1. A public register of government AI systems’ function, performance, and impacts
1. A public register of government AI systems’ function, performance, and impacts
We want the government to make a list -- a 'register' -- of the ways it uses AI.
This list should include:
How AI is being used,
How well it's performing,
How this could impact our government and our lives.
For example, the Ministry of Social Development is using relatively simple rules-based AI to check if someone applying for a student loan is already on the benefit or not. This particular system is 99.9% accurate, and any potential errors can be redirected from the automatic system to a human for review.
Another example is that Health New Zealand has endorsed the use of "ambient AI scribes" like Heidi Health, allowing staff to automatically transcribe conversations with patients. This system been shown to reduce documentation time by 10 minutes per patient, and reduces after-shift admin by 81%. However, it can still produce errors, which is why the outputs are considered draft only and must be verified by a medical professional before they can be saved into a patient's medical records.
All of this information would be in one place, making it easy for you to find and use. A register could also make it easier to understand the information, because information is presented in the same format each time.
If you're worried that recording every way AI is used is too much paperwork... it doesn't have to be! Some types of AI aren't very risky. For example, many automatic phone menus use AI to understand what you have said, and choose where to send your call. If a use of AI is clearly low-risk, it could be OK to report less information about it. But systems that have a material impact on the public, from welfare to housing to legal rights, must be able to receive the necessary public scrutiny.
At this stage, we don't know exactly how to define 'low-risk' and 'high-risk' uses of AI. That's why we're also asking for public input. We, the New Zealand public, should have time and space to figure out what types of AI use in government we are comfortable with.
If you want to know about which government agencies are using AI, right now you have to rely on them to voluntarily disclose this, or for someone (like a journalist) to file an Official Information Act (OIA) request and publish their findings. There are many systems already in place, including Automated Decision Making, which the general public is not aware of and there is very little published about it.
We can't ask questions if we don't know about it - transparency is the first step to trustworthiness. A public register forces proactive disclosure, and it ensures that transparency is part of the system design, not an afterthought triggered by an OIA request or a legal challenge.
We also think that a list of AI systems (functions) isn't enough - we need to see that testing has demonstrated how well these systems work (performance), and that agencies have identified how these systems will affect people (impacts). This aligns with our next idea around impact assessment.
The United Kingdom recently launched the Algorithmic Transparency Recording Standard Hub
Canada has an AI Register for government AI systems (see here for an introduction)
The Netherlands has an Algorithm Register (see here for a list of all algorithms)
Scotland has a Register of AI Systems in the public sector
The European Union AI Act will require a EU database for high-risk AI systems (in late 2027)
New Zealand Police publishes a Technology Capabilities List, and the Ministry of Justice has a list on their Algorithm Charter page
The Public Health Communication Centre Aotearoa has an opinion piece arguing for a register.
A report from NZ’s Maxim Institute (May 2026) recommends creation of an AI register for the NZ public service (see p25).
Dr Alexandra Andhov (University of Auckland) wrote this op-ed in Newsroom (May 2026).
Taylor Fry’s report on the NZ Algorithm Charter (2021) recommends NZ ‘develop an annually updated register of algorithms covered by the Charter’ (see e.g. their summary on p33).
The University of Otago's AI and Law in New Zealand Project called for a public register of predictive algorithms back in 2019!
New South Wales’ recent report on Automated Decision-Making in government recommends ‘planning and piloting solutions such as disclosure registers, as a means of staying abreast of international AI governance developments’ (p33).
The NZ Public Service AI Work Programme includes an item to develop a "Central repository of AI registries, patterns and tools", although it is unclear if this will be made public, or contain the type of information we think is important, as the primary objective is to drive AI usage.
While we fully expect the details to be debated before anything is implemented, here are some elements based on international models that could be included:
Objective: What problem is the AI system trying to solve?
Capability: How does the system actually work? This could contain both process elements and technical elements.
Provenance: What datasets were used to train this system, and where did the system itself come from? How confident are we that the system can reflect New Zealand's unique demographic and cultural context?
Human-in-the-loop: At what stages do humans review the machine outputs, and can they intervene?
Performance: Has the system been tested, and how often is it reviewed for accuracy and acceptable levels of error?
Accountability: Who is specifically held accountable if the system produces discriminatory or flawed outcomes?
Impact: Who does this system affect (positively or negatively)? Is there a link to the impact assessment?
Not necessarily. We are open to a framework that appropriately recognises the level of risk, and low-risk administrative tools may not need to be included. For medium or high-risk systems that have a material impact on the public, slowing down might not be the worst thing if it means we think about it more!
In most cases, the information we are asking to be published is already held by the agencies when they design and procure these systems. There is minimal new work, we are just asking them to publish that information in a centralised, plain-language way.
Even then, narrow exceptions consistent with existing law would be required, for example for systems that are used for national security, or where release of the information would prejudice the maintenance of the law or endanger public safety. Commercial sensitivity should not be a valid reason to withhold the existence of a tool, noting that no source code or proprietary secrets need to be published.
Innovation flows faster when there is certainty. The government can progress with confidence that they are taking the appropriate steps to build trust with the public and develop the requisite social licence. Deploying AI in secret and then having it exposed later leads to public backlash and legal challenges, which slow down the pace of government innovation far more than publishing information that already exists into a public register.