AI enabled healthcare

The case for information sharing is strong. Clinicians need timely access to pathology, imaging, medicines and discharge information. Consumers should not have to repeat their history or undergo duplicate investigations. The Australian Government’s Sharing by Default impact analysis identified incomplete information as a constraint on participation, decision making and coordination (*1). 

From 1 July 2026, specified pathology and diagnostic imaging reports must generally be uploaded to My Health Record unless an exception or extension applies. Consumers can request that reports not be uploaded, and later delete, hide or restrict access to reports already present. 

Those controls matter, but their reach is limited. They apply to My Health Record, not necessarily to provider systems, copied information, downstream conclusions or trained models. My Health Record is also not a complete health record (*2). 

Permission drift begins when authority granted for one context is carried into another, such as service improvement, research, model development or supplier reuse. Each approval may appear reasonable, while no single forum sees the full chain from original understanding to downstream effect. 

A 2025 scoping review identified inadequate consent, privacy and security concerns, and unauthorised sharing as barriers to using health information in AI models. Better consent, stronger data protection and ethical standards were key facilitators (*3). Different purposes need defensible authority and a clear explanation. 

Permission can also drift over time. A conclusion can continue to influence care after consent is withdrawn, information is corrected, evidence evolves or the approved purpose ends. Australian privacy guidance states that consent must be current and specific, may be withdrawn, and cannot support future use or disclosure after withdrawal (*4). 

Health services need to know the basis for each use, whether it still applies and how existing outputs should be handled. Consent, care obligations, legislation, expectations and research arrangements can each lead to a different answer. Continuation should be reviewed, not assumed. 

Where care information is later used for model development, consent to receive a service should not be treated as consent for AI training. Leaders should separate authority for care, secondary use and model development, then test whether downstream analytics can still infer what the patient sought to limit (*5). 

Most privacy controls assume sensitivity sits in identifiable documents or fields. AI weakens that assumption because sensitive attributes may be reconstructed from visible or shared information. 

Healthcare has always relied on inference. Clinicians interpret symptoms, histories, results and context to make judgements under uncertainty. AI does not make inference new. It makes inferences easier to reuse across workflows, organisations and decisions. The governance challenge is accountability once consequential inferences become institutional assets. 

Consider a patient who restricts mental health notes because of stigma concerns. A decision support tool may still process medicines, appointment frequency and billing information. No single field contains the restricted diagnosis, but the pattern can operate as a proxy. 

Australian regulatory guidance already recognises this risk. Inferred, incorrect or artificially generated information about an identifiable person can be personal information and generating it through AI constitutes collection under the Australian Privacy Principles. AI-related secondary uses may also be difficult to establish as reasonably expected (*6). 

Recent research shows why this matters (*7). Attackers can sometimes determine with near-perfect accuracy whether a person’s medical data had trained an AI model, even where aggregate testing appeared secure. Risk was higher for underrepresented groups, showing that average results can mask individual privacy risks. 

The point is practical. Health information risk should be assessed by what data can reveal, not only by how source fields are labelled. In Lindenapotheke, the Court of Justice of the European Union held that pharmacy-order information could constitute health data where it revealed information about an identifiable person’s health (*8). For Australia, classification can turn on derivability, not only direct disclosure. 

Fujitsu’s work with the Madrid Health Ministry illustrates the point. The solution integrates genomic and clinical information, using machine learning to assess whether genetic variants are clinically relevant (*9). These use cases can deliver clinical value but require governance that tracks sensitive conclusions as they are generated, reused and acted upon. 

Consumer evidence reinforces the point. In an Australian qualitative study, mental health service users wanted integrated electronic records limited to relevant information and relevant professionals. Although small, the study shows that technical availability does not create legitimacy (*10). 

Once information enters an AI-enabled environment, deletion is rarely a single act. A source record might be removed, a prediction corrected and a derived feature retained, while the model may still reflect the original data. Each raises a different test of safety, feasibility and authority. 

The distinction matters clinically. Removing a result without understanding where it influenced treatment could create safety risk. Retaining an outdated or contested inference could also cause harm. Australian guidance states that clinicians remain accountable for AI outputs informing decisions or records, and should monitor accuracy, relevance and downstream effects (*11). 

Machine unlearning aims to remove specified training data influence without retraining the full model. Some research argues it should be designed into high-risk clinical AI where consent may be withdrawn, evidence may change or bias may be identified. It is not guaranteed. Still, changed authority should trigger a defined process, with technical limits disclosed before deployment (*12). 

The impact extends to consumers. A disputed inference may have influenced triage, referrals or treatment, and consumers may need to know where an incorrect conclusion has propagated. The 2025 Australian Digital Inclusion Index found persistent gaps in access, affordability and digital ability across older people, First Nations people and people with lower educational attainment. Challenge pathways must remain accessible beyond digital channels (*13). 

For Aboriginal and Torres Strait Islander peoples, inference governance also involves collective interests, cultural authority and self-determination. A 2025 CSIRO scoping project identified Indigenous Data Sovereignty, cultural appropriateness, AI health literacy, and Indigenous-led design and oversight as responsible AI priorities (*14). Individual consent may be insufficient where models produce community-level conclusions.

Most health services already rely on risk scores, triage indicators, care-gap recommendations and decision-support outputs. If they cannot identify where those outputs are reused, which workflows depend on them, or how they would be corrected if challenged, the governance problem already exists. 

An inference register is one way to make that problem visible. It should connect source categories, purpose, model version, derived outputs, affected workflows, receiving systems and accountable leader ownership. 

Inference stewardship also changes the investment case. A 2025 systematic review found that AI interventions often reported economic benefits, while indirect costs, infrastructure investment and equity considerations were frequently underreported (*15). Traceability, monitoring, correction, supplier cooperation, revalidation and retirement are lifecycle costs, not administrative extras. 

Four leadership decisions follow. 

1. Separate the things that need to be governed. Distinguish source information, person-level inferences, decisions, derivatives and learned model influence. Assess withdrawal or correction requests against each item. 

2. Assign accountable ownership. A named clinical or operational leader should approve intended use, evidence thresholds, affected populations and operating conditions, with review triggered by withdrawal, correction, purpose change, new safety evidence or material model updates. 

3. Design for traceability and response. Architecture should preserve source and history, track model versions and allow outputs to be restricted or retired. Procurement should test source identification, reuse prevention, correction, unlearning, secure environments, guardrails and post-change evidence. 

4. Make inference risk visible to Boards. Directors and leaders should monitor high-impact systems with named owners, disputed-output review times, unresolved corrections or withdrawal requests, uses beyond approved purpose and models unable to respond to changed authority or evidence. 

The leadership test is simple. If a patient withdrew consent or challenged an inference, could the health service identify what processing would stop, what conclusions remained, which decisions were influenced and what remedies were available? An organisation should not discover after an incident that no one retained the authority, evidence or means to reconsider the output. 

Fujitsu’s Consulting arm, Uvance Wayfinders can help health leaders establish inference stewardship across clinical governance, privacy, operating model design, procurement and service delivery. We trace high-impact use cases from source data to supported decisions, identify gaps in authority, ownership and response capability, and define practical controls for service design, clinical assurance, workforce responsibilities and implementation. 

Fujitsu differentiates through its ability to connect this advisory work with data, AI, cloud and cyber security capabilities. This allows governance decisions to be translated into information lineage, access management, model monitoring, secure environments and lifecycle controls, rather than remaining separate policies or registers. 

Fujitsu’s sovereign AI direction also supports greater control over sensitive information, models and agents through deployment in controlled environments, incremental learning, security monitoring and runtime guardrails (*16). Together these capabilities help health services move from responsible AI principles to demonstrable control over how consequential inferences are created, used, challenged and managed.