledgment has not been followed by a funded policy commitment or a government partner willing to build the marketplace infrastructure alongside Copyright Licensing NZ’s proposal.

Comparable frameworks are operating at sector level elsewhere. EBSCO Information Services launched a licensed intermediary platform for AI and academic publishing in May 2026, connecting peer-reviewed journals to AI tools under a model where publishers retain rights and creators receive attribution for content referenced in AI outputs. A licensed scholarly content exchange between publishers and AI platforms shows that prompt-level attribution mechanics can work at scale. Copyright Licensing NZ’s broader proposal for the creative sector targets the same logic applied to a wider content class, with the same missing ingredient: a government partner committed to building the market rather than waiting for it to form without one.

An Architecture Built to Swap

Hutcheson’s structural argument for why the litigation risk is manageable, at least for central government, is about how the underlying systems are designed. He has observed a shift in the past five years toward public sector IT architectures that can swap AI models in and out as suppliers change or as legal outcomes demand. “As long as things are architected properly from government, we will be able to switch those models out,” he said. Some agencies have already demonstrated this by shifting workloads between Microsoft and Anthropic without structural disruption.

Smaller businesses lack that flexibility. For central government, the more immediate risk is less about being stranded with a shut-down vendor and more about paying inflated subscription costs once copyright liability resolves into mandatory licensing regimes, while no New Zealand-side creator compensation infrastructure exists yet to redirect any portion of those fees toward the creators whose work the models trained on.

If the Munich Higher Regional Court confirms the GEMA ruling on appeal, and if US courts begin narrowing fair-use protections for AI training data, the licensing costs Hutcheson forecasts become embedded in every contract Wellington signs from that point. The public service will be renewing subscriptions at higher prices into a market reshaped by verdicts made in jurisdictions it observed rather than influenced.

If courts continue splitting and no binding global standard emerges within three years, the government may close its efficiency window before the full bill arrives. The 8,700 jobs will be gone either way. The question is whether they are replaced by tools that cost what the strategy assumed, or by tools whose price is being set right now in courtrooms Wellington did not enter.