Artificial intelligence systems integrated with physician oversight generally improve diagnostic performance and workflow in controlled studies, although evidence for patient outcomes, governance, and real-world implementation remains limited, according to a scoping review published in npj Digital Medicine.  

Three implications for successful integration of AI in health care were identified: evaluation must be more task- and context-specific; collaboration requires consideration to human and organizational factors; and accountability and patient safety must be treated as core evaluative concerns and not secondary ethical reflections. 

Researchers from the University of Oxford in the UK and collaborating institutions reviewed empirical studies of human-artificial intelligence collaboration (HAIC) in health care published between January 2015 and October 2025. Using Joanna Briggs Institute methodology and PRISMA-ScR reporting guidelines, they screened 17,463 records from five databases and included 140 studies. The review addressed three research questions. 

How is effectiveness defined and evaluated across clinical task contexts? 

Diagnostic interpretation accounted for the largest evidence base, representing 88 of 140 studies. Screening and triage comprised 20 studies, therapeutic decision-making 26 studies, and administrative and documentation workflows 11 studies, with some studies spanning multiple categories. Overall, 86 studies directly compared physician-only performance with physician-AI collaboration, whereas 54 used qualitative, survey, or other non-comparative designs. 

What technical, human, and organizational determinants shape successful human-AI collaboration? 

Across all included studies, 121 (86%) reported positive-direction findings according to each study’s own predefined outcome measures, such as diagnostic performance, efficiency, or user acceptance. Sixteen studies reported neutral or mixed findings, and three reported negative findings. However, the researchers cautioned that these figures should not be interpreted as pooled estimates of effectiveness because study designs, clinical tasks, and outcome measures varied substantially. 

Among studies that empirically compared physician-only with physician-AI performance, 76 reported positive outcomes. Studies also found that poor workflow integration or excessive false alarms increased review burden and disrupted performance. 

Evaluation methods varied according to clinical task. Diagnostic studies emphasized task-level accuracy, sensitivity, area under the receiver operating characteristic curve, efficiency, or inter-rater agreement. Screening and triage studies focused more on workflow efficiency, time-to-decision, prioritization, and resource utilization.  

Therapeutic decision-making studies more frequently incorporated downstream patient outcomes, workflow measures, clinician behavior, implementation metrics, and qualitative assessments of usability and trust. Administrative workflow studies focused primarily on documentation time, workflow fit, and acceptability. 

Trust emerged as a recurring but inconsistently measured determinant of successful collaboration. Only 57 studies explicitly assessed trust. Self-reported surveys were the most common measurement used in 41 studies, followed by qualitative assessments in 19 studies, behavioral measures in 17 studies, and calibration metrics in six studies. Researchers found that increased self-reported trust did not necessarily correspond with appropriately calibrated reliance on AI recommendations. 

Explainability also produced mixed results. Fifty-three studies examined explanation methods such as heat maps, confidence scores, or example-based reasoning. Thirty-eight studies reported that explanation methods improved interpretability, diagnostic performance, or both, whereas 22 found that certain explanation formats increased cognitive workload or encouraged overreliance by making AI recommendations appear more certain than warranted. 

Training and workflow integration consistently influenced successful implementation. Thirty-five studies reported that clinician training on AI tools, uncertainty information, and explanation formats was associated with better diagnostic performance and more appropriate reliance on AI, particularly among resident physicians and novice users. Eleven implementation-focused studies also highlighted the importance of seamless electronic health record integration, visible leadership, and engagement of multiple stakeholders during deployment. 

What are the ethical implications for accountability and patient safety?  

Ethical and governance issues were frequently discussed but rarely evaluated empirically. Twenty-seven studies addressed accountability. Many concluded that physicians retain responsibility for diagnostic and treatment decisions despite AI involvement. Thirty-one studies examined patient safety risks, identifying automation bias, confirmation bias, and inappropriate reliance on incorrect recommendations as potential hazards. Suggested safeguards included explainability, clinician training, and mechanisms allowing physicians to override AI recommendations.  

The study had several limitations. As a scoping review, the study mapped existing evidence rather than performing a meta-analysis or formal risk-of-bias assessment. The review included only English-language publications and did not systematically search gray literature. Most evidence came from controlled diagnostic studies rather than real-world clinical implementations, and relatively few studies evaluated patient-level outcomes or long-term effects on clinician performance and skill retention. 

"Our findings suggest that future HAIC research should progress from short-term reader studies and toward longitudinal and in-situ evaluations of how collaboration reshapes clinical work, patient outcomes, and professional responsibility over time," wrote first author Joshua Strong, a doctoral student with the Institute of Biomedical Engineering at the University of Oxford, and colleagues. 

The researchers reported no conflicts of interest. 

(Editor’s note: The researchers noted that the study manuscript will undergo further editing before final publication, and there may be errors present that affect the content. All legal disclaimers apply.)