Researchers from University Hospital Limerick and the Research Directorate of HSE Mid West argue that hospitals can harness powerful AI while keeping sensitive health information under their own control
Hospitals should not have to choose between embracing artificial intelligence and protecting sensitive patient data, according to researchers from University Hospital Limerick and the Research Directorate of HSE Mid West.
In a new paper published by the Journal of Medical Internet Research, Dr Azrin Muslim and Professor Declan Lyons propose a model of “sovereign AI” that would allow hospitals and health services to use powerful artificial intelligence tools while retaining control of sensitive healthcare information.
The paper, Sovereign Language Models for Regional Health Research under the European Health Data Space, argues that AI language models can increasingly be operated within healthcare organisations’ own controlled computing environments, rather than requiring sensitive information to be sent to external commercial AI services.
This could allow researchers to use AI to review scientific literature, analyse information, generate computer code and support medical research — while confidential patient information remains under the control of the health service responsible for it.
Dr Muslim said the rapid arrival of AI in healthcare makes the question of who controls both the technology and the data increasingly important.
“Healthcare has an enormous opportunity to benefit from AI, but we should not have to give up control of sensitive patient data in order to realise those benefits. The technology is developing to the point where the AI can increasingly come to the data, rather than the data having to be sent to the AI.”
A potential opportunity for regional health systems
The researchers argue that this approach could be particularly important for public health services and regional hospitals, which collectively hold enormous quantities of valuable long-term clinical information but may not have the computing resources available to the world’s largest academic medical centres.
Advances in computing hardware and increasingly capable open AI models mean that sophisticated language models can now potentially be run on infrastructure controlled by healthcare and research institutions themselves.
The paper also describes how hospitals could collaborate in major national and international research projects without pooling individual patient records in a central database.
Using an approach known as federated research, participating hospitals can perform the same analysis locally and share the resulting statistics, while individual-level health records remain within the institution responsible for them.
This could allow regional health systems to contribute to research involving very large populations while retaining local control of their patients’ underlying data.
AI must support — not replace — expertise
The authors caution that keeping AI systems locally controlled does not make the technology infallible.
Large language models can confidently produce incorrect information, invent scientific references and make research findings appear more certain than the evidence warrants. The paper also highlights the risk that excessive dependence on AI could undermine the development of expertise among the next generation of medical researchers.
Professor Lyons said: “AI can produce sophisticated code, analyses and summaries extraordinarily quickly, but speed is not the same thing as understanding. We need researchers who can question these outputs, recognise errors and understand the methods behind them. Investment in AI therefore has to go hand in hand with investment in people and scientific expertise.”
The authors describe a potential risk of “never-skilling” — where researchers become capable of producing increasingly sophisticated analyses using AI without developing the methodological knowledge required to critically evaluate them.
They argue that skills in statistics, study design, critical appraisal and scientific reasoning will therefore become more important in the AI era, not less.
Who controls the AI — and who controls the data?
Rather than unrestricted use of commercial AI or attempts to prohibit the technology, the paper proposes a middle path: AI systems under institutional control, sensitive data kept within governed environments, transparent records of how AI is used, and qualified humans remaining responsible for the science.
The technology is not proposed as an autonomous system for making decisions about individual patients. Instead, its purpose is to help researchers work more effectively while maintaining scientific, ethical and professional oversight.
The central question facing healthcare, the researchers argue, may therefore no longer simply be “Should we use AI?”
Increasingly, it is “Who controls the AI — and who controls the data?”
As health services in Ireland and internationally seek to realise the research potential of clinical data while protecting patient confidentiality and public trust, the authors argue that sovereign AI could provide a practical way forward.