The following articles are from the Centre for Data Ethics and Innovation's July 2026 newsletter. The newsletter comes out quarterly, full of news, commentary, opinion, and education. Sign up by emailing dataethics@stats.govt.nz
Kia ora koutou,
In my role as Government Chief Data Steward, I'm delighted to open the June edition of the Centre for Data Ethics and Innovation newsletter. The Centre plays a vital role in shaping a shared understanding of ethical data use across New Zealand.
Through practical guidance, strong connections with people such as yourselves, and support for integrating data ethics into real work programmes, the team helps ensure data is used responsibly and in ways that build public trust. I’m proud of the impact this small team continues to make, and I hope you find this edition insightful.
The theme for this quarter’s newsletter is ‘Opening up Stats’. To me, opening up stats is about more than making information available. It is about creating space to understand the people and communities behind the data.
Numbers matter, but on their own they can never tell the full story of people’s lives, aspirations, and experiences. If we want our work to have real meaning, we need to listen as carefully as we measure. That means engaging early, understanding different perspectives, and ensuring communities have opportunities to influence the questions we ask, the way we interpret findings, and the decisions that follow.
When we do this well, data becomes more than something we collect and analyse. It becomes a tool for connection, learning, and shared understanding. It helps us recognise context, identify what matters most, and create value that is relevant, practical, and grounded in lived experience.
Data tells stories, but those stories are strongest when they remain connected to the people they represent and shared in a way that is easily understood. The greatest value of data is not in reducing people to numbers, but in helping us see them more clearly.
For Stats NZ, ‘opening up’ in this way is both an opportunity and a responsibility. It helps ensure our work is not only statistically robust, but also meaningful, trusted, and useful for the communities we serve across New Zealand.
On that note, right now Stats NZ is establishing Pacific, rainbow, and homeless community advisory groups. These groups, alongside the recently established Crown-Māori Statistical Design Forum, are being established as part of Stats NZ’s work to modernise the census and other official statistics, and they mark a significant step in Stats NZ’s commitment to a modern, inclusive census.
I trust you will enjoy this edition.
Ngā mihi,
Colin
I’ve been thinking a lot about the difference between information and data.
This came up in my day job related to a question I was asked; does the agency I work for need to know if someone is part of a rainbow community? For the day-to-day interactions kaimahi have with a person, no, but for understanding the needs of the communities who fall under the rainbow umbrella, yes. This is the difference between information and data.
Information is what is directly collected from a person and should be stored for the purposes of ensuring that how kaimahi interacts with them is in response to who they are and their needs.
Examples of this could include recording pronouns in a client record, or in a health setting recording if it is possible for them to get pregnant or develop certain cancers.
Whereas data is at a population level, it should be anonymised and used to show trends and possible issues populations are facing, while also noting that quantitative data will show that there is an issue—for example, rates of homelessness—but not why there is that issue.
Stats NZ has a longstanding commitment to working alongside Pacific statistical leaders and Tangata Moana communities across New Zealand and the Pacific. At the heart of this work are the people behind the data, particularly communities whose voices are too often underrepresented or seldom heard.
For Pacific statistics offices, the question is not simply whether AI should be used, but how it can be used responsibly while protecting the trust communities place in official statistics. In small island settings, even anonymised information can sometimes point back to individuals, families, or communities. A single confidentiality breach can damage public confidence for many years.
Used well, AI can support statistical work by helping with analysis, summaries, and insight generation. But it must never come at the expense of confidentiality, quality, transparency, or Pacific ownership of Pacific data. Public AI tools can create risks if sensitive survey, census, or administrative data is uploaded into systems hosted offshore or used in ways people did not expect.
Pacific statistical leaders also raised concerns about bias and misrepresentation. Many AI systems are trained on global data that may not reflect Pacific social structures, customary land arrangements, extended family systems, migration patterns, or cultural concepts. Without local knowledge and statistical judgement, outputs may be convincing but wrong.
The message from Canberra was clear: AI should augment, not replace, the expertise of Pacific statisticians. Responsible use means strong governance, staff awareness, community trust, and a continued commitment to ensuring Pacific people remain visible, respected, and in control of how their data is used.
After a decade in the community sector, I've observed a familiar pattern in "community consultation": we ask people with lived experience to share their stories as evidence, then remove those people from the process of tackling the problem. It's an extractive practice and one that reproduces the same hierarchies that created poor outcomes in the first place. As a result, we build services that don't work, technologies that cause harm, and policies that fail the people they're meant to serve.
A person who survives within a system - health, housing, social services - understands it differently from someone who studies it from the outside, and a person who holds no power in a system experiences it differently from those who do hold power. Lived experience isn't a soft complement to formal knowledge, but a distinct, legitimate, and irreplaceable analytical lens.
For government to work effectively with communities, it should treat experiential knowledge with the respect it deserves. This means bringing lived experience into the room where solutions are built, not just where problems are described and it means asking not only who is in the room, but who is absent - because the perspectives least likely to be heard are often the most valuable.
Tim Young with Ingrid Jones,
My Life My Voice
Every time I research a route before a trip, or decline an invitation because the access question feels too uncertain to resolve, I am experiencing a data failure. The information that would make that uncertainty unnecessary does not exist because it was never collected. The same absence now lives inside AI systems used to depict, describe, and define disability for a world that mostly learns about it from a distance. The cost is not abstract. It is disabled people continuing to disappear from the systems built to represent reality, first from the data, then from everything built from it.
Before my family goes anywhere new, I do reconnaissance. I check whether there is accessible parking close to the venue, whether it has rear-entry or side-entry access, whether there is a ramp, where it is, and where it leads. I try to find out whether the bathrooms are accessible, whether the path from the car to the door has steps, and whether there are kerb cuts or parked cars blocking the way. This research can take longer than the outing itself. And at the end of it, I still cannot always be sure. Sometimes I arrive and something listed as accessible is not. Sometimes I find out at the door.
This is the invisible labour of being a disabled person navigating a world that was not designed with you in mind, and it compounds. I have missed speaking events I was invited to because the stage was not accessible. I have sat awkwardly near the entrance of boardrooms because they were not large enough for a wheelchair. I have accepted awards from the ground, out of view, because there was no ramp to the stage. After enough of these experiences, something shifts: you spend less time seeking opportunities in new places, and more time doing what you can from home.
None of that experience shows up in any dataset. That invisibility, the way absence becomes self-reinforcing, uncounted, and therefore unaddressed, runs through every aspect of the disability data problem, including its urgent new frontier: AI systems that increasingly shape how the world sees and represents disabled people.
Without Data, There Is No Reality
Data is how we know whether something is real. It is how we distinguish genuine progress from the appearance of it, set targets we can be held to, and surface problems that are invisible to people not living them. When data is absent, the default is not neutrality. It is the perspective of whoever is in the room and speaking loudest. For disability, that has rarely been disabled people.
The consequences are visible across every domain. Employment rates among disabled people are often aggregated so broadly that distinct experiences disappear. Homelessness, poverty, educational outcomes, where disability data is collected at all, it is often too coarse to inform specific interventions and too infrequent to show whether anything is changing. Accessibility infrastructure, like the footpaths and building entrances I research before every unfamiliar trip, is largely unmapped. Planners make investment decisions based on who complained most recently rather than on a clear picture of where barriers are concentrated and who they exclude.
The result is a policy environment full of aspiration and short on accountability. You cannot set meaningful targets without a baseline. You cannot demonstrate improvement without something to measure against. Good data changes the conversation from “we believe this is a problem” to “here is where it is, here is its scale, and here is what fixing it would cost and deliver.”
The planning work I do before a family outing should not be necessary. It is necessary because no one has systematically collected the data that would make it unnecessary. That is a choice, even if it has never been made explicitly.
The Same Absence, Inside the Machine
The same failure that produces unmapped footpaths and missing detail in statistics is now producing something more pervasive: AI models that do not know how to represent disabled people accurately because the data used to train them did not show them.
When image and video generation models are prompted to depict people with physical disabilities, the results show what was missing from their training. Wheelchair users appear with arm positions that make no sense for self-propulsion. A blind person navigating a street appears without a cane or guide dog unless the prompt explicitly asks for one, and even then the result is often physically wrong. For most people, generating a convincing image of someone walking down a street is effortless. For a wheelchair user, the model can fail at something as fundamental as what their arms are doing.
These are not aesthetic glitches. They are data failures. They show that accurate, varied, and authentic depictions of disabled people, in motion, in context, using assistive technology correctly, were not present in sufficient quantity or quality when these systems were built. The model learned from what existed. What existed was inadequate.
The problem extends beyond images. When language models are asked about disability, how to describe it, design for it, communicate with or about disabled people, their answers reflect the same skew. The perspectives shaping the training data, evaluation prompts, and feedback processes have largely been non-disabled. So the model learns to reflect the world as non-disabled people have imagined it, rather than as disabled people actually inhabit it.
The Invisible Author Problem
This matters because most people using these tools, for images, video, marketing, education, or social media, have little or no direct experience of disability. They are not equipped to notice when an output is wrong. They may not know how a wheelchair user’s arms should be positioned, or that a blind person would not usually be standing at a crossing without a navigational aid. They accept the output as reasonable because nothing in their experience tells them otherwise.
That content circulates. It shapes how disability is perceived by people who had no idea they were getting it wrong. AI does not just reflect existing misrepresentation; it industrialises it.
This is the double disappearance at the heart of the data problem. Disabled people are first absent from the data systems are trained on. Then they are absent or inaccurately present in the outputs those systems produce. Because most people consuming those outputs have no reason to question them, the inaccuracy compounds invisibly, at scale.
I think about the speaking events I missed, the boardrooms I sat at the edge of, and the family trips that required hours of research to feel confident enough to attempt. None of that is visible to the AI model generating an image of a disabled person going about their life. Because it was never in the data, it will never be in the output, unless someone decides it should be.
What Getting It Right Would Change
Fixing this is not only about representation, though that matters. It is about what becomes possible when AI systems are trained on adequate data and evaluated by people with direct lived experience.
When a model learns from sufficient, accurate examples, its outputs change. A generated image of a wheelchair user shows correct posture and arm mechanics. A blind person navigating a scene has a cane or guide dog, held and used in a way that reflects reality. Crucially, none of that requires the person generating the content to know these details themselves. The model teaches by example, at the point of creation. It raises the default standard of how disability is depicted across the content it helps produce.
That matters because these tools are becoming ubiquitous. Many people creating content about disability will never consult a disabled person. If the tools they use have learned from disabled people, accuracy becomes the path of least resistance. Getting it right becomes the default.
But that only happens if the data exists. Disabled people need to be present in training sets: in genuine diversity, in real contexts, using real assistive technology correctly. Disabled people also need to be embedded in evaluation processes, reviewing outputs and identifying failures. That is not an equity add-on. It is a technical prerequisite for systems that work. Lived experience of disability is domain expertise.
In April, Stats NZ released An overview of disability data in Aotearoa New Zealand, a new report developed to make it easier to find, understand, and use disability data in Aotearoa New Zealand.
It includes an overview of key disability data sources, important indicators, and guidance on how to use them. It is based entirely on publicly available data from a selection of key population surveys as at 1 January 2026.