Artificial intelligence (AI) is changing how clinicians and healthcare teams gather data, make assessments, document care, and convey their findings. Clinical AI tools can draft notes, summarize records, suggest possible diagnoses, prioritize risks, and recommend treatments. An ethical assessment of AI tools from an Australian healthcare perspective has asserted that unsettled AI technologies used in patient care must protect safety, remain transparent and accountable, and preserve professional responsibilities [1]. The World Health Organization guidance on large multi-modal models similarly warns that general-purpose AI in health can propagate false content, amplify bias, promote automation bias, compromise privacy, and be poorly governed in clinical decision-making [2].

The gap between the technology available to clinicians and clinicians’ preparedness to use it has widened. In a survey of 39 primary care clinicians in metropolitan Australia, average AI familiarity was just below 2 on a 5-point scale; 64% intended to use AI in their work to some extent, while concerns centered on privacy, AI-induced errors, hallucinations, and integration with existing workflows [3]. This suggests willingness to adopt AI in the absence of mature cognitive and ethical frameworks for safe use. Large language models have ingested extensive medical knowledge but retain gaps in safety, factuality, and clinical value [4]. GPT-4 has shown strong performance on clinical reasoning tasks compared with practicing physicians [5]. However, in a randomized clinical trial, access to a large language model that performed well in isolation did not improve clinicians’ diagnostic reasoning beyond conventional resources [6]. A 2026 Science study likewise reported that a large language model exceeded clinicians’ baseline performance on selected reasoning tasks and showed the value of clinical adjudication [7]. These findings create an empirical tension: clinicians may over-rely on AI and weaken independent reasoning, or fail to use AI when it could improve care.

“Brain fog” is used in biomedical literature to describe subjective cognitive symptoms such as reduced attention, memory, language fluency, and mental clarity, although the term remains variably defined [8]. It is also prominent in post-coronavirus disease 2019 (COVID-19) literature as a patient-reported cluster of persistent cognitive symptoms [9]. In this article, “AI brain fog” is not a clinical diagnosis. It is an operational concept for reduced independent reasoning that may arise when clinicians repeatedly offload clinical synthesis, recall, and judgment to AI. “AI brain fry” has a more direct source: Bedard et al used the term for mental fatigue associated with extensive use or monitoring of AI tools in a workplace study of 1,488 United States workers [10]. Cognitive load theory provides a framework for this concern because reasoning may deteriorate when working-memory demands exceed human capacity [11].

In the literature reviewed, AI brain fog and AI brain fry are not established as a paired classification and are not routinely used interchangeably. AI brain fog is an author-proposed synthesis of literature on brain fog, cognitive offloading, dependency, automation, and deskilling; AI brain fry adapts a workplace term to healthcare workload and monitoring. Their distinction concerns the dominant direction of cognitive demand: too little active reasoning in fog versus too much information and oversight in fry. They may coexist, and both are conceptual educational tools intended to support observation, discussion, and research rather than diagnoses or validated syndromes. Differentiating them matters because the immediate response differs: fog calls for re-engagement with independent reasoning, whereas fry first calls for reduction and prioritization of cognitive load.

This article was designed as a narrative conceptual review rather than a systematic review or meta-analysis. The aim was not to provide pooled effect estimates, but to develop a transparent conceptual synthesis of emerging cognitive risks associated with AI in healthcare. The review focused on two non-diagnostic constructs: AI brain fog, referring to diminished independent clinical reasoning associated with excessive cognitive offloading to AI, and AI brain fry, referring to cognitive overload and fatigue arising from excessive AI-generated outputs, alerts, verification work, and monitoring demands.

A structured narrative search strategy was used to improve transparency and reproducibility. Literature was identified through PubMed/MEDLINE, Embase, Google Scholar, and relevant publisher websites. The searches combined terms related to AI and healthcare with terms related to cognition, clinical reasoning, education, and workload. Search terms included “artificial intelligence,” “healthcare,” “clinical reasoning,” “cognitive offloading,” “automation bias,” “alert fatigue,” “healthcare education,” “medical education,” “nursing education,” “large language models,” “AI dependency,” “clinical decision support,” “cognitive load,” “ambient AI scribes,” “diagnostic reasoning,” and “human-AI interaction.” Searches were supplemented by backward citation searching (backward snowballing) of relevant articles and by foundational cognitive-science literature when it directly informed the conceptual framework [12].

The review prioritized peer-reviewed empirical studies, systematic reviews, narrative reviews, conceptual analyses, and major policy or governance documents relevant to AI in healthcare. Greater emphasis was placed on literature published between 2021 and 2026, reflecting the rapid development of generative AI and large language models in clinical and educational settings. Earlier studies were included when they provided foundational concepts, particularly cognitive offloading, memory, automation bias, cognitive load theory, out-of-the-loop performance, and clinical reasoning. Non-healthcare AI literature was considered only when it provided a mechanism plausibly transferable to healthcare, such as workplace AI-related fatigue or human-AI workload effects.

Studies and documents were eligible for inclusion if they addressed at least one of the following domains: AI-assisted clinical reasoning, cognitive offloading, automation bias, alert fatigue, healthcare education, clinician or healthcare-worker workload, AI-generated documentation, clinical decision support, human-AI interaction, or governance of AI in healthcare. Articles were excluded if they did not address healthcare or a directly transferable cognitive mechanism, did not examine cognitive or professional implications of AI use, or focused solely on technical model performance without implications for reasoning, education, workload, or governance.

The synthesis proceeded in three stages. First, relevant literature was grouped into thematic domains: human clinical cognition; cognitive offloading and dependency; automation bias and out-of-the-loop performance; cognitive load and alert fatigue; healthcare education; ambient AI documentation; clinical decision support; and healthcare AI governance. Second, these domains were mapped to the proposed constructs of AI brain fog and AI brain fry, including areas of overlap. Third, recurring risks were translated into a cognitive resilience framework for clinicians, educators, and health services, including human-first reasoning, AI as a second opinion, structured verification, cognitive-load control, AI-free competency assessment, and reflective accountability.

Because this was a narrative conceptual review, formal risk-of-bias assessment and meta-analysis were not undertaken. The review does not claim to exhaustively identify all studies, establish prevalence or causality, or validate AI brain fog or AI brain fry as diagnostic entities. Evidence drawn from non-clinical settings is used to generate testable mechanisms, not to imply that effects have already been demonstrated in healthcare. The proposed terms are conceptual and educational tools rather than formal clinical diagnoses.

Clinical reasoning involves more than retrieval of facts. It includes problem representation, pattern recognition, management of uncertainty, formulation and testing of hypotheses, and reflection and adaptation. Dual-process models remain relevant because clinicians move between rapid pattern recognition and slower analytical reasoning. Cognitive load can affect diagnostic performance, particularly in intensive care and other settings where data are incomplete and time is constrained [13]. AI alters this cognitive context by retrieving information, condensing records, and suggesting diagnostic possibilities, while also influencing what clinicians notice, remember, and question. Previous work on medical synthesis has described how AI may help integrate complex information while raising questions about standardization, interpretability, and professional judgment [14].

Automation can also change the clinician’s role from active synthesis to passive monitoring. In other safety-critical domains, this “out-of-the-loop” condition has been associated with reduced situation awareness and slower recovery when automation fails [15]. Direct transfer to healthcare requires prospective testing, but the mechanism is relevant where a clinician must rapidly resume unaided reasoning after an erroneous output. The question is therefore not whether AI should assist cognition; its role will continue to expand. The more important question is whether AI use strengthens or weakens the reasoning habits needed when cases are ambiguous, data are incomplete, or AI output is wrong.

AI brain fog is defined here as a decline or underuse of independent clinical reasoning associated with habitual reliance on AI in situations where active synthesis would ordinarily be required. Observable manifestations may include an incomplete problem representation, inability to recall or reconstruct the reasoning path, a narrow differential diagnosis, inability to justify a clinical plan without reopening the AI output, and failure to identify disconfirming evidence. The term should be used cautiously because brain fog is an established but variably defined clinical descriptor, including in COVID-19 literature [8, 9]. AI brain fog does not describe a neurological syndrome; it is an educational and research concept.

Cognitive offloading underpins this concern. The “Google effect” describes how people may remember where information can be found while retaining less of the information itself [16]. Risko et al defined cognitive offloading as the use of an external system to reduce demands on internal cognitive resources [17]. Offloading can be efficient and adaptive; the risk arises when the external tool replaces the rehearsal, retrieval, or synthesis required to maintain competence, rather than removing low-value work while preserving judgment.

Clinical care may be especially vulnerable because clinicians routinely work with incomplete and contradictory information. A learner who asks AI to generate a differential diagnosis before framing the problem offloads the generative step. An emergency or primary care clinician who accepts an AI summary may miss a contradiction in the record. A radiology reader who sees an AI label before an independent search may narrow attention prematurely. A nurse, nurse practitioner, or physician assistant who follows a prompted assessment may document individual elements without integrating the patient’s trajectory. Conversely, a team can appropriately offload low-value retrieval while retaining responsibility for identifying the most dangerous diagnosis not to miss. The concern is therefore not offloading itself, but the transfer of cognitive authority for tasks that require professional synthesis [15, 17].

A survey of 319 knowledge workers found that greater confidence in generative AI was associated with less self-reported critical thinking, while the locus of effort shifted from information generation toward verification, integration, and task stewardship [18]. This non-clinical study does not establish AI brain fog in healthcare, but it supports a plausible mechanism for reduced engagement. Healthcare education is particularly relevant. Lim et al argued that AI dependency may diminish clinical intuition, empathy, and diagnostic confidence [19]. Learners require deliberate practice before their internal checks are sufficiently developed to use external supports safely; AI should therefore be sequenced to follow, rather than replace, initial problem representation and reasoning.

AI brain fry is defined here as cognitive exhaustion caused by the volume, fragmentation, and monitoring demands of AI-assisted work. Bedard et al used the term for fatigue arising when people use or monitor AI tools beyond their cognitive capacity [10]. In healthcare, AI may reduce drafting or retrieval while increasing the need to review draft notes, reconcile multiple outputs, and verify recommendations. Cognitive load theory predicts that limited working-memory capacity can be consumed by verification, integration, and revision tasks [11]. Long, poorly prioritized summaries, low-value alerts, and disjointed or contradictory recommendations may therefore produce hasty decisions, missed signals, alert dismissal, and reliance on default options.

Clinical decision-support systems already demonstrate related workload effects. A systematic review found that alerts which disrupt workflow may be overridden, silenced, or ignored [20]. Combined with automation bias, fatigue can create risk in both directions: excessive trust in a recommendation and disengagement from useful warnings [21]. In a randomized clinical-vignette study involving hospitalist physicians, nurse practitioners, and physician assistants, standard AI modestly improved diagnostic accuracy, but systematically biased AI substantially reduced it; commonly used image-based explanations did not eliminate the harm [22]. This finding supports an interprofessional, rather than physician-only, account of AI-related cognitive risk.

Fatigued clinicians may accept a recommendation after only superficial confirmation—checking that the output is present, complete, or superficially plausible—rather than assessing its validity. In this review, structured verification means checking patient identity and context, evidentiary source, internal consistency, medication and dose accuracy, contradictory data, uncertainty, and fit with the patient’s values and circumstances. Completing a review checkbox is not equivalent to validating a recommendation.

Ambient AI scribes illustrate the benefit-burden trade-off. Ma et al reported reductions of 6.89 min in average daily documentation time, 5.17 min in after-hours documentation, and 19.95 min in total daily electronic health-record time [23]. Generated notes still require review, especially where medication changes, symptom chronology, or patient preferences are concerned. Much of the current ambient-scribe evidence is physician-centered. A nursing commentary has highlighted risks of omission, hallucination, and bias when nurses are excluded from design and oversight [24]. The literature reviewed did not directly test whether early-career nurses who depend on clinical decision-support prompts have difficulty narrating an unprompted assessment. This is a plausible but currently unverified training concern and should be studied rather than assumed.

The practical distinction is based on the dominant observable problem, not on a formal diagnosis. AI brain fog is most likely when the clinician appears cognitively under-engaged: the output is accepted with little reconstruction, the plan cannot be explained without returning to AI, or important alternatives were never generated. AI brain fry is most likely when the clinician appears overloaded: output volume, interruptions, or monitoring demands lead to fatigue, skimming, delayed review, or indiscriminate dismissal. These patterns are consistent with the literature on brain fog, AI-related fatigue, cognitive load, cognitive offloading, out-of-the-loop performance, alert fatigue, and automation bias [8–11, 15, 17, 20–22]. A proposed bedside or supervisory probe is to ask the clinician to restate the problem, working diagnosis, dangerous alternative, and reasons for accepting or rejecting the AI output without reopening it. Difficulty doing so suggests fog; restoration of performance after reducing output and interruptions suggests fry.

The constructs can occur together. Drawing on cognitive-load, alert-fatigue, automation-bias, and cognitive-offloading mechanisms, a long, unprioritized AI-generated summary may first produce brain fry; the resulting fatigue may then prompt shallow acceptance of the recommendation, producing a fog-like failure to reason independently [11, 17, 20, 21]. Fog may reflect a gradually learned habit, whereas fry may present as an acute state, but neither distinction is absolute. Co-occurrence should therefore be managed in sequence: reduce the immediate cognitive burden, then re-establish independent reasoning and high-risk verification.

Based on these mechanisms, the authors propose that when fog is suspected, AI consultation should be paused while the clinician independently restates the problem representation, differential diagnosis, disconfirming evidence, and patient-specific risk. When fry is suspected, outputs should be triaged or limited, low-value alerts suppressed, review responsibility assigned, and protected review time provided. These responses are consistent with evidence on cognitive load, cognitive offloading, out-of-the-loop performance, alert fatigue, and automation bias, although they require prospective clinical validation [11, 15, 17, 20–22]. When both are present, cognitive load should be reduced first, followed by an AI-free restatement and structured verification of medications, diagnoses, and follow-up. If safe independent review cannot be completed, the case should be escalated rather than processed through further AI output. The distinguishing features, areas of overlap, and corresponding remediation and prevention strategies for AI brain fog and AI brain fry are summarized in Table 1 [11, 15, 17, 20–22].

Click to view

Table 1. Conceptual Distinction Between AI Brain Fog and AI Brain Fry

AI-related cognitive risks are unlikely to be distributed equally across the healthcare system, and they are not limited to physicians. Health professions students may be vulnerable because diagnostic, assessment, and illness scripts are still developing. Early-career clinicians who diagnose or prescribe—including junior doctors, nurse practitioners, and physician assistants—may be vulnerable to premature closure under time pressure. Experienced clinicians may face gradual deskilling in rare or complex tasks if AI replaces repeated practice. Nurses, allied health professionals, pharmacists, diagnostic professionals, and administrative staff may experience brain fry through alerts, documentation, standardized recommendations, or automation at scale. Evidence for the exact prevalence and gradient by profession or seniority remains limited; Table 2 is therefore hypothesis-generating. Its mechanisms are supported by research on out-of-the-loop performance, alert fatigue, automation bias across physicians, nurse practitioners and physician assistants, and AI in nursing education [15, 20, 22, 25].

Click to view

Table 2. AI-Related Cognitive Risks Across Healthcare Groups and Experience Levels

Healthcare education literature supports this cautious, interprofessional approach. A systematic review of large language models in medical education identified opportunities for teaching, assessment, and feedback alongside concerns about accuracy, bias, integrity, and learner dependence [26]. A systematic review of 15 studies involving 1,464 nursing students and professionals found potential educational benefits but also emphasized the absence of longitudinal evidence and robust measurement [25]. Virtual-patient tools can support clinical reasoning when they require active engagement rather than passive retrieval of answers [27]. Across professions, a practical educational rule is that AI should create more reasoning, not less: learners should formulate an assessment or differential before viewing AI output and should be able to explain why they accepted, modified, or rejected it.

AI brain fog and AI brain fry should not be used to dismiss AI. The stronger argument is that benefits require safeguards. Several perioperative studies illustrate why clinicians may value AI-assisted modelling [28]. A causal machine-learning study of 3,017 adults undergoing non-cardiac surgery found heterogeneous renal effects associated with phenylephrine, with age, baseline renal function, and American Society of Anesthesiologists (ASA) physical status acting as important effect modifiers [29]. A study of 67,134 surgical procedures identified nonlinear fentanyl–propofol dose–response patterns associated with postoperative intensive care unit (ICU) admission risk [30]. Tabular foundation modelling has also shown promise in perioperative classification, although performance varied by sample size, outcome incidence, and calibration [31]. Related work has proposed TabPFN for real-time precision modelling in surgical health economics [32]. These studies support the claim that AI can reveal patterns that may be difficult to detect unaided.

The ethical risk emerges when accuracy creates dependence—the accuracy-dependency paradox. As AI becomes more useful, clinicians may grant it more cognitive authority; as authority shifts, independent reasoning may be practiced less often. The pattern may differ by specialty. In radiology, an accurate triage label can speed review but may narrow the independent visual search when it is wrong. In emergency and hospital medicine, an AI-generated differential may improve breadth but also anchor the team on an early frame. In perioperative and critical care, a risk model may reveal nonlinear associations while weakening attention to the physiological rationale if the score becomes the endpoint. In primary care, automated summaries may reduce inbox workload yet conceal contradictions or medication changes. In nursing and pharmacy, escalation or verification tools may speed recognition but also encourage template-driven assessment or superficial approval. These examples are mechanism-based illustrations, not evidence that every specialty will experience the same effect [15, 22].

There is no single correct organizational response. Depending on the task, evidence, and benefit–risk profile, a health service may retain, redesign, restrict, pause, or withdraw an AI tool. Where the tool is retained, a proportionate goal is to preserve clinical agency rather than reflexively weaken useful capability. Interfaces can ask clinicians to record a working diagnosis before viewing advice, identify disconfirming evidence, state uncertainty, and name the patient-specific factor that could make a recommendation unsafe. The appropriate balance should be evaluated empirically for each role and workflow.

Cognitive resilience is the ability of clinicians and teams to use AI while preserving independent reasoning, attention, accountability, and patient-centered judgment. It requires individual habits, interprofessional educational design, and organizational governance. Table 3 [17, 19–22, 25, 26, 33] presents a practical framework for embedding cognitive resilience in AI-assisted healthcare.

Click to view

Table 3. Cognitive Resilience Framework for AI-Assisted Healthcare

This framework reframes AI literacy. Healthcare professionals need to know not only how AI works, but how it changes their own thinking. AI literacy should include prompt formulation, bias recognition, uncertainty appraisal, structured verification, awareness of cognitive offloading and overload, and the ability to stop using AI when a task requires direct human attention. These competencies should be tailored to scope of practice and experience rather than delivered as a single physician-centered curriculum.

System design must complement individual vigilance. Health services should determine whether AI reduces workload or redistributes it into less visible verification and monitoring. Potential metrics include time spent reviewing generated notes, alert volume, output length, changes in diagnoses or actions after AI consultation, override and dismissal patterns, near misses, corrected documentation errors, clinician confidence without AI support, and disparities across professions, settings, and patient groups. These measures can distinguish genuine assistance from a shift of cognitive work to verification.

Operational responsibility should be explicit, consistent with governance guidance emphasising patient safety, transparency, accountability, and preservation of professional responsibility [1, 2]. Each AI tool should have an executive sponsor, a named clinical service owner, and oversight by a multidisciplinary AI governance group. Membership should include physicians, nurses, nurse practitioners, physician assistants, pharmacists, allied health professionals, informaticians, educators, human-factors and quality-safety specialists, privacy and legal representatives, and patient or consumer representatives where appropriate. The clinical owner should define safe use and escalation; educators should define competency requirements; information technology and vendors should provide telemetry and change control; and quality-safety teams should conduct independent audits. This distribution of responsibility also aligns with implementation frameworks that emphasize stakeholder involvement and organizational context [33].

Compliance should be monitored through workflow data and sampled case review, not solely through self-attestation. This is important because alert fatigue [20], automation bias and erroneous AI recommendations [21, 22], and errors or omissions in AI-generated documentation [23, 24] may not be captured by simple attestations. Before deployment, organizations should establish baseline workload, error, and performance measures, while staged pilots should use predefined success criteria and stop rules [2, 33]. Early monitoring may be monthly and can become less frequent only after stable performance is demonstrated. Audits should examine review time, alert burden, unreviewed outputs, changes in management, accepted incorrect recommendations, corrected scribe errors, near misses, and the ability to complete selected tasks without AI [20–24]. Monitoring should support learning rather than punitive surveillance, with rapid feedback to frontline users [2, 33].

Implementation barriers include acquisition and maintenance costs, workflow disruption, interoperability, training time, unequal access to support, vendor lock-in, and the burden of ongoing evaluation; concerns about workflow integration have also been reported among frontline primary care clinicians [3]. The Nonadoption, Abandonment, Scale-up, Spread, and Sustainability (NASSS) framework emphasizes that sustained adoption depends on interacting complexity across the condition, technology, value proposition, adopters, organization, wider system, and change over time [33]. Practical mitigation includes high-risk-first pilots, co-design with each affected profession, protected training time, fallback workflows for downtime or uncertainty, procurement rights to audit logs and performance data, and a recurring budget for evaluation [2, 33]. A tool should be paused, reconfigured, or withdrawn when benefit cannot be demonstrated or when cognitive burden, inequity, or safety risk worsens [1, 2, 33].

The proposed constructs require prospective validation rather than rhetorical acceptance. Studies should test whether they have discriminant validity, whether they predict clinically important outcomes, and whether targeted safeguards reduce risk. Table 4 converts the conceptual claims into measurable hypotheses suitable for educational, simulated, and real-world studies.

Click to view

Table 4. Proposed Measurable Hypotheses for Prospective Validation

Controlled educational studies can test H1 by allowing one group to access AI before forming a diagnosis or assessment and requiring another group to form an independent problem representation first. Outcomes should extend beyond immediate accuracy to explanation quality, delayed retention, transfer to new cases, and performance after AI is removed. Longitudinal cohorts should include medical, nursing, nurse practitioner, physician assistant, pharmacy, and allied health trainees, while measuring baseline knowledge, sleep, burnout, workload, and AI exposure to avoid attributing all cognitive change to AI.

AI brain-fry research should quantify the number, type, length, timing, and priority of outputs received during real work. Validated workload measures such as the NASA Task Load Index can be combined with system logs, review omissions, decision time, eye-tracking where feasible, and qualitative accounts [34]. Ambient-scribe studies should include nurses and other professions and test whether repeated use changes the quality of unprompted assessment and handover. Clinical decision-support trials should compare displays of uncertainty, opposing evidence, and alternative diagnoses, and should measure both appropriate reliance when AI is correct and resistance when it is wrong.

Construct-validation studies should determine whether AI brain fog and AI brain fry are distinguishable from ordinary fatigue, burnout, sleep deprivation, low knowledge, and general digital overload. A credible measure of fog should predict loss of unaided reasoning or retention after controlling for workload; a credible measure of fry should predict acute review failure after controlling for baseline competence. Real-world studies should also test the proposed sequence for co-occurrence: reduce load first, then restore independent reasoning and structured verification. These studies would establish whether the terms add explanatory and practical value beyond existing constructs.

AI brain fog and AI brain fry are conceptual educational tools for naming different cognitive risks associated with AI-assisted healthcare. AI brain fog describes under-engagement and possible erosion of independent reasoning when AI supplies the first-pass assessment or synthesis. AI brain fry describes overload and fatigue produced by excessive outputs, alerts, monitoring, and verification. They are not formal diagnoses, do not have a clinical ontology, and may occur together. Their practical value lies in linking observable problems to different immediate responses: re-engage reasoning for fog, reduce load for fry, and do both sequentially when they coexist.

The central ethical question is not whether AI is proper in healthcare in the abstract, but whether a particular system and workflow preserve the cognitive capabilities required for safe care. Highly accurate AI may influence documentation, prediction, education, and professional practice in subtle ways. When a clinician cannot explain a diagnosis without the AI output, or when a team is desensitized by repeated alerts, responsibility lies not only with the individual but also with the design, implementation, and governance of the system.

Human-first reasoning, AI as a second opinion, structured verification, AI-free competence assessment, role-sensitive education, and workload-aware system design should be evaluated as core components of AI governance. Implementation requires named owners, interprofessional oversight, measurable performance indicators, staged deployment, and willingness to reconfigure or withdraw a tool when harms exceed benefits. Healthcare AI should be judged not only by model performance and time saved, but by whether it helps professionals maintain clear, critical, patient-centered thinking when AI is absent, uncertain, or wrong.

1. Lim ECN, Lim CED. Responsible use of AI in healthcare: An Australian perspective on promise, perils, and professional duties. AI Ethics. 2026;6:10.
   doi
2. Ethics and governance of artificial intelligence for health: Guidance on large multi-modal models. World Health Organization. 2025. https://www.who.int/publications/i/item/9789240084759.
3. Lim ECN, Lim CED. Artificial intelligence use in primary care: Attitudes, concerns, and readiness among health professionals in metropolitan Australia. Journal of Computer and Communications. 2025;13:63-71.
   doi
4. Singhal K, Azizi S, Tu T, Mahdavi SS, Wei J, Chung HW, Scales N, et al. Large language models encode clinical knowledge. Nature. 2023;620(7972):172-180.
   doi pubmed
5. Cabral S, Restrepo D, Kanjee Z, Wilson P, Crowe B, Abdulnour RE, Rodman A. Clinical reasoning of a generative artificial intelligence model compared with physicians. JAMA Intern Med. 2024;184(5):581-583.
   doi pubmed
6. Goh E, Gallo R, Hom J, Strong E, Weng Y, Kerman H, Cool JA, et al. Large language model influence on diagnostic reasoning: a randomized clinical trial. JAMA Netw Open. 2024;7(10):e2440969.
   doi pubmed
7. Brodeur PG, Buckley TA, Kanjee Z, Goh E, Ling EB, Jain P, Cabral S, et al. Performance of a large language model on the reasoning tasks of a physician. Science. 2026;392(6797):524-527.
   doi pubmed
8. Denno P, Hampshire A. Defining brain fog: A transdiagnostic narrative review. European Psychiatry. 2025;68(S1):S602-S603.
   doi
9. Graham EL, Clark JR, Orban ZS, Lim PH, Szymanski AL, Taylor C, DiBiase RM, et al. Persistent neurologic symptoms and cognitive dysfunction in non-hospitalized Covid-19 "long haulers". Ann Clin Transl Neurol. 2021;8(5):1073-1085.
   doi pubmed
10. Bedard J, Kropp M, Hsu M, Karaman OT, Hawes J, Rosen Kellerman G. When using AI leads to “brain fry.” Harvard Business Review. 2026. https://hbr.org/2026/03/when-using-ai-leads-to-brain-fry.
11. Wang P, Liu H, Zou L, Paas F. Overloaded minds and machines: A cognitive load framework for human-AI symbiosis. Artificial Intelligence Review. 2026;59:109.
    doi
12. Greenhalgh T, Peacock R. Effectiveness and efficiency of search methods in systematic reviews of complex evidence: audit of primary sources. BMJ. 2005;331(7524):1064-1065.
    doi pubmed
13. Harris E, Santhosh L. Dual process theory and cognitive load: how intensivists make diagnoses. Crit Care Clin. 2022;38(1):27-36.
    doi pubmed
14. Lim ECN, Cheng NCL, Lim CED. The art of medical synthesis: Where Chinese medical wisdom intersects with artificial intelligence. Journal of Traditional Chinese Medical Sciences. 2026;13(1):51-59.
    doi
15. Endsley MR, Kiris EO. The out-of-the-loop performance problem and level of control in automation. Human Factors. 1995;37(2):381-394.
    doi
16. Sparrow B, Liu J, Wegner DM. Google effects on memory: cognitive consequences of having information at our fingertips. Science. 2011;333(6043):776-778.
    doi pubmed
17. Risko EF, Gilbert SJ. Cognitive offloading. Trends Cogn Sci. 2016;20(9):676-688.
    doi pubmed
18. Lee HP, Sarkar A, Tankelevitch L, Drosos I, Rintel S, Banks R, Wilson N. The impact of generative AI on critical thinking: Self-reported reductions in cognitive effort and confidence effects from a survey of knowledge workers. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems. Association for Computing Machinery. 2025;Article 1121:1-22.
    doi
19. Lim ECN, Lim CED. Uncertain hands: Ethical implications of AI dependency in medical education. AI Ethics. 2026;6:219.
    doi
20. Olakotan OO, Mohd Yusof M. The appropriateness of clinical decision support systems alerts in supporting clinical workflows: a systematic review. Health Informatics J. 2021;27(2):14604582211007536.
    doi pubmed
21. Khera R, Simon MA, Ross JS. Automation bias and assistive AI: risk of harm from AI-driven clinical decision support. JAMA. 2023;330(23):2255-2257.
    doi pubmed
22. Jabbour S, Fouhey D, Shepard S, Valley TS, Kazerooni EA, Banovic N, Wiens J, et al. Measuring the impact of AI in the diagnosis of hospitalized patients: a randomized clinical vignette survey study. JAMA. 2023;330(23):2275-2284.
    doi pubmed
23. Ma SP, Liang AS, Shah SJ, Smith M, Jeong Y, Devon-Sand A, Crowell T, et al. Ambient artificial intelligence scribes: utilization and impact on documentation time. J Am Med Inform Assoc. 2025;32(2):381-385.
    doi pubmed
24. Topaz M. Invisible scribes: can nurses trust ambient AI for clinical documentation? J Contin Educ Nurs. 2025;56(9):358-359.
    doi pubmed
25. Ma J, Wen J, Qiu Y, Wang Y, Xiao Q, Liu T, Zhang D, et al. The role of artificial intelligence in shaping nursing education: a comprehensive systematic review. Nurse Educ Pract. 2025;84:104345.
    doi pubmed
26. Lucas HC, Upperman JS, Robinson JR. A systematic review of large language models and their implications in medical education. Med Educ. 2024;58(11):1276-1285.
    doi pubmed
27. Plackett R, Kassianos AP, Mylan S, Kambouri M, Raine R, Sheringham J. The effectiveness of using virtual patient educational tools to improve medical students' clinical reasoning skills: a systematic review. BMC Med Educ. 2022;22(1):365.
    doi pubmed
28. Lim ECN, Lim CED. Machine learning in perioperative medicine: a comparative review of predictive, causal, and foundation model approaches in surgical data science. AI in Clinical Medicine. 2026;2:e27.
    doi
29. Cui J, Lim ECN, Wu X, Lim CED. Phenylephrine’s heterogeneous renal effects in non-cardiac surgery: a causal machine learning study. Cureus Journal of Computer Science. 2025;2:es44389-025-00018-2.
    doi
30. Cui J, Li W, Lim ECN, Wu X, Lim CED. Stratified causal inference for intensive care unit risk prediction: informatics-based modeling of anesthetic drug combinations. JMIR Form Res. 2026;10:e80294.
    doi pubmed
31. Cui J, He Z, Lim ECN, Lim CED. Precision vs recall in surgical AI: Performance and limitations of tabular foundation model for perioperative classification and risk prediction. Digit Health. 2026;12:20552076261443585.
    doi pubmed
32. Lim ECN, Lim CED. Redefining surgical health economics: The potential of TabPFN for real-time precision modelling. Journal of Computer and Communications. 2025;13:30-40.
    doi
33. Greenhalgh T, Wherton J, Papoutsi C, Lynch J, Hughes G, A'Court C, Hinder S, et al. Beyond adoption: a new framework for theorizing and evaluating nonadoption, abandonment, and challenges to the scale-up, spread, and sustainability of health and care technologies. J Med Internet Res. 2017;19(11):e367.
    doi pubmed
34. Hart SG, Staveland LE. Development of NASA-TLX (Task Load Index): Results of empirical and theoretical research. In: Hancock PA, Meshkati N. (Eds.) Human mental workload. North-Holland. 1988. p. 139-183.
    doi