AI in Healthcare: Addressing AI Technologies in Australia's Healthcare System

Authors: Pushti Shah and Nur Wulan Nugrahani

---

INTRODUCTION: SUMMARY OF FINDINGS

This study analyses the increasing use and adoption of Artificial intelligence (AI) tools in the Australian healthcare system by investigating its applications in Part 1 and its risks and benefits to clinicians and patients in Part 2. The paper also scrutinises the existing legal and professional frameworks to determine why accountability for AI-generated errors is allocated to clinicians and healthcare organisations in Part 3.

AI is now actively integrated in Australian healthcare, but adoption is uneven across operational domains. Back‐of‐house applications (such as administrative automation and clinical documentation aids) are the most widely and successfully implemented, and deliver efficiency gains with relatively low direct patient risk. Middle‐of‐house applications such as clinician decision support and diagnostic AI are emerging through pilots and regulated deployments, and provide incremental improvements in safety and throughput. These systems operate under human oversight and within existing clinical workflows rather than autonomously. Front‐of‐house, patient-facing AI such as chatbots and symptom checkers remains limited to experimentation and specific use cases, reflecting prevailing concerns about trust, transparency, and safety in direct patient interaction.

 AI medical scribes illustrate both the benefits and challenges of current AI use. These back‐of‐house systems reduce the time clinicians spend on documentation by an estimated 20-30% per encounter and curtail after-hours paperwork, thereby addressing clinician burnout and improving efficiency. At the same time, they introduce distinct error profiles (notably omissions of important details and hallucinated content with no basis in reality) which can compromise record accuracy and patient safety if left uncorrected. This means that robust human oversight is essential. Practitioners remain responsible for verifying and correcting any AI-generated documentation. Automation bias (over-reliance on AI outputs) is a recognised risk, and independent verification of AI-generated notes is critical to maintaining quality of care.

These error and bias issues highlight broader equity and implementation challenges. Evidence suggests AI scribes are less reliable for culturally and linguistically diverse patients, Indigenous communities, and rural or remote settings due to accent misinterpretation, lack of culturally informed design, and infrastructure limitations. Frameworks for patient consent and privacy protocols regarding AI involvement in care are still developing, underscoring the need for consistent governance.

Under existing legal and professional frameworks, accountability for AI-generated errors remains with clinicians and healthcare organisations. Current laws treat AI tools as extensions of the clinical team, thus doctors and hospitals bear liability for any patient harm from AI outputs, while developers largely avoid direct legal exposure. This allocation preserves the clarity of clinicians’ duty of care but creates a misalignment: those who design and profit from AI often do not share in its risk. As healthcare AI becomes more pervasive and autonomous, the findings of this paper indicate a growing tension between the system-level benefits of AI and the concentrated accountability borne by front-line providers.

---

PART 1: CURRENT ADOPTION OF AI TOOLS ACROSS DIFFERENT OPERATIONAL DOMAINS IN THE AUSTRALIAN HEALTHCARE SYSTEMS

Artificial intelligence (AI) is increasingly embedded in healthcare systems globally, with Australia now firmly in an active adoption phase rather than early exploration. Australia’s National AI Plan published in 2025 establishes an economy-wide framework designed to ensure that AI delivers broad societal benefits while remaining safe, inclusive, and trustworthy. This policy reflects both growing momentum for AI adoption and heightened sensitivity to risk, accountability, and public trust in healthcare settings.

Evidence indicates that adoption in healthcare is uneven. Back-of-house applications, such as administrative automation and system-level analytics, are the most established and lowest risk; middle-of-house applications, including clinician-facing decision support and diagnostic tools, are growing but tightly regulated; and front-of-house, patient-facing applications remain the most constrained due to concerns around consent, transparency, and safety (Australian Government Department of Health, 2025). Despite increasing policy attention, systematic evidence describing where AI is actually adopted in healthcare operations remains fragmented. 

To respond to this evidential gap, this paper prioritises fact-based research by focusing on documented, real-world adoption of AI tools rather than projected potential or speculative benefit. Using an operational lens and drawing on empirical evidence from Australian hospitals and healthcare organisations, this paper examines the current adoption of AI tools across different operational domains in Australian healthcare systems. By synthesising implementation-focused evidence within the Australian context, this study provides an overview of current practical integrations of AI in healthcare operations and where adoption remains limited or emergent.

1: METHODOLOGIES AND SCOPE OF RESEARCH

A structured literature search was conducted to identify empirical evidence on the adoption and implementation of AI tools in Australian healthcare settings. PubMed and Scopus databases were searched for studies published between 2021 and 2026. Eligibility criteria were established to prioritise real-world evidence of AI adoption. Studies were included if they reported on the implementation, integration, or operational impact of AI systems within healthcare organisations, particularly at the organisational or workflow level. Studies were excluded if they involved model performance only without real-world application, were protocols without reported results, or consisted of opinion or commentary pieces without empirical data.

The search initially identified 385 records (PubMed: 272; Scopus: 113). After removal of duplicates and preliminary screening, 192 records were screened based on title and abstract. A total of 95 reports were sought for full-text retrieval, with 91 reports assessed for eligibility. Of these, 66 were excluded based on the predefined criteria. The final evidence base consisted of 25 studies that met all inclusion criteria and were included in the analysis.

2: BACK-OF-HOUSE APPLICATIONS

Back-of-house AI applications, defined as tools operating in non-patient-facing functions, represent the most mature and widespread domain of AI adoption in Australian healthcare. Evidence suggests that adoption often begins in non-clinical areas where risks are lower and operational benefits are more immediate (Bacchi et al., 2021). In this domain, AI is primarily used to enhance administrative efficiency, reduce clinician workload, and improve financial and data management processes. Compared to patient-facing applications, back-of-house systems are more scalable and are increasingly embedded within routine organisational workflows.

A major area of adoption is administrative automation and data analysis. AI tools are used to process large volumes of unstructured healthcare data to improve both service quality and safety. For example, natural language processing (NLP) and large language models (LLM) such as BERT have been applied to automatically classify patient safety incident reports and outperforming traditional methods in identifying incident type and severity across multiple Australian health systems (Wang & Magrabi, 2025). Similarly, semi-automated text-mining tools such as Leximancer are used to analyse patient-reported experience measures to generate actionable feedback for clinicians (Canfell et al., 2024). Automation also extends to registry management, where AI-assisted systems extract and structure data from electronic medical records (EMRs) in the Australian Brain Cancer Registry to reduce the burden of manual data entry (Satheakeerthy et al., 2026). At the workforce level, “algorithmic workforce” tools support clinicians by managing information overload and tracking patient progress, thus freeing up clinicians’ time for higher-value clinical tasks (Bidargaddi et al., 2025).

Predictive analytics represents another key area of back-of-house AI adoption, particularly in managing hospital demand and capacity. Tools such as the Adelaide Score use clinical data to predict patient readiness for discharge, enabling more efficient discharge planning and reductions in length of stay (Kovoor et al., 2025). Complementary systems, such as the Rosetta platform, analyse free-text clinical notes using NLP to enhance predictive accuracy (Kovoor et al., 2025). In community and mental health settings, AI-driven “journey boards” integrate local and national health data to provide real-time oversight of patient caseloads and identify early signs of deterioration (Bidargaddi et al., 2025). These applications illustrate how AI is embedded in organisational decision-making, supporting resource allocation and system-level coordination rather than direct patient care.

Additional applications focus on documentation support. AI scribe’ tools such as Lyrebird Health automate the transcription of clinical encounters and generate structured notes, referral letters, and summaries, reducing administrative burden and after-hours work for clinicians (Evans et al., 2025). 

Despite relatively advanced adoption, implementation of AI in the back-of-house domain is not without challenges. The quality and standardisation of underlying data is a fundamental barrier, as EMR systems often contain inaccuracies, inconsistencies, and outdated information (Satheakeerthy et al., 2026). Implementation can also introduce short-term “digital disruption,” including reduced efficiency during transition phases and issues such as alert fatigue from automated systems (Canfell et al., 2022). Governance and ethical concerns remain significant, including uncertainty around medico-legal liability (Bidargaddi et al., 2025), risks of algorithmic bias, and data privacy concerns (Satheakeerthy et al., 2026). Finally, cultural and cognitive barriers persist as clinicians navigate potential tensions between algorithmic recommendations and professional judgement, raising concerns about over-reliance on automated systems and the erosion of clinical reasoning skills (Bidargaddi et al., 2025; Satheakeerthy et al., 2026).

3: MIDDLE-OF-HOUSE APPLICATIONS

Middle-of-house AI applications support clinicians by assisting with diagnostic interpretation, risk prediction, triage, and clinical decision-making without direct patient interaction. In Australian healthcare settings, adoption in this domain is selective and uneven, but demonstrably growing, particularly in data-intensive specialties (Australian Government Department of Health, 2025). Across included studies, AI tools are most often deployed through pilot or partial implementations, embedded within existing clinical infrastructure such as Picture Archiving and Communication Systems (PACS) and EMRs. Importantly, these systems are consistently positioned as decision-support tools with human evaluation retained, rather than autonomous decision-makers.

The most advanced adoption is observed in diagnostic imaging and radiology, where several AI systems have moved beyond experimental testing into routine use. Large-scale deployment of Annalise CXR, deep-learning chest X-ray interpretation across an Australian radiology network spanning 106 sites, demonstrated radiologist agreement with AI findings in 86.5% of cases, and led to report changes in 3.1% and altered patient management in 1.4% of cases (Jones et al., 2021). In hospital settings, Aidoc ("BriefCase") for intracranial haemorrhage detection identified previously missed cases in 0.07% of scans, functioning as a safety net rather than a primary diagnostic mechanism (Zia et al., 2022). Other radiology applications, including AI-Rad Companion (AIRC) lung nodule detection, achieved sensitivities of 87.3% comparable to specialist radiologists, although it had a low specificity (12.5%) and tended to miss some larger, clinically significant nodules (Mark et al., 2024). Collectively, these findings indicate that radiology AI delivers incremental diagnostic and safety benefits, with effectiveness strongly shaped by baseline system efficiency.

A second cluster of middle-of-house deployment involves disease screening and clinical triage, particularly in settings where AI can prioritise clinician attention. AI-assisted diabetic retinopathy screening implemented in outpatient clinics and Aboriginal Medical Services achieved diagnostic performance exceeding international regulatory benchmarks, with sensitivity 96.9% and specificity 87.7%, while substantially expanding screening capacity in underserved populations (Scheetz et al., 2021). In public oral healthcare, CoTreat, an AI-based triage system demonstrated 98.3% agreement with dentists’ triage decisions and 100% sensitivity for urgent cases (Nguyen et al., 2026). Similarly, Olympus Endo-AID® CADe, computer-aided detection in colonoscopy increased adenoma detection rates from 29% to 41.9% among endoscopists using the system regularly (Rao et al., 2026). However, inconsistent performance in melanoma classification within Australian primary care showed lower-than-expected accuracy (56.6%) and sensitivity (63.2%) compared to international studies, suggesting these tools require further refinement for the Australian primary care context (Miller et al., 2025).

Middle-of-house AI is also increasingly used for patient monitoring, deterioration detection, and clinical surveillance, where systems operate continuously within clinical workflows. A machine-learning-based early warning system integrated into hospital EMRs was associated with a 19% relative reduction in major adverse events, including death and ICU admission (Bassin et al., 2023). In geriatric inpatient units, AmbIGeM System were associated with fewer injurious falls (approximately 0.036 fewer falls per patient) and lower costs (AUD$4,554 less) in the intervention group (Pham et al., 2023). Complementary documentation and surveillance tools, including AI-assisted wound assessment and natural-language-processing–based pressure-injury detection, improved documentation completeness and outperformed traditional clinical coding methods (Barakat-Johnson et al., 2022, 2024; Pilowsky et al., 2026).

Overall, the middle-of-house domain reflects a measured but meaningful adoption trajectory. AI tools are actively deployed in Australian health services and demonstrate quantifiable benefits in diagnostic safety, screening efficiency, and clinical monitoring. However, their impact is typically incremental rather than transformative, constrained by workflow integration challenges, variable performance across contexts, and the continued necessity for clinician supervision.

4: FRONT-OF-HOUSE APPLICATIONS

Front-of-house AI applications, defined as tools that directly interact with patients, represent the most visible yet least adopted operational domain in Australian healthcare. Commonly discussed tools include chatbots, virtual assistants, and symptom checkers. However, evidence from included studies indicates that routine deployment of these tools within Australian hospitals and health organisations remains limited. Their use is most often confined to pilot programs, research studies, or narrowly scoped interventions rather than embedded, ongoing clinical services. This pattern reflects a cautious approach to patient-facing AI, where direct interaction with consumers elevates concerns around safety, trust, consent, and appropriateness of use.

Empirical studies demonstrate that patient-facing chatbots can deliver measurable short-term benefits in specific contexts. Sharp et al. (2025) evaluated an eating-disorder chatbot delivering a single-session intervention to individuals on outpatient waitlists in Australia and reported reductions in eating-disorder pathology, psychosocial impairment, depression, and anxiety, alongside increased subsequent engagement with in-person treatment. Similarly, Bendotti et al. (2024) described a prototype smoking-cessation chatbot in Queensland that users perceived as supportive, non-judgemental, and helpful in maintaining accountability during quit attempts. Despite these positive findings, such applications remain predominantly adjunctive and experimental.

Adoption in this domain is constrained by a combination of emotional, technical, safety, and structural barriers. Patients and clinicians frequently express concern that AI lacks genuine empathy and may weaken therapeutic relationships, with some users reporting reduced openness or disclosure when interacting with chatbots (Barnett et al., 2021). Technical limitations including failures in sentiment recognition, pattern-matching errors, and usability constraints can disrupt engagement and erode trust towards the use privacy (Bendotti et al., 2024; Sharp et al., 2025). Safety and ethical risks are particularly salient in high-risk sectors like eating disorder treatment, where concerns persist regarding crisis detection, misinformation, algorithmic bias, and data privacy (Bendotti et al., 2024; Sharp et al., 2025). Structural challenges such as high dropout rates, uneven demographic engagement, and persistent financial barriers further limit scalability (Sharp et al., 2025; Vandelanotte et al., 2025). 

This review demonstrates that AI adoption in Australian healthcare exists but varies substantially across operational domains and between clinical and non-clinical functions. These studies indicate that AI is currently embedded primarily as a supportive infrastructure within healthcare organisations rather than as a transformative agent of care delivery. Understanding where AI is already operational and where its adoption remains constrained provides a necessary foundation for future evaluation of its clinical, organisational, and societal impact in the Australian health system.

Our analysis now turns to the clinical risk and safety implications of one of the most widely adopted applications: AI medical scribes.

---

PART 2: CLINICAL RISK AND SAFETY IMPLICATIONS OF AI SCRIBES IN AUSTRALIAN HEALTHCARE

 AI medical scribes are ambient listening systems that transcribe and summarise clinical consultations using speech recognition and large language models. In Australian healthcare, these tools are increasingly deployed to support documentation tasks in general practice and hospital settings, where administrative burden and clinician burnout are persistent concerns. Evidence indicates that AI scribes can substantially reduce documentation time per consultation and decrease after-hours record-keeping.

 At the same time, AI scribes introduce distinct safety risks that differ from traditional human documentation errors. These include omissions of clinically relevant information and the generation of fabricated or inaccurate content known has hallucinations. Such errors can compromise the accuracy of the medical record and, if uncorrected, influence downstream clinical decision-making. This section examines the emerging risk profile of AI medical scribes, focusing on documentation accuracy, automation bias, equity implications, and the conditions under which clinical benefits may be realised or undermined.

1: ERROR PROFILES: HUMAN OMISSIONS VS AI HALLUCINATIONS

Traditional human documentation errors predominantly involve omissions. These omissions include instances where clinicians may fail to record certain symptoms, findings or interventions due to time pressure, cognitive overload, or reliance on memory. Research demonstrates that a considerable portion of information from patient encounters is never recorded; for example, a home healthcare study found that roughly 50% of patient problems discussed and 21% of planned interventions were not documented by human providers (Song et al., 2022). Human clinicians may leave out details or produce inconsistent notes, but they rarely invent clinical facts out of whole cloth. By contrast, AI medical scribes exhibit qualitatively different failure modes. Modern AI scribes leveraging large language models tend to have lower overall error rates (often 1-3% of words or statements inaccurate) than older speech-to-text systems (7-11%) (Song et al., 2022). Yet they introduce distinctive errors such as ‘hallucinations’ – content that sounds plausible but has no basis in the actual conversation – and misinterpretations of what was said (Draper et al., 2025). In a comprehensive 2025 evaluation of seven commercial AI scribe products in simulated primary care consultations, omissions accounted for approximately 845 of all AI errors (with a median of 1-6 omissions per consultation), while hallucinations were less frequent (~7-8% of errors) but often more insidious and clinically significant (Draper et al., 2025). For instance, one scribe erroneously documented “no chest pain” even though chest pain was never discussed – a fabricated entry that could mislead clinical decision-making (Draper et al., 2025). Importantly, no AI system produced a completely error-free summary, underscoring the necessity of human review of AI-generated documentation (Draper et al., 2025). Indeed, the majority of AI scribe errors, notably omissions, are subtle and require careful human attention to detect, while the rarer hallucinations and factual inaccuracies, though fewer in number, can be even more harmful if left uncorrected (Draper et al., 2025).

2: PATIENT SAFETY RISKS AND AUTOMATION BIAS

The patient safety risks posed by AI documentation errors differ in nature from those arising from human documentation. Human errors most often involve omissions, which result in gaps in the record that may cause delayed or suboptimal care such as a missed drug allergy or an unrecorded symptom contributing to misdiagnosis. By contrast, AI hallucinations insert false facts into the medical record, actively misinforming clinicians who later rely on those notes (JAMA Network, 2024). Earlier voice-recognition systems have incorrectly transcribed critical phrases with serious consequences. In one case, “normal vascular flow” was recorded as “no vascular flow”, leading to an unnecessary invasive procedure (Song et al., 2022). In another, a transcription error miscommunicated tumour location and contributed to surgery being performed on the wrong site (Draper et al., 2025). Although modern AI medical scribes feature improved accuracy and contextual understanding, these examples demonstrate the persisting risk to patient safety if clinicians do not carefully review AI-generated records. The Royal Australian College of General Practitioners (RACGP) acknowledge AI scribes’ potential benefits but explicitly warn that they “can make mistakes, misinterpret or omit important information, or even hallucinate content”, and emphasises that uncorrected errors “can have serious patient safety consequences” (AHPRA, 2024).

Automation bias which refers to the tendency to place too much trust in automated systems even when they make errors, is one major concern. Experienced clinicians may be lulled by the polished, comprehensive appearance of an AI-generated note and become less vigilant in scrutinising it (Parasuraman & Manzey, 2010). The RACGP guidance cautions that as AI scribes gain popularity, general practitioners risk “becoming over-reliant on their use and paying less attention to critical clinical details or forgoing the vital process of checking the output… resulting in errors that could affect patient safety" (AHPRA, 2024).

3: OPERATIONAL BENEFITS AND CLINICIAN EXPERIENCE

In alleviating administrative burden, AI scribes can contribute positively to care quality and clinician well-being, provided they are used prudently. By handling much of the note-taking, these tools allow clinicians to engage more fully with patients during consultations (Evans et al., 2025). A multicentre study in 2025 found that after one month of using an ambient AI scribe, clinician burnout rate dropped from 51.9% to 38.8%, and doctors saved nearly one hour per day in after-hours documentation (Evans et al., 2025). However, if clinicians or systems treat AI notes as reliable ‘out-of-the-box’, the efficiency gains could be offset by new errors and the efforts needed to correct them. Moreover, there is a risk that healthcare organisations might respond to efficiency gains by simply increasing clinical workloads, such as scheduling more patients per day, rather than giving clinicians more breathing room (Bidargaddi et al., 2025). In such cases, the intended relief to clinician burnout could diminish. Thus, the operational impact of AI scribes is context-dependent. They appear most beneficial in settings that implement them thoughtfully, ensuring that increased efficiency translates into improved patient interactions and not merely productivity pressure (AHPRA, 2024; Evans et al., 2025). Crucially, regulators and professional bodies emphasise that AI scribes are an aid, not a replacement for clinical skill (AHPRA, 2024). As a result, doctors must continue to use their judgement and maintain thoroughness when using AI scribes (AHPRA, 2024).

4: EQUITY AND IMPLEMENTATION CHALLENGES

 Several important issues relating to AI scribes remain underexplored in the Australian context, especially regarding equity and diverse practice settings. One major gap concerns performance for culturally and linguistically diverse (CALD) populations. Speech recognition systems often perform less accurately for non-native English speakers and for speakers with certain accents. Error rates in speech recognition systems have been reported up to 11% points higher on strong non-native accents compared to native speakers (Akano et al., 2024; Nguyen et al., 2026). Given that more than a quarter of Australians speak a language other than English at home, an AI scribe may systematically mis-transcribe consultations with CALD patients, potentially leading to significant omissions or mistakes in their records (ABS, 2024). Without adaptation or careful awareness of a user’s background, these tools may exacerbate existing communication barriers and health disparities. 

Indigenous health is another area of concern. Aboriginal and Torres Strait Islander patients often use Aboriginal English or languages that standard AI models are not trained on, raising the risk of misinterpretation or culturally inappropriate transcription (CSIRO, 2025). CSIRO (2025) underscores the necessity of Indigenous data sovereignty and culturally informed AI design as currently the mainstream AI tools lack these considerations. If introduced in Indigenous communities without such safeguards, an AI scribe might produce inaccurate or insensitive documentation, undermining patient trust and care quality.

Rural and remote healthcare contexts also face unique hurdles. Many remote clinics have bandwidth limitations, so cloud-based AI scribes might be unreliable or unusable. Additionally, scribes validated in big-city clinics might not handle the broader scope of rural general practice consultation (Bressan et al., 2025). Notably, the methodology used in this project did not identify any Australian study that evaluates AI scribe performance in rural or Indigenous settings, which may leave a gap in practitioner knowledge of these tools.

In addition, patient consent and privacy issues loom large in the context of AI scribe use in healthcare. Both RACGP (2024) and AHPRA (2024) advise that patients must be informed and give consent if an AI scribe is recording their consultation, especially since audio recording and notes may be processed using external servers, including those located overseas. Failure to obtain consent could breach privacy laws or even surveillance device legislation in some jurisdictions (AHPRA, 2024). In practice, approaches vary: some clinics obtain written consent (as recommended by medical defence organisations), others rely on signage or verbal disclosure, and a few early accounts indicate some patients were not properly informed of AI scribe use at all (The Conversation, 2025). In summary, safe and fair adoption of AI scribes will require consideration of contextual factors, such as adapting for CALD speech, ensuring culturally safe use in Indigenous care, securing robust data protection measures, and standardising patient consent protocols, none of which are fully resolved yet in Australia.

5: DOCUMENTED EVIDENCE OF AI-RELATED HARM

Despite growing concern, confirmed cases of patient harm directly caused by AI in healthcare remain relatively few to date, likely due in part to underreporting and early intervention. A 2026 systematic review identified 295 documented AI-related health incidents globally between 2012 and 2025, with most cases in the US and UK and none explicitly in Australia (Denecke et al., 2026). The authors emphasise that these figures almost certainly underestimate the true extent of harm, as most health systems were found to recommend unsafe or incorrect cancer treatments during internal tests (though these were caught by oncologists before reaching patients (Ross & Swetlitz, 2018). A recent analysis by the US’ Food and Drug Administration found logged 108 reports (2019-2023) of suspected AI or machine-learning errors in devices, including missed radiology findings and dosing miscalculations (Handley et al., 2024). The legal implications are also emerging. The case of Lowe v Cerner Corp. in the USinvolved a negligence lawsuit against an EHR software vendor after a patient suffered brain damage when a doctor’s monitoring order was lost due to a design flaw. The Court allowed the claim against the software maker to proceed, signalling that in egregious cases, healthcare IT companies might share liability when their products contribute to harm. In Australia, documented harm has thus far arisen more in consumer-facing contexts. For example, an ABC investigation reported a case where an AI mental health chatbot encouraged a 13-year-old experiencing depression to commit suicide (McLennan, 2025). Such incidents highlight the extreme consequences that can ensue outside of the formal healthcare system if AI tools malfunction without proper safeguards. To date, there are no public reports of patient injury directly tied to an AI scribe in Australia, but given the identified error risks, experts caution it is plausible that incidents could occur if oversight lapses (AHPRA, 2024). 

AI medical scribes evidently do provide real benefits to the Australian healthcare system by reducing administrative load, potentially improving documentation completeness, and freeing clinicians to focus more on patients. At the same time, they introduce  new qualitative risks, from hallucinated content and transcription errors to amplification of biases or communication gaps, that require careful management. Whether AI scribes can deliver a net critical benefit will depend on how use of this tool is implemented and governed. With robust human oversight, standardised consent and privacy protections, validation across diverse patient groups, and alignment with professional standards, these tools can likely improve efficiency and even safety by ensuring more complete documentation without endangering patients. However, without these measures, AI scribes could undermine safety and equity, introducing errors that clinicians miss and reinforcing disparities (Chin et al., 2025; Price et al., 2019). In effect, their net benefit remains uncertain and conditional. The requirement for clinician vigilance leads to the question of who bears responsibility for AI-related errors that cause harm. The next section examines how Australia’s current legal frameworks allocate responsibility among clinicians, healthcare organisations, and AI developers, and why, at present, the burden of AI risk falls squarely on clinicians and organisations rather than on the makers of the AI.

---

PART 3: LEGAL ACCOUNTABILITY FOR AI SCRIBE ERRORS — HOW AUSTRALIAN FRAMEWORKS ALLOCATE RESPONSIBILITY

The integration of AI tools into healthcare has not altered the foundational structure of legal accountability in Australia. Under existing negligence law and professional regulation, clinicians and healthcare organisations remain responsible for patient harm arising from clinical care, including care supported by AI systems. AI tools are treated as instruments within clinical workflows rather than independent actors.

 This section examines how Australian legal frameworks allocate responsibility when AI medical scribes contribute to clinical error. It shows that liability continues to concentrate on clinicians and institutions, while developers are largely insulated through regulatory classification, contractual arrangements, and evidentiary distance from patient harm. The analysis highlights how this allocation preserves patient protection through human oversight but also generates tensions between technological control and legal responsibility.

1: BACK-OF-HOUSE AI TOOLS: CLINICIANS AS ACCOUNTABLE GATEKEEPERS

The prevailing legal position in Australia regarding back-of-house healthcare AI tools such as administrative automation tools is that clinicians act as accountable gatekeepers for the AI they use. AI medical scribes are the paradigmatic example. As discussed in Section 2, these tools generally fall outside the definition of a regulated medical device because they support documentation rather than provide therapeutic or diagnostic functions (AHPRA, 2024; Australian Government Department of Health, 2025). Accordingly, AI scribes are not subject to TGA regulation (AHPRA, 2024).

This regulatory gap has practical consequences. In the absence of formal safety or efficacy obligations under therapeutic goods law, AI scribe developers manage risk primarily through contract, using disclaimers and indemnity clauses that shift responsibility to clinician users or healthcare organisations (HealthLegal, 2023; Meridian Lawyers, 2024). At the same time, Australian professional regulators have held that using AI does not dilute a practitioner’s personal responsibility for patient care. AHPRA’s 2024 guidance states that “regardless of what technology is used… the practitioner remains responsible for delivering safe and quality care,” and that  “the practitioner is responsible for checking the accuracy and relevance of records created using generative AI” (AHPRA, 2024). The RACGP adopts the same position, emphasising that “GPs are ultimately responsible for ensuring the patient health record is accurate and up to date” when AI scribes are used (AHPRA, 2024).

These positions reflect settled negligence principles. A clinician’s duty of care extends to all aspects of patient management, including accurate record-keeping and verification of information relied upon in decision-making (Rogers v Whitaker (1992) 175 CLR 479; Medical Board of Australia, 2020, cl 8.4). Practitioners cannot evade liability by attributing errors to tools, just as they cannot escape responsibility by delegating tasks to inadequately supervised assistants (Albrighton v RPAH (1980) 2 NSWLR 542, 568). In legal terms, the AI scribe is treated as a tool under the clinician’s control. Thus, clinicians remain responsible for interpretation and verification of outputs.

This allocation of responsibility is illustrated by the following hypothetical. Dr Green, a GP, uses an AI scribe during consultations. When Ms Lee presents with subtle chest symptoms, the AI scribe hallucinates an entry stating, “no chest pain”, a statement Ms Lee never made. Dr Green fails to notice the error and signs off the record. Relying on the erroneous note, he attributes the symptoms to anxiety and does not investigate further. Ms Lee later suffers a heart attack that timely testing could have prevented.

Dr Green owed Ms Lee a duty to exercise reasonable care in diagnosis and record-keeping (Rogers v Whitaker). A court would likely find a breach of duty if Ms Lee made a negligence claim,as a competent GP would have clarified the presence of chest pain or questioned the unexplained negative finding generated by the AI (Civil Liability Act 2002 (NSW) s 5B). The harm, a preventable cardiac event, is causally linked to that failure. Liability would fall on Dr Green and, vicariously, on his clinic employer. The AI developer would not be directly liable. As the scribe is only a general documentation tool, the clinician’s failure to verify a critical detail is treated as the proximate cause of harm (Nolan & Matulionyte, 2023).

Regulators and insurers reinforce this conclusion. AHPRA (2024) and medical indemnity providers emphasise that AI does not alter the fundamental rule that practitioners must apply independent judgment to all outputs. One insurer summarises the position succinctly: “Healthcare professionals are ultimately responsible for reviewing and amending any AI-generated health records to ensure accuracy” (Avant, 2025). Australian negligence law therefore provides no ‘automation defence’ to practitioners over-relying on AI. Any substandard care resulting from such over-reliance constitutes negligence like any other deviation from professional standards (Chin et al., 2025). In short, current frameworks allocate virtually all responsibility for back-of-house AI errors to the clinician user.

2: MIDDLE-OF-HOUSE AI TOOLS: REGULATED BUT STILL UNDER HUMAN OVERSIGHT

Middle-of-house AI tools, including clinician-facing diagnostic support systems and predictive algorithms, typically have a more direct clinical impact and are more likely to qualify as ‘medical devices’ under the Therapeutic Goods Act 1989 (Cth). Where marketed for therapeutic purposes, these tools are subject to TGA regulation and must be listed on the Australian Register of Therapeutic Goods (Australian Government Department of Health, 2025; TGA, 2023). Examples include AI systems for chest radiograph analysis (such as Annalise CXR) or sepsis risk detection in hospital inpatients.

Regulatory approval imposes safety and quality obligations on developers and, in theory, enables product liability claims where a safety defect in an AI device causes harm (Competition and Consumer Act 2010 (Cth) Sch 2 (ACL) s 9). However, in practice, liability for errors involving middle-of-house AI continues to lie primarily with clinicians and healthcare organisations, because these systems are deployed as decision-support tools with human oversight. Clinicians are expected to critically interrogate AI outputs and correlate them with clinical context.

If a radiologist misses a fracture that both they and an AI system failed to identify, the claim will be one of professional negligence. Conversely, if an AI flags a lesion and the radiologist unreasonably ignores it, liability likewise rests with the clinician for failing to act on available information. In both cases, the clinician’s conduct is assessed against the professional standard of care (Prictor, 2022). Manufacturer liability may arise only where the AI tool is demonstrably unsafe across its intended use, such as systematic failures no reasonable user could detect. Even then, patients typically pursue claims against clinicians and hospitals first, with any contribution claims against vendors remaining secondary (Chin et al., 2025).

To date, no Australian case has held an AI software provider directly liable for patient harm, whereas clinicians are routinely sued for diagnostic errors. TGA approval does not alter this reality. As AHPRA makes clear, “TGA approval does not change a practitioner’s responsibility to apply human oversight and judgment” (AHPRA, 2024). AI systems are therefore treated as part of the clinical workflow, not as independent legal actors. Even in rare scenarios where a vendor might be drawn into litigation, such as the US case Lowe v Cerner (2022), such outcomes remain exceptional. In Australia, in the absence of comparable precedent, clinicians cannot rely on regulation or approval to shield them from negligence liability when AI is involved (Nolan & Matulionyte, 2023).

3: FRONT-OF-HOUSE AI TOOLS: MINIMAL ADOPTION UNDER LIABILITY UNCERTAINTY

Front-of-house AI tools, such as patient-facing symptom checkers, triage chatbots, and virtual assistants, remain the least adopted category in Australian healthcare. A key reason is the uncertain and potentially expansive liability exposure they entail (Australian Government Department of Health, 2025). Where patients interact directly with AI without continuous human moderation, it becomes unclear who owes the duty of care for the advice provided.

If a hospital-endorsed chatbot erroneously advises a patient not to seek urgent care and harm results, a court may find that the healthcare provider owed, and breached, a duty to take reasonable care in deploying and supervising the system. Because this terrain remains legally untested, providers have tended to avoid widespread deployment, limiting front-of-house AI to tightly controlled pilots or research settings. Disclaimers stating “this is not medical advice” are common, but their legal effectiveness is uncertain.

Most patient-facing AI currently functions only as an adjunct to clinician review, such as intake questionnaires or appointment triage tools. This caution is driven not only by legal uncertainty but also by ethical and reputational concerns about trust, empathy, and safety (Cohen et al., 2023). Where harm does occur, patients’ most viable claims are likely to lie against the healthcare organisation, under theories of direct or vicarious negligence, rather than against remote developers with whom the patient has no relationship. This liability uncertainty has itself constrained adoption, leaving front-of-house AI marginal within Australia’s health system. Until clearer statutory frameworks emerge addressing performance standards, consent, and responsibility, front-of-house AI is likely to remain limited, with liability presumed to rest on service providers rather than developers.

4: NEGLIGENCE LAW FUNDAMENTALS: WHY LIABILITY FALLS ON CLINICIANS

The above patterns reflect fundamental doctrines of negligence which have not been upended by AI’s arrival. First, the duty of care is non-delegable in effect – health professionals owe duties to their patients and cannot avoid them by citing reliance on a device or algorithm (Rogers v Whitaker). Courts approach AI mishaps by examining the clinician’s actions rather than treating the AI as an independent party with its own duty (Price et al., 2019). Second, the standard of care for using AI is judged by what a competent peer would do. Currently, that means exercising caution and scrutinising AI output. The Bolam-style “peer professional opinion” defence in civil liability legislation (e.g. Civil Liability Act 2002 (NSW) s 5O) is unlikely to excuse poor oversight of AI, since authoritative bodies like AHPRA and the RACGP urge careful verification, not blind trust. Unless or until using AI without checking becomes widely accepted (which is doubtful), failing to verify AI output will likely remain a breach of the standard of care. Third, causation in negligence still focuses on the human contribution. In multi-factor scenarios (AI error + human error), Australian courts use the ‘but for’ test and common-sense principles (Wallace v Kam (2013) 250 CLR 375). If a patient’s harm could have been avoided but for the clinician’s lapse (e.g. not catching an AI’s mistake), causation is satisfied. The presence of AI doesn’t break the chain; at most it is considered part of the factual matrix. Finally, vicarious liability ensures that even if an individual clinician is faultless, but the system (managed by the hospital) failed, the institution can be brought to account. For example, if a hospital mandated use of a poorly performing AI and discouraged clinicians from double-checking it, the hospital’s systemic negligence could be found. But typically, a plaintiff will sue both the individual clinician and the hospital (the latter being vicariously liable for the former’s negligence). The upshot is that for a patient it’s often unnecessary to pursue the AI’s maker, because the treating doctor and hospital collectively provide adequate avenues for redress under established law.

5: CONTRACTUAL AND INSURANCE FACTORS REINFORCING CLINICIAN LIABILITY

 In practice, contractual arrangements and insurance policies further cement that if an AI causes harm, the healthcare provider, not the developer, will bear the cost. Many AI vendors, as noted, have contracts requiring users to assume liability. Hospitals and clinics that agree to these terms effectively indemnify the developer, meaning if a patient tries to sue the developer, the contract shifts that burden back to the provider. Meanwhile, medical indemnity insurance (compulsory for practitioners) typically covers negligence claims against the doctor but does not cover product liability or contractual indemnities (Avant, 2025). So, if a doctor or practice has, for instance, agreed to hold the AI company harmless, their standard insurance might not cover the resulting liability, leaving them financially exposed. This stark reality is leading some Australian medicos to avoid or carefully limit AI tool use, or to seek explicit assurances from insurers (Avant, 2025).

Thus, as things stand, the risk of AI-related litigation effectively falls on the same parties as any other medical error: clinicians and their employers (who are insured accordingly). AI developers, by contrast, usually are not carrying comparable liability insurance for patient injury, reinforcing their peripheral role in claims. On a policy level, this misalignment – that those creating the high-risk tech are not necessarily on the financial hook when it fails – raises concerns about whether incentives for safety are optimally placed (Chin et al., 2025). But legally, it is a consequence of the current framework that sees technology as part of the clinician’s toolbox.

6: EMERGING PERSPECTIVES AND THE NEED FOR REFORM

The concentration of liability on clinicians and providers in Australia’s current approach has drawn critique. Some commentators worry about a “responsibility trap”: clinicians are expected or pressured to use AI to improve efficiency, yet if it goes wrong, they personally shoulder all blame (Matulionyte & Baker, 2021). This could lead to defensive behaviours, e.g. avoiding beneficial tech due to liability fears, or conversely, using it but over-ordering tests to double-check everything.

Comparatively, jurisdictions like the European Union are moving in a direction that would spread responsibility more evenly. The EU’s AI Act (adopted 2023) will require AI developers to incorporate human oversight mechanisms and proactively address known risks, effectively placing some legal duty on developers to facilitate safe use (European Commission, 2023, p.7). The EU’s updated Product Liability Directive (2024) will make it easier for injured patients to sue AI product manufacturers via strict liability for defects and presumptions that lighten the evidentiary burden in complex AI cases (European Commission, 2023, p.7-9). These reforms reflect a normative stance that AI creators should not be entirely immunised from liability and that patients should not be worse off in proving a case just because AI’s involved.

In Australia, no AI-specific legislation has been passed to date. The Department of Health and Aged Care’s 2025 Regulatory Review did highlight that current laws might leave uncertainty and recommended clarifying how AI tools are classified and how legal liability is handled (Australian Government Department of Health, 2025). Any changes, however, are still under consideration. For now, the combined effect of negligence law, regulatory guidance, contracts, and insurance means the risk of AI-related errors is de facto carried by clinicians and their organisations. From a patient-safety perspective, this keeps end-users vigilant. From an innovation perspective, it may slow adoption or put unfair pressure on clinicians. Until reforms are enacted, the prudent course for Australian healthcare providers is to treat AI tools as having zero liability shield, i.e. use them carefully, document decisions, inform patients, and always double-check critical outputs (AHPRA, 2024; Avant, 2025). Only with such caution can the benefits of AI scribes and similar tools be harnessed without compromising legal and ethical duties. Ultimately, under current frameworks AI has not altered ‘who answers for harm’, that stills remains to be the human carer, but it has raised new questions that future law and policy will need to address to ensure a fair and effective allocation of responsibility as AI becomes more pervasive in care.

 AI has progressed from a potential disruptor to an operational reality in Australian healthcare, but its adoption remains uneven and selective. Non-patient-facing tools are well integrated and delivering clear efficiency gains, while clinician-assistive systems are emerging under careful oversight in tightly controlled conditions. By contrast, direct patient-facing AI remains largely confined to pilots and niche uses due to unresolved safety and trust concerns.

AI medical scribes exemplify these dynamics. They significantly reduce documentation time and after-hours work for clinicians, alleviating burnout and improving workflow efficiency. At the same time, they introduce new risks: AI-generated notes can omit important details or include fabricated hallucinations, requiring rigorous human verification to prevent patient safety lapses. Automation bias can further erode clinical vigilance if practitioners over-rely on polished AI outputs. These findings underscore that human oversight remains essential, a theme that extends across all current AI applications.

The evidence highlights important equity and context challenges. AI tools often perform less reliably or appropriately for culturally and linguistically diverse patients, Indigenous communities, and in resource-constrained settings, raising questions about inclusivity and safe use. Additionally, deployment is outpacing the development of standardised consent procedures, data governance, and risk management practices.

Legally and ethically, accountability for AI outputs still resides with human carers. Under existing law and professional standards, clinicians and healthcare institutions bear responsibility for any harm associated with AI use, whereas technology developers largely avoid direct liability. The safety net for patients thus remains the human duty of care, ensuring oversight but also creating a misalignment between those who control and benefit from AI and those who shoulder its risks.

These findings point to emerging tensions in how Australian healthcare will manage AI’s growth. Maximising AI’s benefits while maintaining public trust will depend on aligning technical innovation with clear, fair accountability. As AI becomes more pervasive and autonomous, addressing this mismatch between technological control and legal responsibility will be critical to the sustainable and legitimate integration of AI in healthcare.

---

REFERENCES

 Albrighton v Royal Prince Alfred Hospital (1980) 2 NSWLR 542 (NSW Sup Ct).

Australian Bureau of Statistics. (2022, June 28). Cultural diversity: Census. ABS. abs.gov.au

Australian Government Department of Health and Aged Care. (2025). Safe and responsible artificial intelligence in health care – Legislation and Regulation Review (Final report). Canberra, ACT: Commonwealth of Australia.

Australian Government Department  of Industry, S. and R. (2025). National AI Plan.

Australian Health Practitioner Regulation Agency (Ahpra). (2024). Meeting your professional obligations when using artificial intelligence in healthcare (Guidance document). Ahpra. (Accessed via Ahpra public guidance resources.)

Avant Mutual. (2025, January 7). AI scribes and patient consent. (Risk advisory). Retrieved from https://www.avant.org.au/resources/ai-scribes-and-patient-consent

Bacchi, S., Gluck, S., Koblar, S., Jannes, J., & Kleinig, T. (2021). Improving the accuracy of stroke clinical coding with open-source software and natural language processing. Journal of Clinical Neuroscience : Official Journal of the Neurosurgical Society of Australasia, 94, 233–236. https://doi.org/10.1016/j.jocn.2021.10.024

Barakat-Johnson, M., Jones, A., Burger, M., Leong, T., Frotjold, A., Randall, S., … Coyer, F. (2022). Reshaping wound care: Evaluation of an artificial intelligence app to improve wound assessment and management amid the COVID-19 pandemic. International Wound Journal, 19(6), 1561–1577. https://doi.org/10.1111/iwj.13755

Barnett, A., Savic, M., Pienaar, K., Carter, A., Warren, N., Sandral, E., Manning, V., & Lubman, D. I. (2021). Enacting ‘more-than-human’ care: Clients’ and counsellors’ views on the multiple affordances of chatbots in alcohol and other drug counselling. International Journal of Drug Policy, 94. https://doi.org/10.1016/j.drugpo.2020.102910

Bassin, L., Raubenheimer, J., & Bell, D. (2023). The implementation of a real time early warning system using machine learning in an Australian hospital to improve patient outcomes. Resuscitation, 188, 109821. https://doi.org/10.1016/j.resuscitation.2023.109821

Bendotti, H., Lawler, S., Ireland, D., Gartner, C., & Marshall, H. M. (2024). Co-designing a smoking cessation chatbot: Focus group study of end users and smoking cessation professionals. JMIR Human Factors, 11, e56505. https://doi.org/10.2196/56505

Bidargaddi, N., Patrickson, B., Strobel, J., & Schubert, K. (2025). Digitally transforming community mental healthcare: Real-world lessons from algorithmic workforce integration. Psychiatry Research, 345. https://doi.org/10.1016/j.psychres.2024.116339

Bressan, T., Bakke, A., & Zelek, B. (2025). AI scribes in rural and remote primary care: An antidote to physician burnout or Pandora's box? Rural and Remote Health, 25(1), Article 9430. https://www.rrh.org.au/journal/article/9430/

Canfell, O. J., Chan, W., Pole, J. D., Engström, T., Saul, T., Daly, J., & Sullivan, C. (2024). Artificial intelligence after the bedside: Co-design of AI-based clinical informatics workflows to routinely analyse patient-reported experience measures in hospitals. BMJ Health & Care Informatics, 31(1), e101124. https://doi.org/10.1136/bmjhci-2024-101124

Canfell, O. J., Meshkat, Y., Kodiyattu, Z., Engstrom, T., Chan, W., Mifsud, J., Pole, J. D., Byrne, M., Raders, E. Van, & Sullivan, C. (2022). Understanding the Digital Disruption of Health Care: An Ethnographic Study of Real-Time Multidisciplinary Clinical Behavior in a New Digital Hospital. Applied Clinical Informatics, 13(5), 1079–1091. https://doi.org/10.1055/s-0042-1758482

Chin, K. Y., Blake, M., & Rizzi, M. (2025). Civil liability for AI clinical tools in Western Australia: A critical overview. University of Western Australia Law Review, 53(2), 121–156. 

Denecke, K., Gabarron, E., Petersen, C., & Merolli, M. (2026). Global landscape of AI-related incidents in healthcare (2012–2025): A systematic review. Journal of Medical Internet Research, 28(3), e199. 

Draper, T. C., Cox, T., Lamb-Riddell, K., Moretti, L. A., McCormick, J., Trowell, S., Kiely, J., & Luxton, R. (2025). Clinical AI scribes in primary care: Accuracy, error severity and implications for clinical practice. BMJ Digital Health & AI, 1(1), e000092. https://doi.org/10.1136/bmjdhai-2025-000092

Evans, K., Papinniemi, A., Ploderer, B., Nicholson, V., Hindhaugh, T., Vuvan, V., Cowley, N., Tariq, A., & Thomson, H. (2025). Impact of using an AI scribe on clinical documentation and clinician-patient interactions in allied health private practice: perspectives of clinicians and patients. Musculoskeletal Science & Practice, 78, 103333. https://doi.org/10.1016/j.msksp.2025.103333

Goodman, A., & Australian e-Health Research Centre. (2025). Artificial intelligence for healthcare in Australian Indigenous communities: Scoping project to explore relevance. CSIRO. https://publications.csiro.au/publications/publication/PIcsiro:EP2025-3601

Handley, J. L., Krevat, S. A., Fong, A., et al. (2024). Artificial intelligence–related safety issues associated with FDA medical device reports. npj Digital Medicine, 7, Article 351. DOI: 10.1038/s41746-024-01357-5

Handley, S. H., Bell, D., & Raubenheimer, J. (2024). Adverse events involving artificial intelligence in medical devices: analysis of the U.S. FDA’s database (2019–2023). Resuscitation, 188, 109823. https://doi.org/10.1016/j.resuscitation.2023.109823 (Analysis of FDA adverse event reports linked to AI/ML failures.)

Jones, C. M., Danaher, L., Milne, M. R., Tang, C., Seah, J., Oakden-Rayner, L., … Esmaili, N. (2021). Assessment of the effect of a comprehensive chest radiograph deep learning model on radiologist reports and patient outcomes: A real-world observational study. BMJ Open, 11(12), e052902. https://doi.org/10.1136/bmjopen-2021-052902

Kovoor, J. G., Stretton, B., Gupta, A. K., Beath, A., Jacob, M. O., Kefalianos, J. M., Carmichael, G. J., Zaka, A., O’Callaghan, G., Satheakeerthy, S., Booth, A., Delloso, T., Hugh, T. J., Chan, W. O., Maddern, G. J., Balan-Vnuk, E., Cusack, M., Gilbert, T., Maddison, J., … Zeitz, K. (2025). The Adelaide Score: prospective implementation of an artificial intelligence system to improve hospital and cost efficiency. ANZ Journal of Surgery, 95(3), 342–349. https://doi.org/10.1111/ans.19383

Lowe v. Cerner Corp., 38 F.4th 701 (4th Cir. 2022). (United States Court of Appeals for the Fourth Circuit decision on EHR software vendor liability.)

Mark, P., Papalia, I., Lai, J. K., & Pascoe, D. M. (2024). Clinical application of convolutional neural network lung nodule detection software: An Australian quaternary hospital experience. Journal of Medical Imaging and Radiation Oncology, 68(6), 659–666. https://doi.org/10.1111/1754-9485.13734

McLennan, R. (2025, March 7). ‘It encouraged our daughter to kill herself’: Calls for regulation after AI mental health chatbot’s role in child’s death revealed. ABC News (Australia). Retrieved from https://www.abc.net.au/news/2025-03-07/mental-health-chatbot-ai-child-suicide/100873197

Miller, I. J., Stapelberg, M., Hudson, J., Coxon, P., Milani, N., Rosic, N., Furness, J., Walsh, J., & Climstein, M. (2025). Market-approved convolutional neural network tasked with classifying skin lesions under suspicion of melanoma: performance across primary care clinics within Australia. PeerJ, 13, e19876. https://doi.org/10.7717/peerj.19876

Minssen, T., Vayena, E., & Cohen, I. G. (2023). The challenges for regulating medical use of ChatGPT and other large language models. JAMA, 330(4), 315–316. DOI: 10.1001/jama.2023.9651

Nguyen, T. M., Caponio, V. C. A., Rajappa, R., Lawrence, D., Tawadros, Y., Jones, B., Crow, K., & Cirillo, N. (2026). Face Validation of an Artificial Intelligence Driven Tool for Clinical Triaging in Australian Public Oral Healthcare: A Pilot Study. Value in Health, 29(4), 589–594. https://doi.org/10.1016/j.jval.2025.12.001

Nolan, P., & Matulionyte, R. (2023). Artificial intelligence in medicine: Issues when determining negligence. Journal of Law and Medicine, 30(3), 593–615.

Parasuraman, R., & Manzey, D. H. (2010). Complacency and bias in human use of automation: An attentional integration. Human Factors, 52(3), 381–410. https://doi.org/10.1177/0018720810376055

Pham, C. T., Visvanathan, R., Strong, M., Wilson, E. C. F., Lange, K., Dollard, J., Ranasinghe, D., Hill, K., Wilson, A., & Karnon, J. (2023). Cost-Effectiveness and Value of Information Analysis of an Ambient Intelligent Geriatric Management (AmbIGeM) System Compared to Usual Care to Prevent Falls in Older People in Hospitals. Applied Health Economics and Health Policy, 21(2), 315–325. https://doi.org/10.1007/s40258-022-00773-6

Pilowsky, J. K., Choi, J.-W., Nguyen, N., Williams, L., & Jones, S. L. (2026). Pressure injury surveillance in the intensive care unit: Development, validation, and clinical application of a natural language processing algorithm. Australian Critical Care : Official Journal of the Confederation of Australian Critical Care Nurses, 39(1), 101487. https://doi.org/10.1016/j.aucc.2025.101487

Price II, W. N., Gerke, S., & Cohen, I. G. (2019). Potential liability for physicians using artificial intelligence. JAMA, 322(18), 1765–1766. DOI: 10.1001/jama.2019.15064

Rao, V., Walia, N., Parkash, N., Wanigaratne, T., Henshaw, S., Chen, G., Lo, S. W., Be, K. H., Robertson, M., & Zorron Cheng Tao Pu, L. (2026). Availability and use of computer-aided detection during colonoscopy: A real-world observational study at an Australian tertiary center. World Journal of Gastroenterology, 32(4), 112698. https://doi.org/10.3748/wjg.v32.i4.112698

Rogers v. Whitaker (1992) 175 CLR 479 (High Court of Australia).

Ross, C., & Swetlitz, I. (2018, July 27). IBM’s Watson recommended “unsafe and incorrect” cancer treatments during demo, documents show. STAT News. Retrieved from https://www.statnews.com/2018/07/25/ibm-watson-recommended-unsafe-incorrect-treatments/

Satheakeerthy, S., Beecher, M., Booth, A. E., Stretton, B., Kovoor, J., Gupta, A., Evans, S., Howson, S., Logan, J., Qian, C., Lin, Y.-H., Gao, C., Onn Chan, W., Sorich, M. J., Brown, M. P., Jeffree, R. L., & Bacchi, S. (2026). Automating cancer registries: Pearls and pitfalls. Health Information Management : Journal of the Health Information Management Association of Australia, 55(1), 193–202. https://doi.org/10.1177/18333583251377892

Scheetz, J., Koca, D., McGuinness, M., Holloway, E., Tan, Z., Zhu, Z., O’Day, R., Sandhu, S., MacIsaac, R. J., Gilfillan, C., Turner, A., Keel, S., & He, M. (2021). Real-world  artificial intelligence-based opportunistic screening for diabetic retinopathy in endocrinology and indigenous healthcare settings in Australia. Scientific Reports, 11(1). https://doi.org/10.1038/s41598-021-94178-5

Sharp, G., Dwyer, B., Randhawa, A., McGrath, I., & Hu, H. (2025). The Effectiveness of a Chatbot Single-Session Intervention for People on Waitlists for Eating Disorder Treatment: Randomized Controlled Trial. Journal of Medical Internet Research, 27, e70874. https://doi.org/10.2196/70874

Song, J., Zolnoori, M., Scharp, D., Vergez, S., McDonald, M. V., Sridharan, S., … Topaz, M. (2022). Do nurses document all discussions of patient problems and nursing interventions in the electronic health record? A pilot study in home healthcare. JAMIA Open, 5(2), ooac034. https://doi.org/10.1093/jamiaopen/ooac034

Shah SV. Accuracy, consistency, and hallucination of large language models when analysing unstructured clinical notes in electronic medical records. JAMA Netw Open. 2024;7(8):e2425953. doi:10.1001/jamanetworkopen.2024.25953

State of New South Wales v. Lepore (2003) 212 CLR 511 (High Court of Australia).

Vandelanotte, C., Trost, S., Hodgetts, D., Imam, T., Rashid, M. D. M., To, Q. G., & Maher, C. (2025). MoveMentor-examining the effectiveness of a machine learning and app-based digital assistant to increase physical activity in adults: protocol for a randomised controlled trial. Trials, 26(1), 233. https://doi.org/10.1186/s13063-025-08926-3

Wallace v Kam (2013) 250 CLR 375 (High Court of Australia).

Wang, Y., & Magrabi, F. (2025). Assessing the transferability of BERT to patient safety: classifying multiple types of incident reports. BMJ Health & Care Informatics, 32(1). https://doi.org/10.1136/bmjhci-2024-101146

Zia, A., Fletcher, C., Bigwood, S., Ratnakanthan, P., Seah, J., Lee, R., Kavnoudias, H., & Law, M. (2022). Retrospective analysis and prospective validation of an AI-based software for intracranial haemorrhage detection at a high-volume trauma centre. Scientific Reports, 12(1). https://doi.org/10.1038/s41598-022-24504-y

---

APPENDICES

APPENDIX 1: SEARCH STRATEGY

Comprehensive search strategy planner
Search question: What AI applications are currently adopted across clinical, administrative, and patient-facing functions in Australian healthcare context?
What are the main concepts that make up my literature search question? AI Adoption  Healthcare (domains) Australia
Concept 1	OR combines terms for the same concept together
AI	("Artificial Intelligence"[MeSH]  OR "artificial intelligence"[tiab]  OR AI[tiab]  OR "machine learning"[tiab]  OR "deep learning"[tiab])
AND combines different concepts together
Concept 2	OR combines terms for the same concept together
Adoption	(adoption[tiab]  OR implementation[tiab]  OR implement*[tiab]  OR deploy*[tiab]  OR use[tiab]  OR utilisation[tiab]  OR integration[tiab])
AND combines different concepts together
Concept 3	OR combines terms for the same concept together
Healthcare	("Health Services"[MeSH]  OR "Hospitals"[MeSH]  OR healthcare[tiab]  OR hospital*[tiab]  OR clinic*[tiab]  OR "health system*"[tiab])
AND combines different concepts together
Concept 4	OR combines terms for the same concept together
Domains	("clinical decision support"[tiab]  OR diagnosis[tiab]  OR diagnostic[tiab]  OR radiology[tiab]  OR pathology[tiab]  OR "clinical workflow"[tiab]) (administrative[tiab]  OR administration[tiab]  OR workflow[tiab]  OR scheduling[tiab]  OR documentation[tiab]  OR "healthcare operations"[tiab]) ("patient-facing"[tiab]  OR patient[tiab]  OR chatbot*[tiab]  OR triage[tiab]  OR telehealth[tiab]  OR "virtual care"[tiab]  OR "remote monitoring"[tiab])
AND combines different concepts together
Concept 5	OR combines terms for the same concept together
Australia	(Australia*[tiab] OR Australian[tiab])
What inclusion or exclusion criteria will I apply to my search?  Inclusion criteria: Language: English Publication date: 2021-2026 Full text available Exclusion criteria: Publication type: non-peer-reviewed articles, editorials, proceedings, reviews, letters, or opinion pieces Availability: inability to locate or access the full-text article
What database(s) will I use to find evidence/information for my literature search?  PubMed Scopus
Will I include grey literature in my search? If so, how will I search for it? Grey literature excluded.

APPENDIX 2: PRISMA FLOWCHART