State of clinical AI in 2026

Review State of clinical AI in 2026 BMJ Digital Health & AI: first published as 10.1136/bmjdh-2026-000113 on 29 September 2026. Downloaded from https://bmjdigitalhealth.bmj.com on 25 October 2026 by guest. John Emmett Worth,1 Anastasia Perez ‍ ‍ ,2 David JH Wu,3 Peter Brodeur,4 Emily Tat,5 Arjun Manrai,6 David Wu,7 Chase Walton,8,9 Liam G McCoy,4,10,11 Priyank Jain,12 Rebecca Handler,2 Jason Hom,13 Laura Zwaan,14 Vishnu Ravi,2 Brian Han,2 Kevin Schulman,2 Kathleen Lacar,2 Kameron Black,2 Adrian D Haimovich,15 Adam Rodman,4 Ethan Goh ‍ ‍ ,16,17 Jonathan H Chen16,17,18 To cite: Worth JE, Perez A, ABSTRACT Wu DJH, et al. State of clinical Clinical artificial intelligence (AI) has advanced rapidly, WHAT IS ALREADY KNOWN ON THIS TOPIC Protected by copyright, including for uses related to text and data mining, AI training, and similar technologies. AI in 2026. BMJ Digit Health with frontier large language models now matching ⇒ Artificial intelligence (AI) systems have demon- 2026;2:e000113. doi:10.1136/ strated strong performance on medical licensing or exceeding physician performance on simulated bmjdh-2026-000113 examinations and controlled clinical reasoning diagnostic reasoning and clinical decision-­support tasks. Yet adoption has outpaced the evidence base: fewer tasks, leading to rapid adoption across healthcare. ► Additional supplemental than 5% of cleared U.S. Food and Drug Administration However, prior reviews have typically focused on material is published online only. (FDA) AI/machine learning (ML) devices have undergone narrow subtopics, such as individual specialties or To view, please visit the journal specific model architectures, without synthesis- peer-­reviewed evaluation and prospective randomised online (https://​doi.​org/​10.​1136/​ ing evidence across model capabilities, evaluation trials remain scarce. This narrative review synthesises bmjdh-​2026-​000113). evidence from ARISE’s State of Clinical AI Report 2026 methodologies, human-­ AI interaction and patient across six domains: model performance and human safety factors. JEW, AP, DJW and PB benchmarking, evaluation methodology and benchmark WHAT THIS STUDY ADDS contributed equally. validity, foundational methods including multimodal and multiagent systems, clinical workflow integration, ⇒ This review synthesises evidence across six do- AR, EG and JHC are joint senior patient-­facing applications and applied domain-­specific mains: model performance, benchmarks and evalu- authors. AI. We conducted a targeted literature search of PubMed, ation, foundational methods, AI in clinical workflows, medRxiv and arXiv supplemented by expert nomination, patient-­facing AI and applied AI. Key themes and pit- Received 9 May 2026 falls are explored for each domain. Across domains, Accepted 31 August 2026 including empirical studies of AI in clinical contexts. Across domains, frontier models demonstrate strong clinical a consistent theme emerges: model capabilities reasoning capabilities but exhibit persistent failures in have progressed faster than the research and im- uncertainty calibration, metacognition and robustness to plementation systems needed for safe deployment. distributional shift. Standard benchmarks have saturated, For example, benchmarks are saturating without prompting development of multidimensional evaluation capturing real-­world complexity, current human-­AI frameworks that assess safety, real-­world workflows and collaboration underperforms model-­only baselines agentic capabilities. New trends in foundational methods and real-­world prospective trials, while promising, include converting medical data into tokens and developing remain limited in scope and generalisability. multiagent, multimodal systems. In clinical workflows, HOW THIS STUDY MIGHT AFFECT RESEARCH, early prospective trials in clinical decision support and PRACTICE OR POLICY diagnostic imaging show promising results, though human-­AI collaboration remains suboptimal and risks of ⇒ More research is needed in real-­world clinical set- automation bias and clinician deskilling are emerging. tings. These findings support directing research In tandem, patient-­facing AI is progressing with more investment towards prospective evaluations em- personalised health assistance; with this comes the risk of bedded in authentic workflows, with greater empha- patient overtrust, raising the bar for guardrails. The most sis on safety, equity, metacognitive calibration and credible translational progress occurs on narrowly defined human-­ AI collaboration alongside accuracy. They tasks with clear endpoints, while broader clinical autonomy also support governance structures that require de- awaits advances in evaluation and human-­AI interaction. mographic reporting, bias assessment and ongoing post-­deployment monitoring of performance and © Author(s) (or their employer(s)) 2026. Re-­use clinical outcomes for all clinical AI tools. permitted under CC BY-­NC. No commercial re-­use. See rights and permissions. Published by INTRODUCTION controlled reasoning tasks, yet exhibit persis- BMJ Group. Clinical artificial intelligence (AI) has tent failures in uncertainty calibration (does For numbered affiliations see reached an inflection point. Over the past 2 expressed confidence match actual accu- end of article. years, AI systems have expanded across clin- racy), metacognition (eg,‘recognizing the Correspondence to ical reasoning, administrative workflows and limits of one’s own knowledge’) and robust- Dr Ethan Goh; patient-­facing interaction.1 Frontier models ness under distributional shift (performance ​ethangoh@​stanford.​edu now achieve superhuman performance on degradation when patient populations, care Worth JE, et al. BMJ Digit Health 2026;2:e000113. doi:10.1136/bmjdh-2026-000113 1 BMJ Digital Health & AI settings or data formats differ from those seen during models and agentic AI systems. Search terms included training).2 The gap between benchmark performance “large language models in medicine,” “AI in medicine,” and real-­world reliability remains one of the field’s “diagnostic reasoning,” “management reasoning,” “diag- BMJ Digital Health & AI: first published as 10.1136/bmjdh-2026-000113 on 29 September 2026. Downloaded from https://bmjdigitalhealth.bmj.com on 25 October 2026 by guest. defining tensions. nostic error,” “clinical benchmarks” and “patient-­facing The volume of literature has grown accordingly, with AI.” Eligible studies required high-­quality empirical eval- systematic reviews spanning nearly all medical special- uation of AI in clinical contexts with quantitative or qual- ties and the rate of large language model (LLM) studies itative outcomes (eg, diagnostic accuracy, bias, calibra- published increasing rapidly since ChatGPT’s release.3–7 tion, workflow and user performance). Purely technical This expansion in volume, however, has not translated papers without clinician-­ facing or patient-­ facing eval- into depth of evidence. Adoption of AI-­enabled medical uation, editorials and non-­clinical AI studies (eg, drug devices and consumer health is outpacing the generation discovery, biotech) were excluded. of prospective, real-­world evidence validating their safety In addition, sources and recommendations from and effectiveness in clinical workflows.8–12 Regulatory outside the scope of the targeted literature review were oversight has similarly struggled to keep pace, with more solicited from clinicians and AI researchers at academic Protected by copyright, including for uses related to text and data mining, AI training, and similar technologies. than 1200 FDA-­cleared AI/ML devices on the market and institutions via email and social media. Using a Delphi fewer than 5% having undergone rigorous peer-­reviewed process among the authors, these studies were reviewed evaluation.13 and included if they were felt to be of high clinical impact. This review is a narrative rendition of the AI Research As such, inclusion decisions were resolved through discus- and Science Evaluation (ARISE) network’s State of Clin- sion among author-­reviewers to reach consensus on final ical AI Report 2026 to provide clinicians, researchers and study selection. This established a final cohort of 82 policymakers with a critical appraisal across six domains: studies: 40 (49%) identified through database searches, (1) model performance, (2) benchmarks and evaluation, 31 (38%) nominated by the author team on the basis of (3) foundational methods, (4) AI in clinical workflows, expert knowledge and 11 (13%) submitted by external (5) patient-­facing AI and (6) applied AI (figure 1). clinicians and AI researchers. After the final study list was generated, the authors METHODOLOGY conducted an additional Delphi process to establish core The authors conducted a targeted literature search from themes. This consisted of a structured consensus process October 2025 to January 2026 in PubMed, medRxiv, across three meetings to identify, refine and organise and arXiv, supplemented by expert nominations. The the core themes emerging from the included literature. primary emphasis was on general-­purpose and genera- During these meetings, studies were reviewed collectively, tive AI systems, including LLMs, multimodal foundation candidate themes were proposed and iteratively revised Figure 1 Summary of the included literature by area. Areas include model performance, benchmarks and evaluation, foundational methods, AI in clinical workflows, patient-­facing AI and applied AI and demos. AI, artificial intelligence; EHR, electronic health record; SCT, script concordance testing. 2 Worth JE, et al. BMJ Digit Health 2026;2:e000113. doi:10.1136/bmjdh-2026-000113 BMJ Digital Health & AI and disagreements were resolved through group discus- in controlled settings. AgentMD, for example, enabled sion until consensus was reached on the final thematic models to autonomously select from a library of over framework. Our six domains were derived through the 2000 clinical calculators for a given patient, significantly BMJ Digital Health & AI: first published as 10.1136/bmjdh-2026-000113 on 29 September 2026. Downloaded from https://bmjdigitalhealth.bmj.com on 25 October 2026 by guest. Delphi process described above and reflect where the outperforming general-­purpose models in individual risk 2025–2026 empirical evidence was most concentrated; prediction accuracy on the 350-­question RiskQA bench- they are not intended as an exhaustive map of clinical AI. mark (87.7% vs 40.9%).16 Furthermore, the combination Several important areas, notably equity and algorithmic of deep learning models with continuous wearable data bias, regulation and governance, data privacy and secu- has been shown to augment clinical insight. In some rity and implementation science, warrant dedicated studies/instances, models detected surgical patient dete- review and fall outside the scope of this synthesis (online rioration up to 17 hours in advance and proactively iden- supplemental figure S1). tified metabolic precursors of insulin resistance.17 18 One study applying LLM-­ based frameworks to population-­ scale health data found that the predicted ‘biological RESULTS: THE SIX DOMAINS OF CLINICAL AI age’ outperformed established ageing biomarkers like Model performance telomere length or epigenetic clocks.19 Protected by copyright, including for uses related to text and data mining, AI training, and similar technologies. In 2025, frontier LLMs have shown capabilities in controlled clinical reasoning tasks, including diag- Persistent failure modes nostic reasoning and clinical decision support, that Despite strong benchmark performance, a recurring frequently match or exceed human physician perfor- limitation is the fragility of model performance in unfa- mance. Following these general-­purpose models, special- miliar scenarios. When multiple-­ choice tests replaced ised prediction models can now provide clinicians with the correct answers with ‘None of the other answers’, a insights previously unavailable through traditional statis- study found a drop in performance, ranging from 9% tical methods. These studies typically evaluate the perfor- to 38%, across all models.20 This indicates that near-­ mance of models in different scenarios as a ‘one-­time perfect accuracy on standard benchmarks may reflect measure’, which is distinct from benchmarks, which are statistical pattern-­matching or shortcut behaviour, where deliberately constructed to be reused for evaluation of models learn to associate structural or contextual cues emerging models. of multiple-­choice exam formatting with likely correct Diagnostic reasoning and management answers, rather than robust clinical logic. Moreover, Across studies, newer models have shown marked in tests designed to reflect the dynamism of clinician improvements in diagnostic reasoning, reasoning pres- reasoning when new information is introduced, models entation and disease management capabilities compared demonstrated systematic overconfidence and difficulty with earlier AI systems and physicians. For instance, one updating judgments under uncertainty.21 Even highly evaluation of frontier models on historically difficult optimised systems show significant room for improve- New England Journal of Medicine (NEJM) clinicopatholog- ment in high-­ stakes tasks like medication reasoning, ical conference (CPC) cases found OpenAI’s o3 model where accuracy varies between open-­book and closed-­ achieved 60% top-­1 and 84% top-­10 diagnostic accuracy book environments.15 Some note frontier models as and selected the correct next test 98% of the time.14 In displaying a ‘jagged’ intelligence: superhuman on some another study using CPCs, OpenAI’s o1 model achieved tasks yet failing at others humans find trivial, and prone 78% diagnostic accuracy and selected the next-­best test to collapse when confronted with excessive complexity or 87% of the time.2 Compared with human performance, a out-­of-­distribution data. Overall, for clinicians, the prac- study using 302 CPCs reported 24% top-­1 and 45% top-­10 tical implication is that trust should be calibrated to the accuracy among physicians with unrestricted electronic specific task at hand rather than to a general impression search.14 Newer models now extend beyond diagnosis to of competence, since capability can vary sharply between support disease management by integrating reasoning superficially similar problems. and retrieval capabilities. For example, Google’s Artic- ulate Medical Intelligence Explorer (AMIE) system was Benchmarks and evaluation found to be non-­inferior to 20 primary care physicians In contrast to model performance alone, benchmarks in management reasoning (treatment planning, longitu- are standardised instruments deliberately constructed dinal care), producing more precise plans and outper- for repeated use, enabling consistent comparison of forming physicians on medication-­reasoning questions.15 capability as new models emerge. Near-­perfect scores on medical licensing examinations do not capture the ‘mess- Scalable prediction and longitudinal clinical modelling iness’ of real-­world scenarios; therefore, clinical AI eval- Clinical AI is increasingly enabling scalable prediction uation has begun to shift toward more ecologically valid and longitudinal health monitoring, including assess- assessments of real-­world capability (figure 2, table 1). ments of risk (disease progression, clinical deteriora- tion, adverse event, etc) and population health. Agentic The saturation of knowledge-centric benchmarking systems, or AI models that autonomously plan and In response to benchmark saturation and the need for execute multistep tasks, have demonstrated clinical utility more trustworthy models, clinical AI evaluation has Worth JE, et al. BMJ Digit Health 2026;2:e000113. doi:10.1136/bmjdh-2026-000113 3 BMJ Digital Health & AI transitioned from multiple-­choice benchmarking towards approximating realistic clinical scenarios, these multi- more naturalistic assessments. The current landscape dimensional frameworks have advanced the evaluation overemphasises knowledge-­ based performance while paradigm beyond knowledge testing. Because these BMJ Digital Health & AI: first published as 10.1136/bmjdh-2026-000113 on 29 September 2026. Downloaded from https://bmjdigitalhealth.bmj.com on 25 October 2026 by guest. neglecting realistic deployment scenarios. A systematic frameworks are harder and more realistic than licensing-­ review (n=519 studies from 2022 to 2024) found that exam benchmarks, they avoid saturation and create a LLM evaluations overwhelmingly focused on medical meaningful spread between models of differing capabil- knowledge tasks, with 95% of studies relying on accuracy ities. as the primary evaluation metric.22 The most common tasks in these studies were answering medical licensing Gaps in safety, performance and the metacognitive frontier examination questions (45%) and making clinical diag- Despite the increased sophistication of these assessments, noses (19%), with only 5% of studies incorporating the transition to dynamic utility has exposed critical real patient data.22 The saturation of knowledge-­centric shortcomings. MedHELM’s findings highlight that the benchmarking and lack of task diversity have opened tasks most desired by practitioners to reduce burnout, the door for multidimensional evaluation frameworks such as administration and workflow support, remain the models’ weakest areas.23 Similarly, MedAgentBench used Protected by copyright, including for uses related to text and data mining, AI training, and similar technologies. that seek to represent a holistic view of clinical practice. As benchmarks improve, we find that determining the a virtual electronic health record (EHR) environment ‘right’ answer in medicine is itself difficult: this can often to test agentic capabilities across 300 patient-­ specific require integrating biology, ethics and the values and tasks, revealing that frontier models perform relatively priorities of individual patients; an option that appears poorly on multiturn, clinical tasks compared with their ‘wrong’ on paper may be optimal for a particular patient medical reasoning performance.26 Findings suggest that given unrecorded context; and ‘correct’ answers are not Fast Healthcare Interoperability Resources Application stable as evidence evolves. Benchmarks that assume a Programming Interface (API) navigation, interpretation single determinate answer therefore capture only part of of users’ prompt intent and long-­term planning remain clinical quality. immature for live clinical deployment. Another recurring limitation is the disconnect between benchmark accu- Broadening the horizon (multidimensional frameworks and models racy and clinical safety. The Numerous Options Harm of clinical realities) Assessment for Risk in Medicine (NOHARM) frame- Benchmarks introduced in 2025 expanded the complexity work revealed that in approximately 22% of primary care of evaluation practices to mirror the complexity of actual consultation cases, frontier models produced responses clinical workflows. A key example of this is Medical Holistic that could lead to clinical harm.27 Specifically, models Evaluation of Large Language Models (MedHELM), exhibited ‘sins of omission’, frequently failing to suggest which developed five physician-­validated taxonomies for critical diagnostic tests or treatments that a human physi- frontier models across 121 distinct clinical tasks.23 This cian would deem essential for safety.27 This limitation broader evaluation approach builds on a shift towards is further highlighted by the CRAFT-­ MD benchmark, more realistic, physician-­ verified assessments using where models, which appeared expert-­ level on struc- synthetically generated multiturn health conversations, tured summaries, frequently missed key history details, as evidenced in OpenAI’s HealthBench and Conversa- and diagnostic accuracy dropped substantially when tional Reasoning Assessment Framework for Testing in cases were converted to dialogue (eg, GPT-­4 from 0.82 Medicine (CRAFT-­MD) studies.24 25 Together, they found to 0.63 and to 0.49 without multiple-­ choice cues).25 steady progress from GPT-­3.5 (16%) to GPT-­4 (60%) on Finally, the consensus in 2025 reemphasises that models complex health reasoning and excellent performance lack the ability to self-­assess or recognise when they are in documentation and patient communication. By wrong. The MetaMedQA benchmark revealed that LLMs Figure 2 Clinical AI evaluation methodology spans across four levels with increasing amounts of real-­world validity: benchmark testing of clinical knowledge, reasoning assessment through simulated vignettes, retrospective validation on real patient datasets and prospective clinical trials in live care settings. Each stage increases the amount of methodological rigour and therefore provides more information about safety and proximity to real clinical deployment. AI, artificial intelligence; EHR, electronic health record. 4 Worth JE, et al. BMJ Digit Health 2026;2:e000113. doi:10.1136/bmjdh-2026-000113 BMJ Digital Health & AI Table 1 Summary table of selected benchmarks from 2025 Country BMJ Digital Health & AI: first published as 10.1136/bmjdh-2026-000113 on 29 September 2026. Downloaded from https://bmjdigitalhealth.bmj.com on 25 October 2026 by guest. Study Benchmark name of origin Purpose Methods 73 Aali et al MedVAL-­Bench USA Evaluate the factual correctness of A dataset of 840 physician-­annotated medical text generation outputs across six medical tasks collected from public medical datasets, with outputs annotated by physicians. Arora et al24 HealthBench USA Evaluate LLM performance 5000 multiturn conversations between a across realistic patient-­doctor model and an individual user or healthcare conversation tasks professional using conversation-­specific rubrics created by physicians. Bedi et al23 MedHELM USA Evaluate LLM performance Developed a benchmark suite that includes for real clinical tasks across 19 existing benchmarks, five reformulated clinical decision support, benchmarks based on existing datasets and Protected by copyright, including for uses related to text and data mining, AI training, and similar technologies. clinical note generation, patient 13 new benchmarks. Used accuracy-­based communication, medical research metrics and an ‘LLM-­jury’ for scores. and administration Dai et al74 Clinical Large-­ USA Evaluate models across Comprised of 4.51 million patient samples Scale Integrative multimodal clinical data in realistic across 2D imaging, 3D video, time series, Multimodal patient scenarios graphs and multimodal data, and models Benchmark testing against a ground truth comparison of expert-­validated answers. Diego et MedPI USA Evaluate LLMs in simulated Simulated data from 366 distinct ‘AI al75 multiturn, patient-­doctor patients’ and 7097 patient-­doctor conversations in primary care conversations were scored against 105 dimensions and judged by LLM judges. McCoy et Script concordance USA Evaluate models’ ability to 750 SCT questions across multiple al21 testing (SCT) produce consistent reasoning and specialties, testing the change in diagnosis answers, especially when new or management depending on added data. information is introduced Scored against expert-­panel responses. Mehandru ER-­Reason USA Evaluate LLM-­based clinical Data from 3984 patients, including clinical et al76 reasoning and decision-­making in notes, progress notes, history and physical the emergency room (ER) exams, consults, echo reports, imaging notes and ER provider documentation, with LLM results evaluated against clinician answers. Nori et al31 SDBench USA Evaluate iterative clinical diagnosis Converted 304 NEJM-­CPC cases into to benchmark diagnostic agents stepwise diagnostic encounters, with (human or AI) decision-­making performance assessed by diagnostic across multiple steps accuracy and total cost Wang et al5 Clinical Safety-­ China Evaluate LLMs against 30 metrics Synthesised 2069 clinical scenario questions Effectiveness Dual-­ with a focus on safety and across 26 specialties and scored responses Track Benchmark effectiveness on safety and effectiveness dimensions, combining into a total score. Wu et al27 NOHARM USA Evaluate the frequency Physician-­validated benchmark of 100 and severity of harm from real primary-­care to specialist consultation AI-­generated medical cases, covering 10 specialties. Given the recommendations patient case description and physician question, the LLM must list a complete management plan in free text. Zuo et al77 MedXpertQA China Evaluate expert-­level medical Included 4460 questions across 17 knowledge and advanced specialties. Evaluated LLMs in two subsets: reasoning text and multimodal. This is a non-­exhaustive list of benchmarks selected by expert review. AI, artificial intelligence; 2D, two-­dimensional; 3D, three-­dimensional; LLM, large language model ; MedHELM, Medical Holistic Evaluation of Large Language Models; NEJM-­CPC, New England Journal of Medicine clinicopathological conference; NOHARM, Numerous Options Harm Assessment for Risk in Medicine. Worth JE, et al. BMJ Digit Health 2026;2:e000113. doi:10.1136/bmjdh-2026-000113 5 BMJ Digital Health & AI consistently fail to recognise their own knowledge limita- preventative intervention. However, as the clinical scope tions, providing highly confident answers even when the of these models expands, the field is beginning to move correct option is absent from a question.28 This failure beyond single-­model architectures towards the complex BMJ Digital Health & AI: first published as 10.1136/bmjdh-2026-000113 on 29 September 2026. Downloaded from https://bmjdigitalhealth.bmj.com on 25 October 2026 by guest. in recall suggests that models cannot yet be trusted to orchestration of specialised AI agents. admit uncertainty, which is a prerequisite for safe clinical translation. Collectively, these findings demonstrate that Multiagent systems: orchestration and interaction challenges better evaluation, specifically frameworks that prioritise Emerging multiagent clinical AI systems suggest that the safety, longitudinal workflows and metacognition, is the way models are organised and interact may matter as essential prerequisite for trustworthy clinical AI. much as the capabilities of the individual models them- selves. Microsoft’s MAI Diagnostic Orchestrator demon- Automated evaluations, an aside strates this principle by simulating a panel of five physi- A substantial share of the benchmark and evaluation cians that sequentially request diagnostic information studies reviewed here rely on automated ‘LLM-­as-­judge’ until reaching high confidence, achieving about 80% scoring, including multiagent designs (eg, HealthBench, accuracy on NEJM case presentations while reducing Protected by copyright, including for uses related to text and data mining, AI training, and similar technologies. CRAFT-­MD) in which language models simulate patient, diagnostic costs by up to 70% compared with off-­the-­shelf clinician and adjudicator. This contrasts with scripted, o3.31 Critically, this performance arises not from a single ground-­truth-­anchored schemes such as NOHARM which superior model but from orchestration: agents collabo- uses correction algorithms where models are constrained ratively select the most informative next diagnostic test in their adjudication allowing repeatable external verifi- rather than exhaustively exploring all options. Multiagent cation. The reliability and potential bias of LLM judges is conversation frameworks reinforce this finding, showing itself a growing area of interest and concern; for now, we that diagnostic accuracy improves when rare-­disease cases suggest that results which depend on automated evalua- are evaluated through structured disagreement among tion should be interpreted with corresponding caution. several specialised agents and a supervising agent (78% vs 58% for standalone models).32 However, this approach Foundational methods has demonstrated shortcomings. Testing 26 multiagent Advancements in foundational methods such as LLM/ configurations on 2400 real emergency department cases AI system architecture and setup suggest that more struc- revealed an ‘optimisation paradox’, where the multiagent tured statistical grounding can expand the utility of clin- model that was assembled from the highest performing ical AI. individual agents (‘best-­of-­breed’ system) actually under- performed. It hit 86% component-­ level accuracy but Prediction: the era of medical event foundation models only 68% diagnostic accuracy, well below a less-­optimised Foundation models such as Delphi-­2M and Epic’s Cosmos system’s 77%.33 This gap in performance has been attrib- Medical Event Transformer (CoMET) can convert a uted to breakdowns in the flow of information between patient’s entire medical history into structured, sequen- agents, demonstrating that optimising individual agents tially readable units (a process called tokenisation), does not guarantee end-­to-­end system performance. enabling these systems to predict future clinical events at the population scale. These systems are trained on Multimodal systems: statistical and clinical grounding hundreds of thousands to 100 million patients and have Multimodal clinical AI systems process multiple types been shown to forecast future diagnoses and future clin- of data, such as clinical notes, lab results and medical ical events across thousands of conditions.29 30 Delphi-­2M, imaging, simultaneously. This system of integrating for instance, achieves an average area under the curve diverse medical data through a single model has been (AUC) of 0.76 in predicting the timing and occurrence shown to improve diagnosis, prognosis and treatment of over 1000 diseases, with the unique capacity to simu- selection in some studies. Systems like AMIE and Multi- late realistic lifetime health trajectories up to 20 years modal transformer with Unified maSKed modeling into the future. AUC alone is not enough to base real-­ demonstrate the superiority of this approach, leveraging world decision-­making, which requires external valida- multimodal reasoning to outperform primary care physi- tion and prospective studies. The model’s developers cians in diagnostic accuracy and surpass established note these limitations, cautioning against assuming the molecular biomarkers in precision oncology.15 34 Further- patterns reflect causation and stressing that any individ- more, real-­ time clinical copilots such as EyeFM have ual’s prediction remains heavily influenced by chance. validated these advancements in randomised controlled This caution extends to other models in the space. trials, showing dramatic improvements in referral accu- CoMET uses 115 billion clinical events to autoregres- racy and a significant increase in patient compliance with sively predict subsequent diagnoses and readmissions, follow-­up care.35 Yet, the impressive capabilities of these matching or exceeding the performance of task-­specific models do not protect them from fundamental errors. models without traditional fine-­tuning. By capturing the Notably, one study found that multimodal medical AI natural history of disease through these tokenised time- models can produce highly confident and detailed expla- lines, these models provide a pioneering level of support nations even when their diagnoses are incorrect, at times for anticipatory medicine, surfacing critical windows for describing visual features that are not actually present.36 6 Worth JE, et al. BMJ Digit Health 2026;2:e000113. doi:10.1136/bmjdh-2026-000113 BMJ Digital Health & AI This ‘illusion of readiness’ captures how easily shortcut In diagnostic interpretation, models can serve as a behaviours can be mistaken for genuine clinical and helpful copilot for clinicians. In a prospective, real-­world visual grounding. Overall, the literature suggests that implementation study in Germany across 12 sites, radiol- BMJ Digital Health & AI: first published as 10.1136/bmjdh-2026-000113 on 29 September 2026. Downloaded from https://bmjdigitalhealth.bmj.com on 25 October 2026 by guest. high benchmark performance does not translate into ogists who opted to use AI support achieved a higher clinical readiness, even with maturing coordination of breast cancer detection rate (6.7 vs 5.7 per 1000; 17.6% modalities. relative increase, CI 5.7% to 30.8%) without a signifi- cant increase in recall rate.42 In a primary care spirom- Reasoning models: optimising clinical logic and trust etry interpretation study, clinicians using AI performed While frontier models possess significant mathematical better at identifying the top diagnosis (mean difference, and computational proficiency, clinical reasoning is a 9.0; CI 4.5 to 13.3), establishing a diagnosis of chronic distinct cognitive domain that requires independent opti- obstructive pulmonary disease (mean difference, 15.9; CI misation, considering gaps and promise for improvement. 9.0 to 22.7) or including the diagnosis in the differential Primarily, the literature suggests that current biomedical diagnosis (mean difference, 7.3; CI 3.0 to 11.7), as deter- benchmarks often overestimate true reasoning ability: mined by expert pulmonologists.43 Protected by copyright, including for uses related to text and data mining, AI training, and similar technologies. one study found only 33% of questions require multi- step reasoning, while models lose 40–60% accuracy Administrative and workflow automation when forced to revise incorrect hypotheses.37 To address While much research focuses on diagnostic reasoning important limitations in optimised clinical reasoning, and clinical decision support, the most mature real-­world Med-­ process reward modeling (PRM) introduced a deployments of AI in healthcare today remain admin- medical domain-­specific process reward model. It scored istrative. Several studies showed that ambient scribes each reasoning step for factual accuracy, logical coher- reduce administrative burden, with physicians reporting ence and clinical relevance rather than evaluating only reduced cognitive workload, improved patient atten- final answers. Using a retrieval-­augmented generation-­ tion and less time spent on documentation.44 However, as-­ a-­ judge labelling pipeline where medical evidence measured time savings have been modest, with some supports the scoring of each reasoning chain link, Med-­ studies reporting around 20 s per note.45 46 Accuracy and PRM achieves 72–73% accuracy across seven benchmarks, substantially outperforming open-­source models with the the rates of clinically meaningful errors also remain a strongest gains on reasoning-­heavy tasks.38 concern, though current evidence is largely limited to retrospective or benchmark study designs.18 47 48 Areas such as clinical inbox management and discharge AI in clinical workflows summaries show similar trends of growing adoption but While real-­world AI deployments have improved diag- still require further testing to understand their impact nostic accuracy and reduced administrative burden, the across different clinical contexts. question of how to optimise human-­ AI collaboration, without eroding clinical skills, remains unresolved. Human-AI collaboration remains underdeveloped Even with advances in real-­ world deployment and Emerging evidence from real-world clinical deployment diagnostic interpretation, evidence suggests that a Recent studies suggest rapid improvement in the clin- limiting factor may lie in human-­AI interaction, rather ical reasoning capabilities of AI systems, although real-­ than model capability. In a meta-­analysis of 52 studies, world deployment remains constrained by workflow human-­AI teaming demonstrated higher diagnostic reli- complexity. AI has substantially improved since 2022, in ability than clinicians alone. However, full complemen- many cases outperforming clinicians on diagnostic and management tasks in controlled settings. However, in tarity, or when the combined performance of clinicians some of these studies, clinicians were evaluated without and models working together is greater than each in access to clinical decision support tools they would use isolation, was rarely achieved.49 A randomised controlled in practice, which could overstate AI’s relative advan- trial of 70 physicians showed that a custom collabora- tage.39 40 Clinicians were typically strongest at adapting tive GPT-­4 system designed to create a joint synthesis to inconsistent information and physical exam findings, of human and AI reasoning found improvements in highlighting the challenges that AI systems face in real-­ diagnostic accuracy when AI served as a second opinion world clinical environments.39 In 2025, a Penda Health (82% vs 75%; mean difference 6.8 percentage points, CI and OpenAI large-­scale quality improvement study with 4.0 to 9.6) or first opinion (85% vs 75%; mean difference LLM clinical decision-­support conducted across 15 clinic 9.9 percentage points, CI 4.7 to 15.0), with no significant sites in Kenya (39 849 visits) reported 16% and 13% rela- difference from AI-­alone performance (87%; p=0.20).50 tive reductions in physician-­rated diagnostic and treat- Similar results were observed in another randomized ment errors, though patient-­reported outcomes did not controlled trial (RCT) study, which tested Open AI GPT-­ differ significantly.41 Further studies are needed to assess 4’s ability to aid physicians with management decisions.40 clinical outcomes and generalisability, given the prompt-­ Despite this, human-­AI teaming did not outperform the heavy and context-­heavy design. model alone. Worth JE, et al. BMJ Digit Health 2026;2:e000113. doi:10.1136/bmjdh-2026-000113 7 BMJ Digital Health & AI Risks of automation bias and clinician deskilling that patient-­facing AI can influence health behaviours Alongside the opportunities from human-­AI collabora- and clinical outcomes, though interpretation is limited tion, the risks of routine AI exposure to clinical expertise by study design and potential human bias toward more BMJ Digital Health & AI: first published as 10.1136/bmjdh-2026-000113 on 29 September 2026. Downloaded from https://bmjdigitalhealth.bmj.com on 25 October 2026 by guest. remain largely unexamined. In 2025, research showed that empathetic-­perceived responses. AI literacy courses can improve physician-­AI teamwork.51 However, those benefits may be complicated by risks of Limits of patient-mediated safety validation automation bias. The randomised trial, conducted from Importantly, patients may not reliably distinguish AI-­gen- June to August 2025, reported that physicians exposed erated medical advice from that authored by clinicians to errors scored 14 percentage points (95% CI −18.9 and frequently rate AI responses, even low-­ accuracy to −9.1) lower in diagnostic accuracy than their peers outputs, as equally or more trustworthy and complete. in the control group.51 These findings, together with a Beyond rating AI responses as comparable to physi- retrospective observational study reporting declines in cians, participants reported being just as likely to follow adenoma detection rates among clinicians following potentially harmful low-­ accuracy recommendations as AI-­assisted colonoscopies,52 suggest that skill redistribu- on high-­accuracy AI or physician responses, suggesting that the burden of medical validation could increasingly Protected by copyright, including for uses related to text and data mining, AI training, and similar technologies. tion under sustained AI use follows a pattern consistent with prior waves of technology adoption, rather than a shift onto patients as more technology becomes patient-­ novel risk. The discussion now evolves to identifying facing.55 which critical skills are worth deliberately preserving, Applied AI and on creating safety structures to compensate for those Research on model capability is extensive, with studies that will diminish. Both over-­reliance on automation and across multiple medical specialties routinely showing clinician deskilling warrant further research to under- incremental, task-­ specific improvements. Randomised stand their causes and identify strategies that preserve prospective trials have already commenced and should be clinical expertise alongside AI adoption. the next wave of evidence in late 2026 and 2027 (table 2). Patient facing AI Specialty-tuned models for high-signal clinical tasks Patient-­facing AI shows early promise for improving By 2025, the most clinically translatable AI advances access and health behaviours, but patient preference is came from specialty-­tuned models solving specific prob- an unreliable proxy for safety, and evaluation frameworks lems with clear endpoints. Imaging remained the domi- have yet to catch up. nant use case, with high-­impact studies across pathology, neurology, nephrology, oncology, pulmonology and Patient-facing conversational AI cardiology. Several studies collectively demonstrated that In a randomised, double-­blind study of 159 simulated AI achieved specialist-­level accuracy in coeliac disease patient cases, Google’s AMIE outperformed 20 physi- diagnosis, carotid ultrasound, structural heart disease cians on nearly all measures of diagnostic accuracy, detection and breast cancer malignancy detection.59–62 communication quality and empathy.15 Similarly, special- ised systems continue to demonstrate performance gains The rise of domain-specific foundational models over general models. Google’s Personal Health LLM, a The more consequential development may be the shift Gemini-­based system fine-­tuned for personalised sleep from single-­ task models toward domain-­ specific foun- and fitness recommendations, performed at or above the dation models. LCTfound, a lung CT vision foundation level of human experts and meaningfully outperformed model, achieved an area under the receiver operating base models, underscoring the continued value of special- characteristic curve (AUROC) of 0.95 for rare diseases, isation even as general model capabilities improve.53 performed state-­ of-­ the-­art non-­ small cell lung cancer Beyond diagnostic performance, randomised and (NSCLC) prognosis stratification and received top Dice controlled evaluations tested whether patient-­ facing scores for mediastinal tumour segmentation.63 EchoP- AI influences behavioural and clinical outcomes. In rime, trained on 12.1 million echocardiogram videos and separate studies, participants preferred AI-­ generated cardiologist reports, matched or exceeded task-­specific responses over human-­ generated content in medical models across five health systems (mean AUC 0.85– advice and health coaching contexts, largely because the 0.92).64 Woollie, an oncology-­specific LLM trained on responses were easier to understand, used more positive Memorial Sloan Kettering data, generalised credibly to and encouraging language and provided clearer, more an external UCSF cohort (AUROC 0.97 internal; 0.88 on actionable guidance.54 55 One RCT involving 368 adults the UCSF cohort).65 Merlin, a three-­dimensional vision-­ found that an AI diabetes prevention programme was language model (VLM) for abdominal CT, similarly non-­inferior to human coaching in terms of weight, hemo- demonstrated strong generalisation across 752 individual globin A1c (HbA1c) and physical activity outcomes.56 tasks and over 44 000 external scans, outperforming two-­ Additional applications include patient-­ facing support dimensional VLMs, CT foundational models and off-­the-­ for understanding clinical notes and medical language shelf radiology models.66 The common theme is that a translation, with early studies demonstrating improved single well-­constructed foundation model can reduce the understanding.57 58 These findings provide early evidence need to build new architectures and labelling pipelines 8 Worth JE, et al. BMJ Digit Health 2026;2:e000113. doi:10.1136/bmjdh-2026-000113 BMJ Digital Health & AI Table 2 Summary table of selected randomised controlled trials from 2025 Country BMJ Digital Health & AI: first published as 10.1136/bmjdh-2026-000113 on 29 September 2026. Downloaded from https://bmjdigitalhealth.bmj.com on 25 October 2026 by guest. Study Title of origin AI system used Purpose 70 Unlu et al Manual vs AI-­Assisted USA Large Language Model To compare the efficiency and accuracy Prescreening for Trial Eligibility (LLM) (GPT-­4) processing of AI-­assisted versus manual clinician Using Large Language Models—A unstructured patient data. prescreening for heart failure trial Randomised Clinical Trial eligibility. 78 Wang et al An adaptive AI-­based virtual China Adaptive AI VR sports To evaluate the efficacy of REVERIE reality sports system for system using deep (Real-­World Exercise and Virtual Reality adolescents with excess body reinforcement learning Exercise Research in Education), a weight: a randomised controlled and transformer agents. virtual reality (VR)-­based exercise trial system, in reducing fat mass and improving cognitive and physical health in adolescents Protected by copyright, including for uses related to text and data mining, AI training, and similar technologies. Qazi et al51 AI-­literacy training enhances Pakistan LLMs used as diagnostic To determine if a structured AI-­literacy physician-­LLM diagnostic decision support. curriculum improves physician-­LLM collaboration in a resource-­limited diagnostic reasoning and reduces bias setting: a randomised controlled in healthcare settings. trial Qazi et al79 Automation Bias in Large Pakistan LLMs providing To assess whether AI-­trained Language Model Assisted diagnostic physicians remain vulnerable to Diagnostic Reasoning Among AI-­ recommendations. automation bias when exposed to Trained Physicians: A Randomised erroneous LLM recommendations. Clinical Trial Goh et al40 GPT-­4 assistance for improvement USA GPT-­4 with conventional To evaluate the impact of GPT-­4 on of physician performance on resources physician performance regarding open-­ patient care tasks: a randomised ended patient management reasoning controlled trial tasks. Wu et al35 An eyecare foundation model for China Multimodal vision-­ To evaluate a multimodal foundation clinical assistance: a randomised language foundation model as a clinical copilot for retinal controlled trial model disease screening. Abdallah et TrialMatchAI: an end-­to-­ France Retrieval-­Augmented To develop and validate a RAG-­based al80 end AI-­powered clinical trial Generation (RAG) framework for automated clinical trial recommendation system to framework using matching using entire unstructured streamline patient-­to-­trial Gemma-­2-­2B patient note histories. matching Afshar et A Pragmatic Randomised USA Abridge (Abridge AI) To evaluate the impact of ambient AI al46 Controlled Trial of Ambient scribes on healthcare practitioners’ Artificial Intelligence to Improve well-­being (burnout and fulfilment) and Health Practitioner Well-­Being documentation efficiency in a real-­ world deployment. Lukac et Ambient AI Scribes in Clinical USA Microsoft Dragon To investigate the effects of two al45 Practice: A Randomised Trial Ambient eXperience AI scribes on documentation time, Copilot and Nabla physician psychometrics, usability, accuracy and safety. This list of randomised controlled trials was selected from the State of Clinical AI Report 2026. AI, artificial intelligence. for every clinical question in the domain. With this shift tools for ruling out acute myocardial infarction, and come challenges such as the need for large, high-­quality augmenting the History, ECG, Age, Risk Factors, and data, the risk of limited transferability across clinical Troponin (HEART) score with AI-­ ECG improved net contexts and increased validation complexity. reclassification by 20% across 18 university-­level teaching AI-enabled risk stratification and triage hospitals in South Korea.67 Further prospective research Alongside imaging, specialties are increasingly repur- suggested earlier patient identification in primary care posing abundant, non-­ invasive data to improve risk diabetic kidney disease screening and fewer complica- stratification. In the emergency department, ROMIAE tions in colorectal cancer surgery.68 Operational use achieved accuracy comparable to established clinical cases are also broadening in parallel, including triaging Worth JE, et al. BMJ Digit Health 2026;2:e000113. doi:10.1136/bmjdh-2026-000113 9 BMJ Digital Health & AI patients/referrals, clinical trial eligibility and adjudi- models’ strong performance on standard benchmarks.27 cating adverse events in clinical trials.69–71 Similarly, the persistent inability of models to calibrate their own confidence portends a concerning possibility that the BMJ Digital Health & AI: first published as 10.1136/bmjdh-2026-000113 on 29 September 2026. Downloaded from https://bmjdigitalhealth.bmj.com on 25 October 2026 by guest. Prospective evaluation and real-world deployment systems most likely to cause harm are also the least likely to While frontier models strive for field-­wide transformation, signal uncertainty.28 This concern is compounded in patient-­ real paths to clinical deployment are prospectively studying facing contexts, where patients cannot reliably distinguish models within context and task-­ specific domains. In the AI-­generated advice from that authored by clinicians and largest real-­world evaluation of stroke AI to date, England’s frequently rate low-­accuracy AI outputs as equally or more National Health Service deployed Brainomix 360 across 26 trustworthy, suggesting that safety assurance cannot rely on hospitals and observed a 100% increase in endovascular patient-­level discernment and must instead be built upstream thrombectomy rates at AI-­enabled sites, significantly greater through evaluation, transparency and accountability struc- than trends in hospitals without AI. Transfer time was also tures.55 As that report argues, a coordinated national infra- reduced and AI-­reviewed patients had higher odds of good structure for clinical AI evaluation, analogous to the federal functional outcomes (OR 1.16, 95% CI 1.07 to 1.25).72 investment that enabled EHR adoption, may be necessary to With gains in clinical accuracy, the focus should now be on Protected by copyright, including for uses related to text and data mining, AI training, and similar technologies. close this governance gap. prospective clinical trials to assess patient outcomes and the optimal integration of workflows. Where to go from here Three priorities should guide the field. First, research investment should shift toward prospective, multisite trials DISCUSSION that evaluate AI systems (and claimed capabilities) within Several themes emerge from this synthesis that merit authentic clinical workflows, measuring not only diagnostic explicit consideration: accuracy but also downstream patient outcomes, clinician workflow impact and safety events. Second, evaluation must The benchmark-to-bedside gap is a structural problem evolve beyond accuracy-­centric benchmarks to frameworks A consistent finding across domains is that strong perfor- that assess metacognitive calibration, performance under mance on controlled evaluations does not reliably predict uncertainty, equity across populations and behaviour over clinical utility. Models that achieve near-­perfect scores on longitudinal patient trajectories. Lastly, and most impor- medical licensing examinations fail to recognise when tantly, patient and user safety should be placed at the fore- the correct answer is absent.20 Systems that produce front. In addition to ensuring that these models minimise expert-­ level diagnoses from structured vignettes lose harm to patients, a more pernicious danger is harm to the substantial accuracy when the same cases are presented physicians or healthcare staff using them; care must be taken as unstructured dialogue.25 Multiagent architectures with to avoid physician deskilling or automation bias; hence, individually superior components can underperform future work on safe human-­computer interaction is critical. less-­optimised configurations due to failures in informa- tion flow.33 These studies reflect a fundamental mismatch Limitations between how models are evaluated and how medicine is This narrative review has several limitations. It draws primarily practised. Closing this gap will require evaluation frame- on the ARISE’s State of Clinical AI Report 2026 and a targeted works that highlight uncertainty handling, longitudinal literature search rather than a systematic review with prede- reasoning and performance under incomplete informa- fined search protocols. Study selection was supplemented tion rather than accuracy on static, well-­defined tasks. by expert nomination, which may introduce selection bias towards studies from well-­networked academic institutions. Human–AI collaboration is the critical unsolved problem The review reflects a temporal snapshot of a rapidly evolving Across multiple randomised trials, clinician-­ AI teams field; findings may shift as preprints undergo peer review failed to outperform the AI model alone, despite consist- and new models are released. This review also has limitations ently outperforming clinicians working without AI.40 This of scope. By concentrating on the six domains where recent ceiling effect suggests that current interaction paradigms evidence clustered, we do not address equity and algorithmic do not enable clinicians to effectively leverage AI capabili- bias, regulation and governance or implementation science ties. Compounding this, emerging evidence of automation in depth. Equity in particular remains a first-­order concern bias and deskilling suggests that poorly designed AI inte- as model performance depends on the representativeness gration could erode the very clinical expertise it is meant to of training data, and few studies reviewed report outcomes augment.51 52 Designing human–AI workflows that achieve stratified by population, meriting future dedicated scrutiny genuine complementarity, where each corrects the other’s as the evidence base matures. errors, remains an open research challenge and arguably the highest-­priority prerequisite for safe deployment. CONCLUSION Safety evaluation lags behind capability development What changed in 2025–2026? As expected, models have The NOHARM framework’s finding that frontier models increased in size, scope and ambition. For example, medical-­ produce potentially harmful outputs in approximately one event foundation models matured from proof-­ of-­ concept in five primary care consultations is striking, given these same to systems that tokenise longitudinal records to forecast 10 Worth JE, et al. BMJ Digit Health 2026;2:e000113. doi:10.1136/bmjdh-2026-000113 BMJ Digital Health & AI clinical trajectories (Delphi-­2M, CoMET).29 30 Perhaps even authors reviewed and approved the final manuscript. The corresponding author, more importantly, we see progress in how we evaluate these EG, attests that all listed authors meet authorship criteria and that no others meeting the criteria have been omitted. AI was used for minor grammatical edits. models, prospectively and in mundo: for the first time, a BMJ Digital Health & AI: first published as 10.1136/bmjdh-2026-000113 on 29 September 2026. Downloaded from https://bmjdigitalhealth.bmj.com on 25 October 2026 by guest. Content drafting was manually done by authors first, with subsequent review by AI, large-­ scale prospective trial tested LLM decision support with content suggestions considered or rejected at the discretion of the authors. in routine care.41 This past year, we have also seen some of All final content was reviewed by authors. EG is the guarantor and accepts full the dangers of overreliance on these tools: the first RCT-­ responsibility for the work and/or conduct of the study. grade and observational-­grade evidence of automation bias Funding The authors have not declared a specific grant for this research from any funding agency in the public, commercial or not-­for-­profit sectors. and clinician deskilling emerged51 52; multiagent systems revealed an unexpected ‘optimisation paradox’ in which Disclaimer This review is presented in written narrative form the ARISE network’s ‘State of Clinical AI Report 2026’ presentation, published online January 2026, with added components can degrade rather than improve perfor- which it shares authorship, the six-­domain organisational frameworks and the mance.33 Together these mark a shift from asking ‘what can study selection. This article extends that report by synthesising key insights through models do’ to asking, ‘can we safely, durably and effectively formal narrative analysis, subjecting findings to rigorous peer review and broadening deploy them?’ This is the question the next year of inquiry accessibility of healthcare AI evidence across academic and public audiences. must answer. Competing interests Dr EG receives funding from the Gordon and Betty Moore Foundation, Macy Foundation, Stanford Artificial Intelligence in Medicine and Protected by copyright, including for uses related to text and data mining, AI training, and similar technologies. Clinical AI in 2026 presents a key challenge: model Imaging—Human-­Centered Artificial Intelligence Partnership Grant, Stanford capabilities are advancing faster than the evidence, evalu- Bio-­X Interdisciplinary Initiatives Seed Grants Program (IIP) [R12]. Dr EG ation infrastructure and governance structures needed to reports consulting fees or other compensation from Google, OpenAI, Samsung deploy them safely. Across six domains, we find that the Research America, Roche Diagnostics, Novartis, Hello Heart, Grow Care and most credible translational progress occurs on narrowly Faculty Connection. Dr Rodman reports funding from the Moore Foundation, Macy Foundation, NIH, ARPA-­H, Google and Google DeepMind. Dr Chen reports defined tasks, whether tackled by specialty-­tuned models cofounding Reaction Explorer, which develops and licenses organic chemistry or general-­purpose LLMs, supported by prospective eval- education software, as well as paid consulting fees from Sutton Pierce, Younker uation and clear workflow integration, while broader Hyde Macfarlane and Sykes McAllister as a medical expert witness. He receives clinical autonomy remains premature. The gap between funding from the National Institutes of Health (NIH)/National Institute of Allergy and Infectious Diseases (1R01AI17812101), the NIH–National Center for Advancing what clinical AI can do in controlled settings and what it Translational Sciences Clinical & Translational Science Award (UM1TR004921), can be trusted to do at the bedside is one of the defining the Stanford Bio-­X Interdisciplinary Initiatives Seed Grants Program (IIP) [R12] tensions of this new field. Bridging it will require not [JHC], the NIH/Center for Undiagnosed Diseases at Stanford (U01 NS134358), the only better models but also better trials, better bench- Stanford RAISE Health Seed Grant (2024), the Josiah Macy Jr. Foundation (AI in Medical Education) and the Stanford CARE AI Scholar Fellowship. marks and a better understanding of how best to partner Patient and public involvement Patients and/or the public were not involved in human and machine intelligence. the design, conduct, reporting or dissemination plans of this research. Author affiliations Patient consent for publication Not applicable. 1 The University of Arizona College of Medicine Phoenix, Phoenix, Arizona, USA Ethics approval Not applicable. 2 Stanford Department of Medicine, Stanford, Stanford, California, USA 3 Provenance and peer review Commissioned; externally peer reviewed. Radiation Oncology, Stanford University, Stanford, California, USA 4 Department of Internal Medicine, Beth Israel Deaconess Medical Center, Boston, Supplemental material This content has been supplied by the author(s). It has Massachusetts, USA not been vetted by BMJ Publishing Group Limited (BMJ) and may not have been 5 Department of Medicine, New York-­Presbyterian/Columbia University Irving peer-­reviewed. Any opinions or recommendations discussed are solely those Medical Center, New York, New York, USA of the author(s) and are not endorsed by BMJ. BMJ disclaims all liability and 6 Department of Biomedical Informatics, Harvard Medical School, Boston, responsibility arising from any reliance placed on the content. Where the content includes any translated material, BMJ does not warrant the accuracy and reliability Massachusetts, USA 7 of the translations (including but not limited to local regulations, clinical guidelines, Department of Dermatology, Harvard Medical School, Boston, Massachusetts, USA 8 terminology, drug names and drug dosages) and is not responsible for any error Beth Israel Deaconess Medical Center, Boston, Massachusetts, USA 9 and/or omissions arising from translation and adaptation or otherwise. Medical University of South Carolina, Charleston, South Carolina, USA 10 Massachusetts Institute of Technology Institute for Medical Engineering & Open access This is an open access article distributed in accordance with the Science, Cambridge, Massachusetts, USA Creative Commons Attribution Non Commercial (CC BY-­NC 4.0) license, which 11 Division of Neurology, University of Alberta, Edmonton, Alberta, Canada permits others to distribute, remix, adapt, build upon this work non-­commercially, 12 Department of Internal Medicine, Cambridge Health Alliance, Cambridge, and license their derivative works on different terms, provided the original work is Massachusetts, USA properly cited, appropriate credit is given, any changes made indicated, and the 13 Hospital Medicine, Stanford University School of Medicine, Stanford, California, use is non-­commercial. See https://creativecommons.org/licenses/by-nc/4.0/. USA ORCID iDs 14 The Institute of Medical Education Research Rotterdam, Rotterdam, The Anastasia Perez https://orcid.org/0009-0009-0836-4056 Netherlands Ethan Goh https://orcid.org/0000-0003-1965-025X 15 Department of Emergency Medicine, Beth Israel Deaconess Medical Center, Boston, Massachusetts, USA 16 Stanford Division of Computational Medicine, Stanford University, Stanford, California, USA 17 Stanford Clinical Excellence Research Center, Stanford University, Stanford, California, USA REFERENCES 18 1 Google Slides. State of clinical AI report 2026. Available: https://​ Division of Hospital Medicine, Department of Medicine, Stanford University, docs.google.com/presentation/d/1A-TcHQb5Hg3-0MoiUFV199FI​ Stanford, California, USA DBgmGvYhh9DyEU9C_Po/edit?slide=id.g3a1c1566c13_0_0#slide=​ id.g3a1c1566c13_0_0 [Accessed 02 Jul 2026]. Contributors PB, DJHW, and EG conceived the review. PB led the literature 2 Brodeur PG, Buckley TA, Kanjee Z, et al. Superhuman performance search and synthesis of evidence. JEW, AP, and DJHW drafted the manuscript and of a large language model on the reasoning tasks of a physician. figures. EG, AR, and JHC contributed to interpretation and critical revisions. All arXiv 2025. Worth JE, et al. BMJ Digit Health 2026;2:e000113. doi:10.1136/bmjdh-2026-000113 11 BMJ Digital Health & AI 3 Preiksaitis C, Ashenburg N, Bunney G, et al. The Role of Large 32 Chen X, Yi H, You M, et al. Enhancing diagnostic capability with Language Models in Transforming Emergency Medicine: Scoping multi-­agents conversational large language models. Npj Digit Med Review. JMIR Med Inform 2024;12:e53787. 2025;8:159. BMJ Digital Health & AI: first published as 10.1136/bmjdh-2026-000113 on 29 September 2026. Downloaded from https://bmjdigitalhealth.bmj.com on 25 October 2026 by guest. 4 Li H, Zhang R, Lee Y-­C, et al. Systematic review and meta-­analysis 33 Bedi S, Mlauzi I, Shin D, et al. The optimization paradox in clinical AI of AI-­based conversational agents for promoting mental health and multi-­agent systems. arXiv 2025. well-­being. NPJ Digit Med 2023;6:236. 34 Xiang J, Wang X, Zhang X, et al. A vision–language foundation 5 Wang L, Bhanushali T, Huang Z, et al. Evaluating Generative AI in model for precision oncology. Nature 2025;638:769–78. Mental Health: Systematic Review of Capabilities and Limitations. 35 Wu Y, Qian B, Li T, et al. An eyecare foundation model for clinical JMIR Ment Health 2025;12:e70014. assistance: a randomized controlled trial. Nat Med 2025;31:3404–13. 6 Gong EJ, Bang CS, Lee JJ, et al. Large Language Models in 36 Gu Y, Fu J, Liu X, et al. The illusion of readiness in health AI. arXiv Gastroenterology: Systematic Review. J Med Internet Res 2025. 2024;26:e66648. 37 Thapa R, Wu Q, Wu K, et al. Disentangling reasoning and knowledge 7 Chen SF, Alyakin A, Seas A, et al. LLM-­assisted systematic in medical large language models. arXiv 2025. review of large language models in clinical medicine. Nat Med 38 Yun J, Sohn J, Park J, et al. Med-­prm: medical reasoning models 2026;32:1152–9. with stepwise, guideline-­verified process rewards. arXiv 2025. 8 OpenAI. Introducing ChatGPT health. 2026. Available: https://​ 39 Zeltzer D, Kugler Z, Hayat L, et al. Comparison of Initial openai.com/index/introducing-chatgpt-health/ [Accessed 12 Jun Artificial Intelligence (AI) and Final Physician Recommendations 2026]. in AI-­Assisted Virtual Urgent Care Visits. Ann Intern Med 9 Introducing Perplexity Health. 2026. Available: https://www.​ 2025;178:498–506. perplexity.ai/hub/blog/introducing-perplexity-health [Accessed 17 40 Goh E, Gallo RJ, Strong E, et al. GPT-­4 assistance for improvement Protected by copyright, including for uses related to text and data mining, AI training, and similar technologies. Apr 2026]. of physician performance on patient care tasks: a randomized 10 Microsoft AI. Introducing copilot health. Available: https://microsoft.​ controlled trial. Nat Med 2025;31:1233–8. ai/story/introducing-copilot-health/ [Accessed 12 Jun 2026]. 41 Korom R, Kiptinness S, Adan N, et al. AI-­based clinical decision 11 Advancing claude in healthcare and the life sciences. Available: support for primary care: a real-­world study. arXiv 2025. https://www.anthropic.com/news/healthcare-life-sciences [Accessed 42 Eisemann N, Bunk S, Mukama T, et al. Nationwide real-­world 12 Jun 2026]. implementation of AI for cancer detection in population-­based 12 The 2026 AI index report | stanford HAI. Available: https://hai.​ mammography screening. Nat Med 2025;31:917–24. stanford.edu/ai-index/2026-ai-index-report [Accessed 12 Jun 43 Doe G, Banya W, Edwards G, et al. AI-­Assisted Spirometry 2026]. Interpretation in Primary Care: A Randomized Controlled Trial. NEJM 13 Angus DC, Khera R, Lieu T, et al. AI, Health, and Health Care Today AI 2025;2:AIoa2400804. and Tomorrow: The JAMA Summit Report on Artificial Intelligence. 44 Olson KD, Meeker D, Troup M, et al. Use of Ambient AI Scribes to JAMA 2025;334:1650–64. Reduce Administrative Burden and Professional Burnout. JAMA 14 Buckley TA, Conci R, Brodeur PG, et al. Teaching large language Netw Open 2025;8:e2534976. models to reason like expert diagnosticians. arXiv 2026. 45 Lukac PJ, Turner W, Vangala S, et al. Ambient AI Scribes in Clinical Practice: A Randomized Trial. NEJM AI 2025;2:AIoa2501000. 15 Tu T, Schaekermann M, Palepu A, et al. Towards conversational 46 Afshar M, Ryan Baumann M, Resnik F, et al. A Pragmatic diagnostic artificial intelligence. Nature 2025;642:442–50. Randomized Controlled Trial of Ambient Artificial Intelligence 16 Jin Q, Wang Z, Yang Y, et al. AgentMD: Empowering language to Improve Health Practitioner Well-­Being. NEJM AI agents for risk prediction with large-­scale clinical tool learning. Nat 2025;2:AIoa2500945. Commun 2025;16:9377. 47 Williams CYK, Subramanian CR, Ali SS, et al. Physician- and Large 17 Metwally AA, Heydari AA, McDuff D, et al. Insulin resistance Language Model-­Generated Hospital Discharge Summaries. JAMA prediction from wearables and routine blood biomarkers. Nature Intern Med 2025;185:818–25. 2026;652:451–61. 48 Asgari E, Montaña-­Brown N, Dubois M, et al. A framework to assess 18 Scheid MR, Friedmann B, Oppenheim M, et al. Development clinical safety and hallucination rates of LLMs for medical text and validation of a clinical wearable deep learning based summarisation. NPJ Digit Med 2025;8:274. continuous inhospital deterioration prediction model. Nat Commun 49 Liu P, Zhang J, Chen S, et al. Human-­AI teaming in healthcare: 2025;16:9513. 1 + 1 > 2? Npj Artif Intell 2025;1:47. 19 Li Y, Huang Q, Jiang J, et al. Large language model-­based 50 Everett SS, Bunning BJ, Jain P, et al. From tool to teammate: a biological age prediction in large-­scale populations. Nat Med randomized controlled trial of clinician-­ai collaborative workflows for 2025;31:2977–90. diagnosis. Health Informatics [Preprint] 2025. 20 Bedi S, Jiang Y, Chung P, et al. Fidelity of Medical Reasoning in 51 Qazi IA, Ali A, Khawaja AU, et al. AI-­literacy training enhances Large Language Models. JAMA Netw Open 2025;8:e2526021. physician-­llm diagnostic collaboration in a resource-­limited setting: a 21 McCoy LG, Swamy R, Sagar N, et al. Assessment of Large randomized controlled trial. medRxiv [Preprint] 2025. Language Models in Clinical Reasoning: A Novel Benchmarking 52 Budzyń K, Romańczyk M, Kitala D, et al. Endoscopist deskilling Study. NEJM AI 2025;2:AIdbp2500120. risk after exposure to artificial intelligence in colonoscopy: a 22 Bedi S, Liu Y, Orr-­Ewing L, et al. Testing and Evaluation of Health multicentre, observational study. Lancet Gastroenterol Hepatol Care Applications of Large Language Models: A Systematic Review. 2025;10:896–903. JAMA 2025;333:319–28. 53 Khasentino J, Belyaeva A, Liu X, et al. A personal health large 23 Bedi S, Cui H, Fuentes M, et al. Holistic evaluation of large language language model for sleep and fitness coaching. Nat Med models for medical tasks with MedHELM. Nat Med 2026;32:943–51. 2025;31:3394–403. 24 Arora RK, Wei J, Hicks RS, et al. HealthBench: evaluating large 54 Mantena S, Johnson A, Oppezzo M, et al. Fine-­tuning LLMs in language models towards improved human health. arXiv 2025. behavioral psychology for scalable health coaching. NPJ Cardiovasc 25 Johri S, Jeong J, Tran BA, et al. An evaluation framework for clinical Health 2025;2:48. use of large language models in patient interaction tasks. Nat Med 55 Shekar S, Pataranutaporn P, Sarabu C, et al. People Overtrust 2025;31:77–86. AI-­Generated Medical Advice despite Low Accuracy. NEJM AI 26 Jiang Y, Black KC, Geng G, et al. MedAgentBench: A Virtual 2025;2:AIoa2300015. EHR Environment to Benchmark Medical LLM Agents. NEJM AI 56 Mathioudakis N, Lalani B, Abusamaan MS, et al. An AI-­Powered 2025;2:AIdbp2500144. Lifestyle Intervention vs Human Coaching in the Diabetes Prevention 27 Wu D, Haredasht FN, Maharaj SK, et al. First, do NOHARM: Program: A Randomized Clinical Trial. JAMA 2025;334:2079–89. towards clinically safe large language models. ArXiv 57 Ray M, Kats DJ, Moorkens J, et al. Evaluating a Large Language 2025:arXiv:2512.01241v2. Model in Translating Patient Instructions to Spanish Using a 28 Griot M, Hemptinne C, Vanderdonckt J, et al. Large Language Standardized Framework. JAMA Pediatr 2025;179:1026–33. Models lack essential metacognition for reliable medical reasoning. 58 Kumar A, Wang H, Muir KW, et al. A Cross-­Sectional Study of GPT-­ Nat Commun 2025;16:642. 4–Based Plain Language Translation of Clinical Notes to Improve 29 Shmatko A, Jung AW, Gaurav K, et al. Learning the natural Patient Comprehension of Disease Course and Management. NEJM history of human disease with generative transformers. Nature AI 2025;2:AIoa2400402. 2025;647:248–56. 59 Luo L, Wu M, Li M, et al. A large model for non-­invasive and 30 Waxler S, Blazek P, White D, et al. Generative medical event models personalized management of breast cancer from multiparametric improve with scale. arXiv 2025. MRI. Nat Commun 2025;16:3647. 31 Nori H, Daswani M, Kelly C, et al. Sequential diagnosis with 60 Poterucha TJ, Jing L, Ricart RP, et al. Detecting structural heart language models. arXiv 2025. disease from electrocardiograms using AI. Nature 2025;644:221–30. 12 Worth JE, et al. BMJ Digit Health 2026;2:e000113. doi:10.1136/bmjdh-2026-000113 BMJ Digital Health & AI 61 Jiang H, Zhao A, Yang Q, et al. Towards expert-­level autonomous 70 Unlu O, Varugheese M, Shin J, et al. Manual vs AI-­Assisted carotid ultrasonography with large-­scale learning-­based robotic Prescreening for Trial Eligibility Using Large Language Models-­A system. Nat Commun 2025;16:7893. Randomized Clinical Trial. JAMA 2025;333:1084–7. BMJ Digital Health & AI: first published as 10.1136/bmjdh-2026-000113 on 29 September 2026. Downloaded from https://bmjdigitalhealth.bmj.com on 25 October 2026 by guest. 62 Jaeckle F, Denholm J, Schreiber B, et al. Machine Learning 71 Marti-­Castellote PM, Badrouchi S, Claggett B, et al. Using Artificial Achieves Pathologist-­Level Coeliac Disease Diagnosis. NEJM AI Intelligence to Adjudicate Major Adverse Cardiovascular Events in 2025;2:aioa2400738. Clinical Trials. J Am Coll Cardiol 2026;87:1856–66. 72 Nagaratnam K, Neuhaus AA, Fensome L, et al. Artificial intelligence 63 Gao Z, Zhang G, Liang H, et al. A lung CT vision foundation model imaging decision support for acute stroke treatment in England: a facilitating disease diagnosis and medical imaging. Nat Commun prospective observational study. Lancet Digit Health 2025;7:100927. 2025;17:35. 73 Aali A, Bikia V, Varma M, et al. 2026 Toward expert-­level medical text 64 Vukadinovic M, Chiu I-­M, Tang X, et al. Comprehensive validation with language models. Npj Digital Medicine. echocardiogram evaluation with view primed vision language AI. 74 Dai W, Chen P, Lu M, et al. CLIMB: data foundations for large scale Nature 2026;650:970–7. multimodal clinical foundation models. arXiv [Preprint 2025. 65 Zhu M, Lin H, Jiang J, et al. Large language model trained on 75 Diego FV, Proniakin O, Gruber V-­E, et al. 2026 MedPI: Evaluating AI clinical oncology data predicts cancer progression. NPJ Digit Med Systems in Medical Patient-­facing Interactions. Health Informatics. 2025;8:397. 76 Mehandru N, Golchini N, Garg N, et al. ER-­Reason: a benchmark 66 Blankemeier L, Kumar A, Cohen JP, et al. Merlin: a computed dataset for LLM clinical reasoning in the emergency room. arXiv tomography vision-­language foundation model and dataset. Nature [Preprint 2025. 2026;652:1318–28. 77 Zuo Y, Qu S, Li Y, et al. MedXpertQA: benchmarking expert-­level 67 Lee MS, Shin TG, Lee Y, et al. Artificial intelligence applied to medical reasoning and understanding. arXiv [Preprint 2025. Protected by copyright, including for uses related to text and data mining, AI training, and similar technologies. electrocardiogram to rule out acute myocardial infarction: the 78 Wang J, Qin Y, Wu Q, et al. An adaptive AI-­based virtual reality sports system for adolescents with excess body weight: a ROMIAE multicentre study. Eur Heart J 2025;46:1917–29. randomized controlled trial. Nat Med 2025;31:2255–68. 68 Rosen AW, Ose I, Gögenur M, et al. Clinical implementation of an AI-­ 79 Qazi IA, Ali A, Khawaja AU, et al. Automation Bias in Large Language based prediction model for decision support for patients undergoing Model–Assisted Diagnostic Reasoning among Physicians Trained in colorectal cancer surgery. Nat Med 2025;31:3737–48. AI Literacy — A Randomized Clinical Trial. NEJM AI 2026;3. 69 Gaber F, Shaik M, Allega F, et al. Evaluating large language model 80 Abdallah M, Nakken S, Georges M, et al. 2026 TrialMatchAI: an workflows in clinical decision support for triage and referral and end-­to-­end AI-­powered clinical trial recommendation system to diagnosis. NPJ Digit Med 2025;8:263. streamline patient-­to-­trial matching. Nat Commun17. Worth JE, et al. BMJ Digit Health 2026;2:e000113. doi:10.1136/bmjdh-2026-000113 13