AI Governance: The Lowest-Trust Market You Operate In

Two numbers about how New Zealanders feel about artificial intelligence have been circulating together this year. They do not belong to the same survey. The first says 44 per cent of New Zealanders believe AI's benefits outweigh its risks, the lowest standing of any country measured. It also found 81 per cent wanting regulation and 89 per cent wanting laws against AI-generated misinformation. The second comes from a different instrument entirely, fourteen months older in its fieldwork. It does not mention New Zealand's 2026 data at all.

Put together, those figures read as one damning finding about a sceptical country. Taken apart, what is left is smaller and more recent, and more useful to anyone running an AI programme here. 62 per cent of New Zealanders say they would stop using a product or service over concerns about how an organisation was using AI inside it. 70 per cent of people who have actually used an AI-powered service in the past year have already hit something that went wrong with it.

The 44, 81 and 89 per cent figures belong to a KPMG and University of Melbourne global study, fieldwork run from November 2024 to January 2025 across 47 countries. The 62 and 70 per cent figures belong to One NZ's own AI Trust Report 2026, fieldwork run in March this year, New Zealand only. Merging the two into a single "One NZ AI Trust Report 2026" finding, as the headline currently doing the rounds does, describes this year's New Zealand market with a survey that predates it by over a year and was never run here in the first place.

A board asked to approve next year's AI governance budget on the strength of "New Zealand ranks lowest in the world on AI acceptance" is being asked to act on a number collected before most of the deployments it is governing even went live. That number was never collected in New Zealand at all. The number that should be driving that budget conversation is the one measured this year, in this country, after the systems in question were actually in production. A dated, misattributed figure is the wrong instrument for the decision in front of the board, not a conservative choice.

One NZ's report, titled The Trust Imperative, surveyed 1,001 New Zealanders aged 18 and over between 6 and 20 March 2026, weighted to census data, with a margin of error of plus or minus 3 per cent. It is the second annual edition of the study. The KPMG and Melbourne study surveyed 48,340 people across 47 countries in late 2024 and early 2025. That makes it the right instrument for a cross-country comparison, and the wrong one for a statement about what New Zealanders think today. Both facts can be true at once. An article that wants to say something accurate about trust in AI this year has to use the first number and date the second.

Usage held. Trust did not.

Net AI interaction across the population held at 76 per cent over the past twelve months, down one point on 2025 and not a meaningful change. Regular use of chat-based assistants such as ChatGPT, Claude and Gemini sits at 50 per cent. The population now splits into three groups that barely moved this year: AI Actives at 29 per cent, up three points; AI Dabblers at 48 per cent, down four; AI Avoiders at 24 per cent, up one. Adoption held flat this year, neither collapsing nor surging.

The three groups are more than demographic colour. AI Actives are the 29 per cent who have moved past occasional use into something closer to routine reliance. It is this group's experience, more than the Avoiders', that is generating most of the problem reports below. An organisation that tracks only aggregate usage treats a satisfied Avoider and a frustrated Active as the same data point. But the figure that actually predicts churn sits almost entirely inside the second group.

Trust is not flat. Across every sector the report measures, confidence in organisations to use AI ethically and responsibly fell. Large multinational businesses are the least trusted category of all: 51 per cent of New Zealanders now express low or no trust in them, up four points on last year. Belief that AI will deliver better outcomes for society fell from 43 per cent in 2025 to under two in five this year. And the behavioural consequence is the headline this article leads on: 62 per cent of New Zealanders say they would stop using a product or service over concerns about how an organisation was using AI inside it, a figure essentially unchanged on 2025. Usage figures alone never surface that kind of standing concern.

The experiential data explains why. Among people who had actually interacted with an AI-powered service in the past year, 70 per cent say they hit at least one problem, two on average. 37 per cent found the system could not understand what they wanted, 33 per cent found it difficult to reach a human afterwards, and 31 per cent received incorrect or misleading information. A further 27 per cent found the experience impersonal or frustrating, and 24 per cent had a privacy or data-use concern. Of the people who hit a problem, 70 per cent say the bad experience made them much less likely to use that kind of service again. Some stopped altogether.

This is the divergence I traced from the other direction in The Hallucination Economy: a deployment can look identical on a usage dashboard for two years running while the credence it was supposed to earn moves in the opposite direction. A dashboard measures the performance, not what the audience watching it concluded. One NZ's own instrument has now measured both sides of that gap in the same population, in the same year. They are not moving together.

A second finding cuts against the easy reading that New Zealanders want AI to go away. Forty-seven per cent say they would trust AI-powered products and services more if government set clearer expectations and standards. The specific asks are concrete rather than vague: clear rules on how AI can and cannot be used, at 31 per cent; stronger enforcement of the privacy and data protection law New Zealand already has, at 28 per cent; and ongoing monitoring, at 26 per cent. None of that reads as hostility to the technology. It reads as a public that has already decided governance, not capability, is the open question. It is naming the specific mechanisms it thinks would close it. A further 45 per cent are concerned about the environmental impact of business AI use, rising to 63 per cent among 18 to 24 year olds. Three in ten of those with major sustainability concerns say nothing currently on offer would address them.

The constraint nobody is measuring on a dashboard

None of the above is a New Zealand peculiarity. A public-perception survey published this week in South Korea comes from the SBS Cultural Foundation and Seoul National University's Institute for Future Strategy. It found that more respondents trust AI's judgement on social and policy questions than they trust their own National Assembly, courts or government. On its face that reads like the exact opposite of a low-trust market: a population apparently more confident in AI than in its own institutions.

The same mechanism is at work, read from the other side. The same South Korean respondents who said they trust AI's judgement over their legislature also insisted that ethical judgement and ultimate responsibility have to remain exclusively human, regardless of how capable AI becomes. Read together, the two findings describe institutional confidence falling faster than AI's own reliability is rising. That is sufficient on its own to produce "people trust AI more than parliament" without any change in what the AI actually does. New Zealand's numbers and South Korea's numbers are the same underlying variable, trust in the institution standing behind the system, moving in different directions for different local reasons. Both arrive at the same structural point: the constraint on AI deployment right now is institutional trust, not the model.

The scale of the South Korean inversion is precise. The same survey found that roughly one in three respondents said they trust AI's opinions more than their own judgement. The researchers paired that finding deliberately with the institutional-trust result rather than publishing it alone, because the two numbers describe the same erosion from two different angles: one about who gets believed, the other about who gets deferred to.

The constraint is public licence. No model card or benchmark score reports on it. A board can prove its AI system works, but the One NZ and South Korean data together say that proving the system works is no longer the same project as proving the public still believes the organisation running it.

What this means for a programme that already passed its pilot

Most enterprise AI governance effort right now goes into two places: does the system perform, and does it comply. Both are measurable and both show up on a dashboard. Neither answers the question the New Zealand and South Korean data actually put in front of a board. Does the public using this system, or standing near it, still extend the organisation the benefit of the doubt? A board that reviews only the first two each quarter is reviewing exactly the two things this data says are not where the actual risk sits.

Three questions follow directly from the data rather than from theory. First, what changed since the system went live? Specifically, has anyone checked what customers now believe about how the organisation uses AI, separately from whether the system itself still performs to specification? Usage figures and trust figures move independently, and only one of them is usually on the board pack. Second, who actually answers when something goes wrong? 33 per cent of the people in the One NZ data who hit a problem could not reach a human afterwards. That is an escalation-path failure, not a model failure. It is the kind of failure a capability audit will not find, because the model was never the thing that broke. Third, what would rebuild trust if it were already falling? Decide that in advance, not after the fact. New Zealanders have already told researchers what they want: clearer rules, enforcement of existing law and ongoing monitoring. An organisation that can show it is already doing some version of those three things is answering a question the public is asking before it is asked directly.

None of the three questions needs a new survey instrument to answer. The first is a matter of asking the organisation's own customers directly, rather than inferring sentiment from usage logs. The second is visible in the organisation's own support and escalation data, if anyone has looked at it through this lens rather than as a generic service metric. The third is a decision a board can make before anything goes wrong. That is the only point at which making it is still a choice, not a reaction.

None of this argues for less AI. Seventy-six per cent of New Zealanders interacted with AI in some form in the past year. That figure held steady. The argument is narrower and, for a board, more actionable: the licence to keep deploying is being measured separately from the decision to deploy, by the people the deployment is aimed at. An organisation that only tracks the first number is flying on half the instrument panel.

The same erosion the New Zealand data measures domestically has an institutional counterpart in how openly AI vendors disclose what their own systems actually do. The Foundation Model Transparency Index is an openly published scorecard built by researchers at Stanford, Princeton, MIT and Berkeley. It grades major AI vendors each year on training data, downstream use and post-deployment monitoring, with its full methodology and every vendor's result published for anyone to read. Its third edition, released in December 2025, recorded the average score falling from 58 out of 100 the year before to 40, even as the underlying systems grew more capable over the same period. Transparency and capability are not moving in the same direction. An index the public can read, and no vendor can quietly opt out of, is one of the few mechanisms that puts an actual number on that gap, rather than leaving it to survey sentiment alone.

The same constraint shows up, in a different register, in defence AI procurement. The United Kingdom's Ministry of Defence published its Defence Artificial Intelligence Strategy in June 2022, alongside a companion policy titled, plainly, Ambitious, Safe, Responsible. The pairing is deliberate. The Ministry's own account holds that losing the confidence of the public and of allied partners would make its AI programme less ambitious, not more, because a fielded capability nobody outside the department trusts cannot actually be used. On the sovereignty side, trust is itself a gate, not a layer added after a system clears its technical and legal gates. An ally's willingness to share a capability, or a public's willingness to accept one operating near its own infrastructure, depends on that gate being satisfied first, the same precondition the New Zealand and South Korean data describe one register down.

The gap between a system that works and a public that still believes in it is exactly the gap I set out to name in The Hallucination Economy. It never shows up in a release note. It has to be measured on its own terms, separately from whatever the system itself is doing, or it accumulates until the 62 per cent who said they would leave actually do.

None of this says AI adoption in New Zealand is failing. It says the number worth watching is not the one that has been circulating. It was never the one a usage dashboard reports anyway. Usage held at 76 per cent. Trust in the organisations running the systems fell, in every sector measured, for the second year running. 62 per cent of the public have already told researchers they will act on that by leaving. A programme that tracks the first number and not the second is measuring half the market it operates in.

Has your organisation ever asked its own customers, separately from the system's performance metrics, whether they trust how it is actually using AI, and would you be comfortable publishing what they said?

If your organisation is moving AI agents from pilot to production and nobody outside the vendor has inspected the control plane, message me and I will send the scope and the fixed fee for an independent review.


The views expressed in this article are entirely my own, informed by more than 30 years of professional experience in architecture, security, and technology leadership in New Zealand. I write as director of Te Pono Limited; the views are personal and do not represent the position of any client, any government agency, or the New Zealand government. My commentary on legislation and policy is analytical, drawing on publicly available sources and my professional expertise in architecture, security, and AI governance, and it is politically neutral.


Andreas Hamberger is a New Zealand leader in Architecture & Security and Associate Member of the Institute of Directors. The Hallucination Economy examines what happens when organisations buy, deploy, report on and are audited against AI claims that nobody outside the producer can check. Through Te Pono he provides independent reviews of agentic AI control planes for organisations moving from pilot to production; contact andreas@thehambergerreport.com for the scope and fixed fee.


This article was produced with AI assistance under my direction. Research, drafting and images pass through a pipeline I built and govern: automated gates for source verification, forbidden language and political neutrality, and my own review before anything is published. The tools include Claude, Gemini and Openart. The frameworks, arguments and editorial judgements are mine and are the same discipline I apply to the AI systems I audit for clients. AI accelerated the work; the thinking, and the responsibility for it, are mine.


[1] One NZ. "AI Trust Report 2026: The Trust Imperative." 26 May 2026. https://content.one.nz/17/c5/494374e740b1bfc232f682933ef8/ai-trust-report-2026.pdf

[2] One NZ Media Centre. "Kiwis shift from AI curiosity to scrutiny as adoption grows, new report finds." 26 May 2026. https://media.one.nz/ai-trust-report-2026

[3] KPMG and University of Melbourne. "Trust, attitudes and use of artificial intelligence: A global study 2025," New Zealand country snapshot. https://assets.kpmg.com/content/dam/kpmgsites/xx/pdf/2025/05/trust-attitudes-and-use-of-ai-new-zealand-snapshot.pdf.coredownload.inline.pdf

[4] KPMG. "Trust, attitudes and use of AI." KPMG New Zealand insights. https://kpmg.com/nz/en/insights/ai/trust-attitudes-and-use-of-ai.html

[5] SBS News. "'AI More Trustworthy Than National Assembly, Courts, Government'... First Public Perception Survey." 5 November 2026. https://news.sbs.co.kr/english/article.do?news_id=N1008831124

[6] SBS News. "AI Shaking Core Institutions... Proposal to 'Change Power and Relationships'." 6 November 2026. https://news.sbs.co.kr/english/article.do?news_id=N1008833034

[7] Stanford Institute for Human-Centered AI, Center for Research on Foundation Models. "The Foundation Model Transparency Index." December 2025. https://crfm.stanford.edu/fmti/December-2025

[8] United Kingdom Ministry of Defence. "Ambitious, Safe, Responsible: Our Approach to the Delivery of AI-Enabled Capability in Defence." 14 June 2022. https://assets.publishing.service.gov.uk/government/uploads/system/uploads/attachment_data/file/1082991/20220614-Ambitious_Safe_and_Responsible.pdf

Next
Next

AI Governance: Nobody Opted In, and the Terms Moved Anyway