Who Decides When New Zealand's Public Systems Use AI?

Eight in ten New Zealanders believe AI regulation is needed, according to a survey cited by Victoria University of Wellington senior lecturer in artificial intelligence Andrew Lensen. Another study found only 5% of respondents were strongly confident the government would use AI responsibly. Those numbers sit awkwardly beside the government’s 2025 AI strategy, which deliberately takes a “light-touch approach” and relies largely on existing law and voluntary guidance.

Writing in The Conversation ahead of the 2026 election, Lensen asks who should control the AI systems New Zealand increasingly relies on, and who should benefit from them. His focus is not limited to generative AI. Public agencies already use automated decision-making systems, many based on machine-learning methods that predate ChatGPT, to determine or speed up decisions affecting individuals.

Some uses are routine and helpful. ACC says claims can be approved automatically before a patient leaves a GP appointment, while Inland Revenue can pay refunds without requiring the taxpayer to act. But the consequences of a poorly governed system can be serious. Immigration New Zealand abandoned a model used to identify overstayers for possible deportation after accusations of racial profiling. Across the Tasman, Australia’s unlawful Robodebt programme forced the government to repay A$721 million in welfare debts.

New Zealand has protections under the Official Information Act and Privacy Act, but Lensen argues they provide limited oversight for automated public decisions. The concern is becoming more immediate. Parliament has passed legislation allowing the Ministry of Social Development to automate benefit decisions, while the Law Commission is developing a framework for automated decision-making across government.

There are workable models elsewhere. Canada requires government agencies to assess the effects of automated decision systems. The European Union regulates AI according to risk, and Australia has established an AI Safety Institute. These approaches differ, but each moves accountability beyond optional organisational guidance.

A New Zealand framework would also need to account for te Tiriti o Waitangi. Lensen points to Māori data sovereignty principles, which emphasise community control over information and the benefits derived from it. Te Hiku Media offers a practical example: the non-profit has developed language technology while keeping control of te reo Māori data within the community. That changes the question from whether an AI system works to who sets its terms, holds its data and receives its value.

The article also identifies a wider issue of national control. Many AI systems used here are developed and hosted overseas. That affects personal information, intellectual property and New Zealand’s ability to retain control over important public services. It also means much of New Zealand’s AI spending flows to international companies. Decisions about procurement, tax and infrastructure therefore sit alongside decisions about safety.

Lensen is making a policy argument, not reporting the result of a trial. The examples show that safeguards are possible, but they do not establish which regulatory model would work best here. Nor does public concern alone settle the balance between useful automation and oversight.

What the evidence does establish is that automated decisions are already part of New Zealand’s public sector, and their effects are not abstract. Rules about impact assessment, accountability and data control determine how those systems meet the people subject to them. Lensen’s full analysis in The Conversation sets out the choices facing the next government and the communities whose information and rights are involved.