This systematic review was conducted in accordance with the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses; Checklist 1) guidelines. The review protocol was registered in the PROSPERO (International Prospective Register of Systematic Reviews; registration number: CRD420251208020). A convergent segregated mixed methods systematic review design was used.

A comprehensive literature search was performed across multiple electronic databases, including PubMed, CINAHL, Embase, Web of Science, SinoMed, Cochrane Library, CNKI, Wanfang database, Google Scholar, VIP database, and Scopus, from January 1, 2014, to September 30, 2025. The search strategy combined keywords and Medical Subject Headings related to three core concepts: (1) AI, (2) simulation, and (3) nursing education. Sample search terms included “artificial intelligence,” “machine learning,” “AI,” “virtual patient,” “intelligent tutoring system,” “simulation,” “nursing education,” “nursing student,” and “learning outcomes.” The full search strategy for PubMed is provided in Multimedia Appendix 1.

Studies were included if they met the following criteria: (1) population: nursing students at any educational level (undergraduate, postgraduate, or continuing education) and practicing nurses enrolled in formal educational activities, consistent with a broad definition of nursing education that encompasses both prelicensure training and lifelong professional development. Studies that enrolled mixed cohorts of health professions learners (eg, medical, nursing, and physician assistant students) were eligible if nursing students constituted a defined subgroup and the educational intervention was situated within a nursing-relevant context; (2) phenomenon of interest: use of AI-powered simulation, defined as any simulation modality in which AI algorithms (machine learning, natural language processing, computer vision, GenAI, adaptive systems, or expert systems) actively drive patient behavior, physiological responses, scenario progression, feedback, or debriefing. This includes conversational virtual patients, AI-enhanced VR, adaptive manikins with AI decision engines, and GenAI clinical scenarios; (3) context: any nursing educational setting (university, hospital-based school, simulation center, or online/blended learning environment); (4) outcomes (quantitative): clinical competence, knowledge, skill performance, critical thinking, clinical judgment, self-efficacy, confidence, anxiety, satisfaction, or transfer to practice measured by validated instruments or objective performance scores; (5) outcomes (qualitative): learner or educator experiences, perceptions of realism, usability, acceptability, facilitators, and barriers; (6) study design: randomized controlled trials (RCTs), quasiexperimental studies, cohort studies, mixed methods studies, and qualitative studies (phenomenology, grounded theory, qualitative description, and thematic analysis); (7) article type: original, empirical, or peer-reviewed journal publications; and (8) language: written in English or Chinese.

Exclusion criteria were (1) studies not involving nursing education; (2) simulation using AI only for debriefing analytics without real-time interaction; (3) studies using only rule-based (non-AI) high-fidelity manikins or static virtual patients; (4) conference abstracts, editorials, and protocols; and (5) studies published before January 1, 2014 (to ensure inclusion only of modern AI technologies post–deep learning revolution).

All records were imported into NoteExpress (4.2; Beijing Aegean Software Co, Ltd) software. After duplicate removal, titles and abstracts were independently screened by 2 reviewers (HJ and ZW). Full texts were retrieved and assessed against eligibility criteria by the same 2 reviewers (HJ and ZW), with disagreements resolved by a third senior reviewer (XL).

Data extraction was conducted independently by 2 reviewers (HJ and ZW) using standardized Joanna Briggs Institute (JBI) templates adapted for this review.

The following data were extracted: authors, year of publication, country, study design, sample size, participant characteristics (level of nursing education, age, and gender), detailed description of the AI-powered simulation intervention using the Template for Intervention Description and Replication (TiDier) checklist [32], comparator condition, outcome measures, key findings, and author conclusions.

All extracted data were cross-checked between the 2 reviewers (HJ and ZW), with discrepancies resolved by consensus or consultation with the third reviewer (XL). Study authors were contacted for missing data or clarification where necessary.

Two reviewers (HJ and WS) independently assessed the methodological quality of all included studies. Disagreements were resolved by consensus or arbitration by a third reviewer (XL).

RCTs were assessed using the Cochrane risk of bias 2 tool (RoB 2). Each domain was judged as low risk, having some concerns, or high risk of bias. An overall risk-of-bias judgment was generated for each study.

Nonrandomized studies (quasiexperimental, pre-post, cohort, and interrupted time-series designs) were evaluated using the Risk of Bias in Nonrandomized Studies of Interventions (ROBINS-I) tool [33]. Bias was assessed across 7 domains (confounding, participant selection, classification of interventions, deviations from intended interventions, missing data, measurement of outcomes, and selection of reported results), with overall judgment categorized as low, moderate, serious, or critical risk of bias.

For cross-sectional studies, the 11-item checklist from the Agency for Healthcare Research and Quality (AHRQ) was used, with items rated as “Yes,” “No,” or “Unclear.” Scores of 0‐3, 4‐7, and ≥8 were classified as representing low, moderate, and high quality, respectively.

For qualitative studies, the JBI Critical Appraisal Checklist (10 items) for qualitative research [34] was used to appraise the study’s philosophical perspective, methodology, design, participant representation, data collection, analysis, and interpretation.

For mixed methods studies, the Mixed Methods Appraisal Tool (MMAT) [35] was used to evaluate the appropriateness of the qualitative, quantitative, and mixed methods components, as well as the overall integration of the findings.

The studies included in the analysis showed excessive heterogeneity, making a formal meta-analysis impossible. We adopted a convergent integrated design, integrating quantitative and qualitative findings through joint display tables and narrative synthesis to derive meta-inferences [36].

Figure 1 illustrates the PRISMA flow diagram of study selection process. After removing duplicates, a total of 3181 citations were identified. Following screening of titles and abstracts, 100 full-text articles were assessed for eligibility based on the inclusion criteria. No additional studies were identified through reference list searching. Ultimately, 19 studies were included.

The 19 included studies comprised a total of 1253 participants. Sample sizes ranged from 10 to 247 participants, and publication years ranged from 2020 to 2025. The 19 studies included 3 (16%) RCTs [37-39], 4 (21%) quasiexperimental studies [40-43], 3 (16%) pre-post studies [44-46], 4 (21%) mixed methods studies [47-50], and 1 (5%) cross-sectional study [51], 4 (21%) qualitative studies [52-55]. Most studies were conducted in the United States (n=5, 26%) [44,46,47,50,51], followed by China (n=4, 21%) [39,40,43,45], Singapore (n=3, 16%) [42,48,53], South Korea (n=2, 11%) [41,49], the United Kingdom (n=2, 11%) [52,55], Turkey (n=1, 5%) [37], Canada (n=1, 5%) [54], and one multinational study (n=1, 5%) [38]. Two studies [40,51] enrolled mixed cohorts. Characteristics of the 19 studies are summarized in Table 1.

Three RCTs were included in this systematic review and assessed for quality using the RoB 2 tool. Two of these studies [38,39] were judged to have a low risk of bias, while one [37] was deemed to have some concerns. For the 4 quasiexperimental studies and the 3 pretest-posttest studies, quality assessment was conducted using the ROBINS-I tool. Five of these studies [40,42-44,46] were rated as having a moderate risk of bias, and 2 [41,45] were rated as having a low risk of bias. For the 4 mixed methods studies, the MMAT criteria indicated that 3 studies [48-50] demonstrated high methodological quality, and one study [47] was rated as having moderate methodological quality. For the 4 qualitative studies, the JBI critical appraisal tool was used; 2 [53,54] were assessed as being of moderate quality, and the other 2 [52,55] were of high quality. The one cross-sectional study [51] was evaluated using the AHRQ methodology checklist and was rated as being of moderate quality. Detailed quality assessment results are presented in the Multimedia Appendix 2.

Regarding the types of AI technologies used, 7 [38,39,43,45-47,49] studies applied GenAI or LLMs (eg, ChatGPT [OpenAI]), 5 [37,44,52,53,55] studies used AI-driven virtual patients or mannequins, 5 [40,48,50,51,54] studies focused on AI-enhanced VR or MR simulations, and 2 [41,42] studies incorporated AI chatbots as adjunctive tools. In terms of simulation platforms, 7 [37,46,47,51-53,55] out of 19 (37%) studies used screen-based or web-based simulations, 8 [38-40,45,48-50,54] out of 19 (42%) studies used VR or MR with head-mounted displays, 3 [41,42,44] out of 19 (16%) studies used high-fidelity mannequins, and 1 [43] out of 19 (5%) studies integrated AI as a supplement to traditional simulation methods. Concerning control group designs, 8 [37-39,41-44,54] out of 19 (42%) studies included a parallel control group, comparing AI interventions with traditional teaching, standard simulation, or alternative technologies, while 11 [40,45-53,55] out of 19 (58%) studies had no control group, using pretest-posttest designs or qualitative feasibility assessments.

A notable pattern across the included studies was the predominance of statistically significant positive findings: 8 [38,39,43-46,48,50] out of 11 (73%) studies reporting knowledge or skill outcomes and 8 [39,40,43-45,48,49,51] out of 9 (89%) studies reporting self-efficacy or attitude outcomes demonstrated significant improvements. Due to substantial clinical and methodological heterogeneity, formal statistical assessments of publication bias were not feasible. The near-uniform positive pattern, combined with the predominance of small-sample, uncontrolled feasibility studies, suggests a high likelihood of unpublished null or negative results. This potential publication bias weakens the interpretability of all validity conclusions in this review.

This mixed methods systematic review synthesizes empirical evidence from 19 studies (n=1253) on AI-driven simulation in nursing education. The principal finding is that AI-enhanced simulations show short-term promise for improving foundational knowledge, clinical reasoning, and communication confidence, particularly when delivering real-time feedback and enabling repeated practice. However, the evidence base is heavily weighted toward uncontrolled or quasiexperimental designs, with only 3 RCTs available. Qualitative data consistently highlight learner appreciation for psychologically safe, repeatable practice environments, alongside persistent concerns regarding technical instability, unnatural interactions, and a learner-perceived “authenticity gap.”

Quantitative findings indicate that AI simulations exert a positive impact on cognitive and affective learning outcomes, particularly in the domains of communication confidence and knowledge acquisition. This aligns with the broader perspective in medical education literature, which posits that the primary advantage of AI lies in its capacity to provide sustained, on-demand deliberate practice, overcoming the limitations associated with traditional high-fidelity manikins or standardized patients. Notably, AI simulations that offer real-time feedback, adaptive scenario progression, and opportunities for repeated practice are associated with stronger learning outcomes, consistent with the principles of deliberate practice and experiential learning [56]. Furthermore, AI simulations foster psychological safety among students, allowing them to make mistakes without fear of judgment from peers or instructors. This factor emerges as a dominant explanation for the significant reductions in anxiety and enhancements in self-efficacy reported across multiple studies. These findings are congruent with prior research indicating that students are more likely to engage boldly in nursing practice training within environments characterized by low error costs and greater inclusivity [57-59].

However, the evidence is not entirely consistent and positive. One study found that AI-assisted simulation was inferior to standardized patient simulation in terms of BSE skills, and 2 other studies reported no significant improvement in critical thinking or knowledge. These discrepancies may stem from differences in simulation fidelity, the sensitivity of outcome measures, or the degree of curriculum integration. Importantly, AI simulation may demonstrate differential effectiveness across various types of learning objectives. Existing evidence suggests that AI simulation shows significant advantages in achieving highly structured learning objectives, such as communication skills and procedural knowledge, while facing notable limitations in objectives requiring high levels of emotional empathy, cultural sensitivity, and complex interpersonal interaction. Kotlyar and Krasman [60] indicate that learners’ perceptions of the feedback source directly influence its acceptance and learning outcomes. In learning tasks intensive in emotion and interpersonal interaction, learners tend to trust human feedback sources perceived as possessing “benevolence” and “integrity” over AI systems perceived as merely “competent.” Furthermore, a network meta-analysis in nursing education revealed that while AI simulation was highly effective in enhancing knowledge acquisition (standardized mean difference=1.11), its effect on cultivating practical skills was relatively limited [61]. Research in the field of social-emotional learning similarly found that AI underperformed compared to human educators in achieving deeper instructional goals, such as facilitating guided reflection and the internalization of social-emotional knowledge [62]. Therefore, educators should select appropriate teaching strategies based on the learning objectives. For highly structured tasks such as foundational communication skills training, history-taking, and health education, AI simulation can provide standardized, repeatable practice opportunities. Conversely, for scenarios requiring emotional empathy, cultural sensitivity, and complex clinical judgment, traditional standardized patient simulation and clinical practicums remain irreplaceable. Moreover, it is essential to gradually enhance the naturalness of interaction and emotional expression while ensuring technical stability and usability, avoiding the sacrifice of the learning experience in pursuit of excessive technical complexity [63].

This finding corroborates the views expressed in some qualitative studies, namely that students perceive interactions with AI as lacking emotional depth, nonverbal communication, and the dimension of physical examination. This observation aligns with the results of a recent qualitative study involving nursing graduate students. Research by Jiang et al [64] found that while nursing students held positive attitudes toward GenAI, they commonly raised concerns regarding content accuracy and technical shortcomings. We term this phenomenon the “authenticity gap.” In this review, the “authenticity gap” is explicitly defined as a learner-perceived discrepancy between AI-driven interactions and human clinical encounters, encompassing deficits in emotional resonance, nonverbal cue recognition, and tactile examination dimensions. It emerges primarily from qualitative usability and experiential data rather than direct performance comparisons. Although AI demonstrates proficiency in handling diagnostic logic and conversational flow, it currently falls short in replicating the tactile and empathetic complexities inherent in human care. Furthermore, the study found that technical malfunctions, such as delays in speech recognition, can disrupt the interaction flow and potentially increase the extraneous cognitive load and state anxiety of some learners. This finding is highly consistent with the conclusions of a recent scoping review published by Chan et al [65]. That review identified barriers to the implementation of AI simulations, including system instability, inaccurate speech recognition, and a lack of structured training for instructors. These shared findings further substantiate that the implementation of AI simulation in education is not merely a matter of technological introduction but involves a systemic transformation encompassing pedagogical philosophy, organizational support, and ethical governance.

Numerous previous studies have indicated that learner acceptance is a critical determinant for the successful application of AI technology in the educational field. The quantitative findings of this review demonstrate that across the included studies, usability, acceptability, and perceived usefulness were consistently rated at moderate to high levels, with overall SUS scores falling within the acceptable range. Dimensions of the TAM further corroborated that students recognized the value of AI simulations for learning, particularly in terms of accessibility and structured guidance. Qualitative evidence enriched these findings, untangling the nuanced ways in which learners experience AI simulations. Students valued the safe, low-risk environment where they could attempt tasks repeatedly without fear of judgment. They also appreciated the consistency and repeatability of AI interactions, which helped bridge the gap between theory and practice. However, persistent concerns were noted regarding the unnaturalness of AI dialogues, the absence of nonverbal cues, and technical instability. These issues highlight the necessity for ongoing improvements in natural language processing and affective expression modeling to enhance the authenticity of AI-driven simulations.

Our review findings highlight the potential of AI simulation in nursing education. In the long term, with the advancement of AI technology, its integration into the daily practice of nursing education is inevitable. However, within the current technological landscape, several barriers to successful implementation remain. To facilitate the integration of AI simulation, based on our review results, we offer several recommendations for nursing educators and instructional designers.

The primary strength of this review lies in its mixed methods convergent segregated design. By integrating objective performance data with qualitative learner experiences, we were able to assess not only whether AI simulations work but also how and why they succeed or fail in specific contexts. Furthermore, the inclusion of contemporary GenAI technologies (eg, ChatGPT-driven virtual patients) ensures the findings are highly relevant to the current technological landscape.

Nevertheless, several limitations must be acknowledged. First, the included studies exhibited considerable heterogeneity in terms of AI modalities (screen-based vs VR vs chatbots), intervention durations, and measurement instruments, which precluded a quantitative meta-analysis and limited the ability to determine standardized effect sizes. Second, a significant proportion of the studies relied on pre-post designs lacking active control groups, increasing the risk of bias. Third, all included studies assessed outcomes immediately or within a short postintervention window; no study examined skill retention beyond the immediate educational context or transfer to clinical practice. This constitutes a critical evidence gap. Fourth, a small number of studies included nonnursing learners, which may limit the applicability of some findings to prelicensure nursing education. Fifth, the near-uniform positive result pattern raises concern about publication bias, which cannot be formally tested given study heterogeneity but should be considered when interpreting the overall evidence picture. Sixth, several quantitative studies had small samples, limiting statistical power and precision of effect estimates. Finally, most studies were conducted in high-income countries or regions with robust technological infrastructure; no studies originated from Africa, South America, or low-income countries, and the concentration of studies from Greater China may reflect regional research networks rather than globally representative sampling.

This systematic review indicates that AI-driven simulations show promise for improving nursing students’ cognitive knowledge, clinical reasoning skills, and communication confidence, particularly by providing a safe, repeatable practice environment. Evidence from RCTs and controlled quasiexperimental studies supports these benefits, though the predominance of uncontrolled designs and the absence of longitudinal data on skill retention and clinical transfer warrant caution in interpreting strength of effectiveness claims. Current evidence suggests AI is most effective for highly structured learning objectives (foundational communication and history-taking) and less effective for advanced psychomotor skills and emotionally complex interpersonal interactions.

Although these positive outcomes demonstrate potential for modernizing nursing instruction, the reported negative technical experiences—captured in the “authenticity gap”—necessitate greater consideration and incorporation of input from both students and educators during the design, implementation, and evaluation phases of AI technologies. The evidence supports positioning AI as a complementary tool within a stepped simulation continuum, alongside traditional methods, rather than as a replacement for standardized patients or clinical placements.

The findings of this review should serve as a starting point for more precise research into AI-enhanced simulation, using standardized measures of clinical competence to obtain a more comprehensive overview. This can be achieved through longitudinal and multicenter RCTs, implementation strategies addressing technical barrier management, further research on integrating automated AI feedback with human-facilitated debriefing, and broader resource evaluation (eg, via cost-effectiveness studies of AI scenario development). By addressing these gaps, the potential of AI-driven simulation in nursing education can be fully realized, improving students’ clinical preparedness and ultimately enhancing patient care outcomes.