Sovereign by Design: Ethics and Governance for Australia's National Health AI Infrastructure

I'd like to thank you very much — the organisers, some of whom I see in the room of this event. And I would like to just, on a personal note, offer a counter to the tech bros in Silicon Valley or in Washington State are determining our AI future. I think it's much more accurate to say it's the Kate Chaneys, the teals, the Kimberlee Weatherall, Rebecca and the list goes on. And many of them happen to be women, quite interestingly, with a legal background or a social sciences background who are supercharging our discussion.

And scholars like Lily in the back there, who I just met last night, with a law degree and a strong interest in AI. I would like to begin first slide by acknowledging the traditional custodians of the lands on which we are meeting here today, the Gadigal people of the Eora Nation, and I pay my respects to elders past, present and First Nations people who are in the room today. I'd like to not start with the project that I am going to speak about, but with the geopolitical context of what makes this work urgent.

Because without understanding that context, the governance and the ethics packages that I'm going to unpack today might seem like a compliance exercise, and I contend that they are not. They are responses to a structural shift in how nations think about data, computation and sovereignty, which we've heard a lot about this afternoon — well, all day, actually. And we're in a period where the global AI infrastructure landscape is consolidating around a small number of offshore hyperscalers and proprietary model providers, largely in the United States and China.

Between 2023 and 2025, companies announced plans to invest in what could scale to more than a $100 billion investment in Australian data centres. This is a figure that reflects Australia's geopolitical and strategic position as a geopolitically stable Five Eyes ally with strong regulatory frameworks. In 2024, Australia ranked second globally, just behind the United States, as a data centre investment destination. In December of 2025, the Australian government released its national AI plan, which explicitly identifies sovereign digital infrastructure as a strategic priority for our AI-enabled economy. But here's the tension.

While Australia is attracting infrastructure investment, Australian health researchers are becoming increasingly dependent on offshore cloud platforms and proprietary AI models to do their work. When a researcher trains a clinical prediction model using a commercial large language model API, or processes linked health records on infrastructure hosted outside of the Australian jurisdiction, they are making a sovereignty decision, whether they frame it that way or not. The data may traverse international boundaries. The models' training data, architecture and decision logic are likely to be opaque. The terms of service may permit secondary uses of that data that would not be permissible under a strong privacy law.

And there are serious concerns with key data sets that we very routinely rely on from the United States or from the UK — for example, the UK Biobank. If anybody here is involved in health care life sciences research, the UK Biobank temporarily suspended access to its research platform in April of 2026, after de-identified participant data was found to be offered for sale on an e-commerce platform in China. The incident affected data originating from a resource containing health, genetic and other information from roughly 500,000 participants. The crucial point is that it was not reported as a conventional cyber attack in which hackers penetrated the UK Biobank central infrastructure.

Rather, it exposed a weakness in the governance of researcher access and data export. The question is really whether AI will transform health research. We know that it already has. The question is whether Australia builds the sovereign infrastructure, governance frameworks and workforce capability to ensure that transformation happens on the terms that protect the rights and interests of Australian communities, and particularly Aboriginal and Torres Strait Islander peoples. Which brings me to the ARDC, Australia's national research infrastructure, the Australian Research Data Commons. The ARDC is one of 28 nationally significant research infrastructure facilities funded through the Australian Government's NCRIS, or National Collaborative Research Infrastructure Strategy.

The Australian Research Data Commons and the wider community recently celebrated 20 years, which is an incredible achievement, and a $5.5 billion science investment to secure Australia's economic, environmental and social future. And we rarely hear figures like that $5.5 billion when we're talking about our digital infrastructure. NCRIS represents a sustained national commitment. The Australian government is investing $4 billion from 2018 to 2029. That's 11 years to support national research infrastructure that many of the universities in Australia benefit from. In the most recent funding round, announced in 2026, the ARDC received almost $39 million to strengthen its delivery of digital research infrastructure, to build a connected, ethical and AI-ready data ecosystem.

This is a good news story. What distinguishes the ARDC's operating model is this principle of co-investment. They don't invest as if they're a venture [inaudible] investor or angel investor. The universities typically have to pony up researchers with in-kind contributions or cash contributions of $0.50 on every $0.50 that they spend. The ARDC does not build the infrastructure in isolation. It co-funded projects with Australian universities, research institutions and sector partners, leveraging government investment to create shared national capability. The result is that this infrastructure is owned and governed collaboratively by the research sector.

We've got real skin in the game, not imposed from above. I'm going to now spend a little bit of time talking about the Advanced Analytics and AI Resource Hub, which is a project that I've been involved in. It's a multi-year initiative that began in January of 2024 and will continue through 2028, [inaudible] and I project managed the last year of the project, led by QUT, in partnership with the University of Technology in Sydney and Curtin University. It's about a $700,000 project, with co-investment from the Artik and the Australian National Data Science Network.

We're building, in short, national-scale infrastructure to support the responsible use of AI and advanced analytics for the healthcare research sector. The program proceeded in two phases. In the first half of 2024, we completed the framework development phase, and we basically delivered a nationally informed reference architecture for advanced analytics in healthcare. It looked at federated learning as a practical approach for analysing sensitive health data across distributed environments. And what these two projects revealed is that we have a very fragmented landscape in Australia. We have very limited availability of coordinated infrastructure, tools, training and support for health researchers broadly working with AI.

There were no shared repositories for analytics tools. Training resources that were tailored to the research context were very scarce — plenty on R and Python and general-purpose tools. And critically, there was an absence of governance and ethical frameworks designed specifically for the health AI research contexts — frameworks that researchers have actually used versus just cited. It's interesting to note that the healthcare life science research community is often the first to adopt innovative technologies, especially for pre-competitive research. It was that out of the healthcare life science research community in the US and Europe that formulated the FAIR principles — findable, accessible, interoperable and reusable data.

So it's good to look at certain sectors as the bellwethers for tooling, infrastructure, ethics and governance around new technologies. Our findings directly motivated the second phase, which is infrastructure development, now underway. The three projects that are actively in development are the Frontier Federated Machine Learning Project, which is building a secure federated ML for Australian health researchers. We also are building a National AI Virtual Research Environment — often referred to as a VRE — by UQ and Intersect. And the project where I serve as a project manager and chief investigator is the Advanced Analytics and AI Resource Hub.

Collectively, we're building a national-scale resource hub that provides researchers access to these tools — preferential licensing arrangements, training, curated data assets and trusted analytics environments. And we're delivering it across seven work packages, which is what academics call slice things up into in order to get funding. And we have deliverables. But it's been a really, really quite successful project, and demonstrates sort of the best of collaboration with people who have all been working in this field for many years. In the remainder of this talk, I'd like to focus on the work package that I led — the ethics and governance component.

We did an extensive literature review. And what we found is it's a really cluttered landscape. Everybody in the last three years has put forth a policy, a framework, guidance, etc., including from the key organisations such as the NHMRC — National Health and Medical Research Council — as you would expect. I'll set the scene with the problem of our ethics and governance work. An Australian health researcher — say, running a clinical trial — or even an early-stage PhD student using AI in their work has to navigate, at a minimum, the NHMRC National Statement on Ethical Conduct in Human Research.

They have to comply with the Privacy Act of 1988 and the 13 Privacy Principles since 2025. They also have to respond to the NHMRC Guide for Assessing Research Involving AI, which introduces an AI Questionnaire with screening questions and eight substantive questions structured around seven areas of concern, from human agency and oversight through transparency and explainability to data governance and security. These are people who have an entirely different background than law or ethics and governance. If their research involves [inaudible] Aboriginal and Torres Strait Islander people's data, then they need to take the CARE Principles for Indigenous Data Governance into consideration.

The AIATSIS Code of Ethics. The NHMRC's dedicated ethical conduct guidelines in Keeping Research on Track II, the companion document. They should also be aware of the Voluntary AI Safety Standard and its ten guardrails and they need to swap in the FAIR data principles. And increasingly they need to understand FAIR alone is insufficient, and need to also understand meaningfully the CARE principles. That's at least ten major governance instruments, each with a very distinct scope, terminology and obligations, sitting across multiple regulatory domains. Some are legislative, some are sector guidelines, some are emerging international frameworks.

And if they interact in ways they're not [inaudible] — they interact in ways that are not always obvious. Our primary users are not governance professionals. They're researchers, they're clinicians, they're data custodians or they're research support staff. The question that we hope my colleague and myself to achieve was, how do we make this landscape accessible and actionable? We came up with something that I wouldn't say is particularly unique to us. It's something that I learned about as a tech founder, probably around 2015, which is to segment our market and understand who the product or the service that we were developing was actually for.

We took an archetype-driven approach. The answer that we arrived at was to build a primer, which is not a compliance manual. It's not yet another policy document or a framework, but it's a guided, scenario driven resource that meets researchers where they are and walks them through what they need to know in the context of their day-to-day job. The primer is structured around eight stakeholder archetypes, each representing a distinct role in the Australian health research ecosystem. We have Albert, who's an early-career researcher using AI in his PhD project.

We have Bhavna, a senior professor, advising students and postdocs on the responsible use of technology in epidemiology, her area of expertise. We have Clarissa, a clinical researcher leading a multistate, multi-site trial. Dana, a health data custodian managing population health data sets — for her, all of the work around Indigenous data governance is very relevant. Fatima, who's a health consumer advocate. Grace, a policymaker developing institutional AI guidance. Hugo, a machine learning specialist, building models for clinical deployment. And Iris, a research support professional helping investigators navigate ethics applications.

And that list of A through I is alphabetical. I hope some part of what many of you do is identifiable. That was our hope in building it. Each archetype maps to specific governance instruments and specific obligations, and the design principle is selective depth rather than comprehensive breadth for the people who need to get work done. In the Australian context, Indigenous data is data generated by, about, or for Aboriginal and Torres Strait Islander people, and includes much of what sits in population health data sets and hospital administrative data sets.

This is an example where we need in Australia our own guidance and our own approach. Because these problems are probably not the same problems that maybe Singapore has, or other countries. Or in New Zealand, the Indigenous population treats handling of data quite differently than we do here in Australia. The ARDC published the Framework for the Governance of Indigenous Data in 2026. So if anybody's working in that area, it just came out last month. I highly recommend having a look at it. The lead author is Distinguished Professor Marcia Langton.

It was developed over three years of co-design with Indigenous and non-Indigenous stakeholders. In our primer, Dana is that archetype whose obligations extend across the Ethics and NHMRC guidelines and the CARE principles, as well as Our Knowledge, Our Way guidelines, which came out of CSIRO. We're moving to new models for digital infrastructure. We're moving from the old model that involves vetting a researcher, for example, approving the project, providing access to a data set, relying on contractual institutional controls, because the stuff is just moving too fast for that old model to work.

The new, emerging model is to vet the researcher, approve the project, keep the data inside a controlled environment, monitor activity, have technically controlled outputs with continuously audited access controls. And a lot of this can be automated using AI agents. We are addressing technology sovereignty — namely where the data is processed, by whom and under what jurisdiction. The People Research Data Commons is building infrastructure that keeps sensitive data within sovereign boundaries. And in our governance work, we don't treat governance as an abstract policy layer that gets bolted on to the side.

The choice of analytics environment is a governance decision. The choice of the AI model that we use is a governance decision. The choice of data linkage architecture is a governance decision. The Ethics and Governance Primer is designed to help researchers and data custodians understand these choices as governance choices, and to make them deliberately, not as a bolt-on compliance exercise. They are designed as living documents. They're versioned. They're updated as the governance landscape evolves, which is very, very quickly, as we have heard today, and very responsive, hopefully, to community feedback last night.

In closing, I'd like to return to the question that I opened up with, or is in the title of the talk, which is sovereignty. I started with the observation that companies have announced plans to invest what could scale to more than $100 billion in Australian-based data centres, some owned by Australians, others owned by potentially US interests. This is not philanthropy. It's a market signal. Global hyperscalers are building in Australia because our jurisdiction, our energy potential — especially in states like South Australia — and our geopolitical stability make us a premium location for their infrastructure.

So the question for the research sector is whether we are building alongside of them, or whether we are simply becoming their tenants. Right now, every month that passes without sustained investment in sovereign digital research infrastructure is yet another month in which Australian health researchers become more dependent on commercial platforms that they do not control, processing sensitive data [inaudible] under terms of service they did not write, using models whose training data and decision logic is opaque. These are not hypothetical risks. They are operating conditions of Australian health AI research today.

And the co-investment model shows what is possible. I think it's a great example, and we're very lucky to have NCRIS in this country. With relatively modest government funding leveraged through university partnerships, we've built federated analytics capabilities, a trusted research environment and governance frameworks to be able to use data responsibly. But our program, the Advanced Analytics in Healthcare program, is funded only until 2028. So what happens after that? The infrastructure will not maintain itself. The governance landscape will not stop evolving. The workforce capability that we're building will atrophy if there is no sustained commitment to fund it.

So I'd like to be direct with this audience and say the research sector can't do this alone. We need industry partners — some of whom are in this room — who understand that sovereign research infrastructure is not a cost centre, but rather a competitive advantage. And we need funders who recognise that governance methods, ethics are not overheads that are bolted on to AI projects. They're preconditions for public trust — we've heard a lot about that today — without which no AI system will be deployed in Australian healthcare at scale.

We need government to treat digital research infrastructure with the same strategic seriousness that it applies to defence and energy infrastructure, because AI-enabled economy data infrastructure is national infrastructure. Our Ethics and Governance Primer and the governance matrix are open access resources designed as living documents. And we welcome any feedback from communities that our archetypes represent. And I'd like to leave you with this thought. The tools we have built are only as durable as the infrastructure on which they sit. If we do not fund, partner and build sovereign digital research infrastructure now, while the window is still open, we will find ourselves governing AI systems that we no longer control, with data that has already left the Australian jurisdiction, serving communities whose trust we will not easily win back.

And we saw that with COVID-19. So thank you very much, and I'd love to take any questions you might have.

Greg Sadler

Thank you so much. Just want to ask — [inaudible] you've kind of covered this, Bernadette. But I think people don't appreciate that [inaudible] if we don't invest in this infrastructure. Increasingly, all forms of science are reliant on developing AI models. Every researcher is essentially a GPU user, and Australia is not investing. The entirety of NCRIS is not enough to upgrade our supercomputing facilities. So what do you think is going to happen in terms of our research rankings if we don't actually start to invest in infrastructure?

Obviously you've outlined what the real challenges are for a lot of

Bernadette Hyland-Wood

our health data, but more broadly. Thank you, Sue, for the question. I don't think I'm in a position to predict what will happen to our research rankings, but I would say that we were facing an existential crisis in tertiary education with the loss of a lot of students. We have interesting policies regarding the number of international students that our universities are able to accept. We haven't gotten our house in order in terms of the use of generative AI for assessments, [inaudible] learning and teaching. So we have a lot of areas.

There are a lot of pieces to the puzzle, but I think I would just broadly say that small investments measured in the tens of millions of dollars in the AI Safety Institute and other institutions are insufficient relative to some of our Indo-Pacific partners who are all in. And I agree with the idea of the previous speaker around middle powers collaborating. I think that's a very powerful model. And the US right now is very divided. It's very frantic, and we should take pride in how much technical innovation comes out of this country.

Near where I live in Brisbane, between Montville and the Gold Coast, is the cryptocurrency development capital of the Southern hemisphere. We have these pockets of really in depth and they mostly work for US companies earning US dollars and live a very nice lifestyle between Maleny and the Gold Coast. So I just say we have a lot of institutions. We have, in some ways, a very positive environment to start a company, because you don't have to worry about health care insurance in this country. In the US, that's a very scary thing to go out and be an entrepreneur, because usually you get your health care coverage from your employer.

So we have a lot going for us. And right now the US is very distracted. So this is a great time for us to really pull up our big girl pants and big boy pants and get out there and rock the world in collaboration with our Indo-Pacific neighbours.

Greg Sadler

Thank you. We also had a question from Jennifer. I'll try and paraphrase Jennifer's question, but I'm out over my skis here in terms of expertise. Oh, sure.

Audience question

I'm interested in the possibility that the government is considering to give researchers compulsory access to data from certain kinds of services that are regulated under the Online Safety Act, which is pretty much any digital service that enables access to the internet and includes generative AI services, platforms that distribute AI models and so on. What kinds of structure or frameworks are necessary to ensure that that kind of access is handled in an ethical and productive manner, in your opinion?

Bernadette Hyland-Wood

Let's just say that if you are a curriculum developer, educator, producer of content on YouTube or any number of social media platforms, any kind of sector-specific training — be it for high school students, tertiary education, researchers in health care, hospital administration, etc. — is a huge market. We talk about capability building, but it's a completely, in this world, untapped market. There are lots and lots of opportunities. But first, I think we have educators that need to be educated in how to use this. Three years ago, lecturers and professors were told it's up to you as to whether or not you allow the use of generative AI tools in your class.

For me, who's worked in the field of AI for 25 years, I'm completely comfortable with that. But for many of my colleagues, many of whom are very technical and they have PhDs in physics, they were terrified by this. There's a couple of core groups that we need to educate about how to use AI safely and responsibly. Public sector, public servants. So there are lots of groups, but it's going to take quite a bit of uplift in education to do this, and to do this well. And I think that the earlier session where there was scenario planning, where they gave them a mock board and scenario, was absolutely great training, and we should be doing more and more of that.

There are opportunities for people who have a lot of background in this to do committee and board training. These are busy people who have expertise and other things, but they need to understand how to do scenario planning and have an emergency escalation plan for what happens when they get hacked. And they will. And we will have another pandemic, whether it's in two weeks, two months or two years. We need to really beef up our national health infrastructure so that we don't have a repeat of what happened with COVID-19. Our whole response as a nation was largely based on data from Israel and the UK, because they have really good data-sharing capabilities.

It wasn't because New South Wales or Victoria actually shared their data, because they're so competitive with each other.

Greg Sadler

Thank you very much, Doctor Hyland-Wood. Let's give a round of applause.