Submission to the Joint Select Committee on Artificial Intelligence

Six recommendations on health data, an Australian health model, regulation by actual use, and AI in medical training.

Dr Samantha Pillay OAM

Published by the committee as submission 153 on 24 September 2026. Read it on the Parliament of Australia website.

Introduction

I am a urological surgeon in Adelaide. I have practised for 25 years and was the first woman to train in adult urology in South Australia. I am a member of the Royal Australasian College of Surgeons' AI in Surgery Advisory Group, a published author and an AI filmmaker. This submission is my own and does not represent the College.

It covers terms of reference (c), (e), (f), (g), (i) and (j), where my experience as a clinician, author and filmmaker gives me something to add.

I used AI tools for background research, fact-checking and editorial feedback. The arguments and the writing are my own. Every figure has been checked against its primary source, numbered in the text and listed at the end.

Summary of recommendations

1. Data centre standards should carry not only energy, water and land conditions, but conditions that Australian health, government and public-model data stays in Australia, is not used to train any model without Australian approval, and a set of computers inside the building is reserved for Australian public use.
2. Australian health data should be held by Australian-owned and controlled providers, on their own servers, in Australia, encrypted with keys held here. Location alone should not count as compliance because a foreign owner answers to its own country's laws as well as ours.
3. The Commonwealth should fund an Australian health model: a free open model, trained on Australian guidelines and data, tested here, run here, updated under Australian control, and offered to the public through healthdirect or similar.
4. Indigenous data sovereignty principles, agreed by Aboriginal and Torres Strait Islander leaders in 2018 and recognised by the Commonwealth in 2024, should be written into the AI standards, with funding for Aboriginal and Torres Strait Islander organisations to hold and govern their communities' data on their own servers.
5. AI tools should be regulated by their actual use at scale, not their marketing claims, so that tools used for diagnosis or treatment come under the TGA whatever their fine print says. Suppliers of clinical AI should be responsible for their products under the Australian Consumer Law, and the AI Safety Institute should test those tools, publish results, and take public complaints.
6. The Commonwealth should fund AI-assisted simulation, feedback and video review across health training, so scarce supervisors can train more people to the same standard, with the same tools extended to training and ongoing support for rural and remote doctors. Copyright reform should treat learning from published work as learning, not copying, and make sure a model built for public health in Australia can still be trained on published medical research.

Part A. Sovereignty. Terms of reference (e) and (f)

A1. Data centres. The existing rules cover the building. They do not yet cover what is inside it.

A data centre is a building full of computers. Every time an Australian asks AI a question or a hospital stores a scan, that work happens in one of these buildings. Most of them are not in Australia. If the buildings are overseas, our information lives under another country's laws, and another country's company sets the price and decides whether the service stays on.

The question is not whether Australia should have data centres. We already depend on them. The question is whether they sit here, under our rules, or somewhere else, under someone else's.

On 15 July 2026 the Prime Minister committed to mandatory standards for large data centres: pay for new power supply to match what they use, pay their full share of grid connection, put at least as much energy into the grid as they take out, minimise water use, and be sited away from new housing.1 On 26 August National Cabinet agreed the approach, and the Commonwealth intends to legislate the standards in early 2027.2 I support all of it.

Every one of those obligations is about the building: its power, its water, its block of land. None is about what happens to Australian information inside it, or who controls it. The Prime Minister said Australia can be much more than a data warehouse for AI products made overseas.1 The speech says sovereignty. The standards do not yet.

If all Australia gets is rent, we have gained nothing. Approval of a large data centre should carry four conditions: Australian health, government and public-model data stays in Australia; that data is not used to train any model without Australian approval; a set of computers inside the building is reserved for Australian public use; and the company trains Australian workers.

A2. Australia as the trusted host, on our terms

The Prime Minister listed the assets: space to build, sunlight for renewable energy, critical minerals, a legal and financial system at the top of the global ladder for integrity, and a stable democracy.1

The countries that already host most of the world's data centres are running short of the three things a data centre needs: land, power and water. Ireland is the clearest example. Data centres used about 5 per cent of Ireland's electricity in 2015 and about 22 per cent in 2024, and 97 per cent of them are in Dublin. Ireland never banned data centres. Its regulator rejected a moratorium and set conditions instead, but in a grid already full around Dublin those conditions left new data centres unable to connect there for years, with no ban on paper and something very like one in practice.4 In December 2025 Ireland changed course: new data centres must bring at least 80 per cent of their annual demand as renewable electricity generated in Ireland. Australia's standards should ask the same thing before the grid runs out, not after.3

If Australia writes exemplary rules before the data centres are built we position ourselves as the desired home for foreign investment. When companies need to build offshore, their home voters will ask whether the overseas building meets the standards they demand.

Australia will always buy chips from others. The risk of that dependency is real. Two things reduce it. The chip supply chain needs the minerals Australia has, so the dependence runs both ways. And because we cannot control the chips, we control the two things we can: our data, and the model that runs on the chips.

A3. Where the data sits is not enough. Whose law applies matters more.

A data centre on Australian soil owned by an American company is still an American company. Under the United States CLOUD Act of 2018, a US court can order an American company to hand over data it controls, no matter which country the data is stored in. Australia signed a reciprocal agreement under that Act in 2021.5 Although it is a law-enforcement power, it proves the principle: a foreign company's home law reaches into buildings on Australian soil.

Australia already knows this. The My Health Records Act makes it an offence to hold My Health Record data outside Australia.6 The Commonwealth's Hosting Certification Framework gives its highest certification only to providers that let the Government specify ownership and control conditions.7 The principle exists for government data. It is not applied to health data outside government, where the Privacy Act lets a GP's AI scribe send a consultation overseas as long as the vendor meets equivalent standards.

There is a difference between renting floor space and renting computers. If an Australian company runs its own servers in a foreign-owned building, the landlord does not control the data. If a foreign company runs the servers, it does. For health data the rule should be: an Australian-owned and controlled provider, running its own servers, in Australia, with the data encrypted and the key held here. Then a foreign order gets a box it cannot open.

A4. Own a copy of a free model, teach it Australian health, run it here

Some AI companies release their models free for anyone to download and keep. Once you have the copy, nobody can take it back, raise the price or switch it off. Building a model from scratch costs billions. Adapting a free model to Australian material costs millions, and most of that is people and data, not computers.

The model out of the box starts off having learned mostly from American and European material. Ask it about back pain and it may name the wrong drug, follow the wrong guideline, and know nothing about Medicare or the PBS. It knows least about the Australians who need it most. The fix is to teach the model our guidelines, our drug names, our funding rules and our data.

Health should be first. It is where a wrong answer causes harm and where our guidelines differ from America's. If we can do it for health, we can do it for anything.

Year one is the hard year: find the authoritative Australian documents, clean them, and write the tests that prove the model gives safe Australian answers. From year two, when a better free model appears, we download it and re-run the teaching and the tests we already built. The same as a clinical guideline: the first version takes years, the revision takes months. Nothing changes without our approval. No silent updates.

The argument is control, not savings. This will not be cheaper than a subscription for some years. What it gives is a model that is ours, tested to our standard, that cannot be withdrawn, repriced or altered by a company in another country. You do not run national health infrastructure on something another country's company can turn off. The free models come from American and Chinese companies. Sovereignty means the model is chosen, tested and controlled by Australia, not made here.

Where it runs: test on the supercomputers Australia already owns at NCI in Canberra and Pawsey in Perth, then run at scale on the reserved computers inside commercial data centres on Australian soil. Australia can afford a sovereign facility for health and government data. Where the public reaches it: healthdirect, which is publicly funded,8 is already the national front door for health questions. The Australian health model is its next job: answers only from Australian guidelines and PBS information, chest pain sent to triple zero, every question logged and audited the way hospitals already are.

A5. Indigenous data. The community holds it, or it does not happen.

The Australians whose data is most under-represented in every existing model are also the Australians with the strongest reasons to refuse. That distrust is the correct response to what was done in the past. If Australia builds a health model and Aboriginal and Torres Strait Islander people are absent from it, the model will fail them first, and it will be Australia's own doing.

I am not speaking for Indigenous Australians. I am a clinician saying the national health model will be unsafe without them, and the best way to have them in it is for them to control their part of it.

The principles already exist. Aboriginal and Torres Strait Islander leaders agreed a set of Indigenous data sovereignty principles in 2018: the right to control how data about their communities is collected, stored, analysed and shared, including the infrastructure it sits on.9 In 2024 the National Indigenous Australians Agency published a framework for Commonwealth agencies to apply those principles to the data they hold.10 The model that works is simple: a local organisation owns the data, local people are trained to collect and analyse it, and the data never leaves. Communities keep their data on servers they control, in Australia, and if they want a model, they take the same free model and train it on their own data on their own servers. Nothing leaves. The Commonwealth framework should become part of the AI standards, with funding so communities can do this.

Part B. Health. Terms of reference (c), (g) and (i)

B1. Training more health workers without lowering the standard

Training capacity across health is limited by the time of the people who supervise, not by the number of people who want to train. That is true of nurses, allied health, general practitioners and surgeons alike. You need clinicians to make clinicians, and the thing in short supply is the thing that produces more of it.

AI can be used to augment training through simulation, testing, performance analysis and review. Simulation with automated feedback improves surgical skill: a 2026 systematic review of 14 studies and 814 trainees found a moderate to large benefit.11 For example, in depth analysis of a video of an operation performed by a surgical trainee, augmenting the work of supervising surgeons who are then able to train and supervise more trainees. This model expands supervision, training and ongoing CPD to reach rural and remote areas.

AI literacy for health workers and patients can support our health care system to meet the expanding needs of our ageing population.

B2. Patients ask AI first

My patients already arrive having asked AI about their symptoms. I am not against this, and I could not stop it if I were. More than 300 million people a week ask ChatGPT about their health, by OpenAI's own count.12 That is the front door to Australian healthcare now, and nobody here governs it.

The law asks what a product is marketed as, not what people use it for. Medical device regulation is triggered by the manufacturer's stated intended purpose.13 A chatbot whose fine print says it is not intended for diagnosis or treatment sits outside the TGA while millions use it for exactly that.

In Australia we strictly regulate commercial influence over health advice. A drug company cannot fly me to a resort to influence what I prescribe. Prescription medicines cannot be advertised to the public. Yet an algorithm can steer millions of health decisions at once, shaped by commercial interests the patient never sees, with no register, no code of conduct, and no regulator able to look inside. The same risk we regulate in doctors, at far greater scale, with no rules at all.

Three changes, all within Commonwealth power. First, when a product is used at scale for health decisions, regulate it for what it is used for, not what it claims. Second, regulate where the companies are reachable: they have Australian offices, revenue and government contracts. Require transparency about training data, no model update reaching Australians before it is tested here, and training on Australians' personal data to be opt-in. Third, offer a better door, through healthdirect, as in A4.

Australia has done this before. When the under-16 social media law took effect in December 2025, platforms removed access to 4.7 million accounts within weeks, according to the eSafety Commissioner.14

B3. Who is responsible when AI is used in care

Clinicians remain responsible for the care they provide. Nothing in this submission moves that. Guidance from the Office of the Australian Information Commissioner, the federal privacy regulator,15 confirms that organisations remain responsible under the Privacy Act for personal information they put into or get out of AI tools, and Ahpra guidance16 makes practitioners responsible for safe care whatever tools they use. No argument with either.

The companies selling the tools know this, and their terms are written to match. Heidi, an Australian AI scribe, states in its own terms that the platform is provided as is, without warranties of any kind, that it may not be free from errors, and that the clinician is responsible for verifying every output before relying on it.17 That is the industry position: the product may be wrong, and the clinician carries it. The more AI a clinician uses, the more risk they carry and the less the vendor accepts.

A surgeon is responsible for the operation. The company that made the instrument is responsible if the instrument breaks. Nobody lets the instrument maker sign a contract saying otherwise. A company that sells a defective AI scribe currently can.

The Australian Consumer Law is Commonwealth law and the vehicle. Suppliers of AI used in clinical care should be responsible for their products: disclose what the model was trained on, notify users before the model changes and give them time to check it, and be unable to contract out of liability for defects.

B4. Someone independent must test the tools

Nobody in Australia tests the AI tools that patients and clinicians use. The TGA does not, because they are not classed as medical devices. The companies test their own products, choose which results to publish, and nobody checks. None of the tests are Australian. Every other product that can hurt a patient is tested independently before it is used. AI in health is the exception. Testing means paying people to try to make the tool give a dangerous or biased answer, over and over, recording what happens, repeating it every time the model changes, and publishing the results so a doctor can read one page instead of running their own trial.

The Government has an Australian AI Safety Institute, established this year with $29.8 million over four years. It sits inside the Department of Industry, is advisory rather than a regulator, and its job is to evaluate emerging AI capabilities and advise government.18 Commonwealth agencies are already required, under the Digital Transformation Agency's policy in force since December 2025, to provide a pathway for the public to report AI safety concerns.19 So there is a body and there is a complaints box. Nothing in its remit covers the tools used in clinical care or testing against Australian health standards. The complaints box covers government use only.

Give the Institute, or a body independent of both government and industry, the job of testing AI tools used in clinical and public-facing settings before and after each model change, publishing the results, and taking complaints from the public and clinicians, with the power to reach privately supplied models. Fifty complaints saying the same thing is evidence. This is also what makes B3 workable. Doctors do not have to evaluate the tools if someone credible already has.

Part C. Copyright. Term of reference (j)

C1. Learning is not copying

I am a published author and a filmmaker, and my income does not depend on either. The committee will hear from artists whose living does depend on their work, and their concern deserves respect. I write from both sides of the table, and from medicine.

A musician does not pay for every song they learned from. An author does not pay for every book they read. A surgeon does not pay the surgeons who trained them. I spent fourteen years learning medicine and surgery from everyone who came before me, and nobody traced my skill back to its sources. Every human creator's work comes from a lifetime of learning from the work of others. We do not allow them to copy. We do allow them to learn. I ask the committee to say what makes the machine different before the law treats it differently.

The distinction the law needs is between copying and learning. A machine that reproduces a song or a book has broken existing copyright law and should be pursued under it. A machine that imitates an identifiable artist's voice or likeness and sells it has caused the harm the Prime Minister described, and Australia can legislate against that cleanly at the output end. A machine that learned patterns from millions of published works and produces something new has done what every musician, writer and surgeon has done.

Published is different from stolen. Work obtained illegally or never made public is theft in anybody's language. Work that was published was offered to the world to be read and learned from, and its author was paid the way authors are paid, when the work was bought. The unsettled question is what counts as fair use of published work. If Parliament decides payment for training is required regardless, collective licensing is the only workable mechanism: one licence fee paid to a collecting society that distributes it to authors, the way a café pays one music licence fee to a collecting society rather than negotiating with every songwriter. If the law bans training on published work altogether, the knowledge in books and research becomes unavailable to AI, and the people who lose most are those who could never afford the expert.

Health is the cost nobody is counting. These systems learned medicine from published research. If learning from published work is theft, the Australian health model in A4 is built from stolen goods, and the price is paid by the patient who could never afford the consultation. The Copyright Act needs an express rule that published scientific and medical research remains available for training models used for public health purposes in Australia.

Closing

Thank you for the opportunity to make this submission. I would welcome the opportunity to give evidence to the committee.

Dr Samantha Pillay OAM

Sources

1. Prime Minister of Australia, AI in Australia's interests, speech, University of Sydney, 15 July 2026. https://www.pm.gov.au/media/ai-australias-interests-0
2. Prime Minister of Australia, Meeting of National Cabinet, communique, 26 August 2026. https://www.pm.gov.au/media/meeting-national-cabinet-26-august-26
3. Commission for Regulation of Utilities (Ireland), Large Energy Users connection policy, decision paper CRU/2025/236, 12 December 2025. https://www.cru.ie/publications/27891/
4. RTÉ News, Data centres get to grips with EirGrid's Dublin pause, 16 January 2022. https://www.rte.ie/news/business/2022/0116/1273819-data-centres-eirgrid/
5. Department of Home Affairs, Australia-US CLOUD Act Agreement, signed 15 December 2021; US Department of Justice, The Purpose and Impact of the CLOUD Act, white paper, April 2019. https://www.homeaffairs.gov.au/about-us/our-portfolios/national-security/lawful-access-telecommunications/australia-united-states-cloud-act-agreement
6. My Health Records Act 2012 (Cth), section 77, Requirement not to hold or take records outside Australia. https://www.legislation.gov.au/C2012A00063/latest
7. Department of Home Affairs, Hosting Certification Framework, Service Providers. https://www.hostingcertification.gov.au/service-providers
8. Healthdirect Australia, Our shareholders. https://about.healthdirect.gov.au/who-we-are/shareholders
9. Maiam nayri Wingara Aboriginal and Torres Strait Islander Data Sovereignty Collective, Key principles, 2018. https://www.maiamnayriwingara.org/mnw-principles
10. National Indigenous Australians Agency, Framework for Governance of Indigenous Data, May 2024. https://www.niaa.gov.au/sites/default/files/documents/2024-05/framework-governance-indigenous-data.pdf
11. Godbole GH, Hawkins D, Ewool K, Camacho MHB, Karim R, Patel B. Educational impact of automated feedback systems in surgical training: a systematic review with quantitative synthesis. Journal of Surgical Education 2026;83(4):103879. https://doi.org/10.1016/j.jsurg.2026.103879
12. OpenAI, Launching Health in ChatGPT, 23 July 2026. https://openai.com/index/health-in-chatgpt/
13. Therapeutic Goods Administration, Understanding how we regulate software-based medical devices. https://www.tga.gov.au/resources/guidance/understanding-how-we-regulate-software-based-medical-devices
14. eSafety Commissioner, Platforms restrict access to 4.7 million under-16 accounts across Australia, media release, 16 January 2026. https://www.esafety.gov.au/newsroom/media-releases/platforms-restrict-access-to-47-million-under-16-accounts-across-australia
15. Office of the Australian Information Commissioner, Guidance on privacy and the use of commercially available AI products, 21 October 2024. https://www.oaic.gov.au/privacy/privacy-guidance-for-organisations-and-government-agencies/guidance-on-privacy-and-the-use-of-commercially-available-ai-products
16. Ahpra and National Boards, Meeting your professional obligations when using Artificial Intelligence in healthcare. https://www.ahpra.gov.au/Resources/Artificial-Intelligence-in-healthcare
17. Heidi Health, Scribe Usage Policy and Clinic Agreement, Standard Terms and Conditions. https://www.heidihealth.com/en-gb/legal/scribe-usage-policy
18. Department of Industry, Science and Resources, Australian Government response to the Senate Select Committee on Adopting AI, April 2026; Ministers for Industry, Science and Resources, Establishment of Australian AI Safety Institute, media release. https://www.industry.gov.au/publications/australian-government-response-senate-select-committee-adopting-artificial-intelligence-ai-report
19. Digital Transformation Agency, Policy for the responsible use of AI in government, version 2.0, in effect 15 December 2025, Preparedness and operations. https://www.digital.gov.au/ai/ai-in-government-policy/preparedness-and-operations