Is it time to introduce an artificial intelligence curriculum in undergraduate medical education? medical students’ perspectives: a systematic review and meta-analysis

Abstract

Background

Although artificial intelligence (AI) is increasingly transforming healthcare systems, its systematic integration into undergraduate medical education (UME) remains limited. Despite widespread recognition of AI’s potential to enhance clinical practice, AI-related competencies are still inadequately embedded in most medical curricula, even though medical students generally express positive attitudes toward AI.

Objectives

This systematic review and meta-analysis synthesizes the current literature on medical students’ attitudes toward teaching about AI and clarifies the challenges, opportunities, and potential approaches for incorporating AI into the UME curriculum. These findings underscore the need for competency-based AI integration in UME worldwide.

Methods

Following Preferred Reporting Items for Systematic reviews and Meta-Analyses (PRISMA) guidelines, a systematic literature review was conducted using the PubMed, Web of Science, Scopus, and Cochrane Library databases between January 2000 and February 2025, revealing 26 studies (n = 20,963 students) assessing attitudes and challenges related to learning AI. Meta-analysis calculated the pooled proportions of student support. A random-effects model with Freeman-Tukey transformation was used; heterogeneity and bias were assessed via standard statistical methods.

Results

The systematic review included 26 studies (n = 20,963 medical students). Meta-analysis revealed that 16,278 participants (78.0%) held positive attitudes towards AI curriculum integration. Despite this strong support, a significant ‘optimism-competence gap’ was identified: while 17,420 participants (83.1%) agreed on the necessity for AI training, only 7,630 participants (36.4%) felt confident in applying AI in clinical practice. Subgroup analyses confirmed consistent support across geographic regions. The analysis indicated extreme heterogeneity (I² = 98.5%). The primary implementation challenges, derived from the systematic review, were curricular overcrowding (reported in 13 out of 19 studies, 68%), lack of faculty expertise (in 12 out of 23 studies, 52%), and ethical concerns (in 9 out of 22 studies, 41%).

Conclusion

A standardized AI curriculum is urgently needed to bridge the optimism-competence gap. Prioritizing faculty training, interdisciplinary collaboration, and ethical frameworks will ensure that future physicians can harness AI’s potential while mitigating risks. Our findings, particularly the quantified optimism-competence gap and the influence of unmeasured institutional factors, provide a new evidence base for developing these prioritized interventions.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12909-026-09630-9.

Background

Digital tools and learning platforms have transformed education, and artificial intelligence (AI) is at the forefront of this revolution [1, 2]. With roots tracing back to the late 1950s via artificial neural networks [3], AI has evolved into a field where machine learning (ML) and deep learning (DL) empower systems to perform tasks that mimic human learning and problem-solving [4, 5]. As learners across all educational levels increasingly interact with digital experiences [6], the potential of AI to enhance clinical practice becomes ever more apparent.

Despite its transformative promise, the integration of AI into medical education remains limited. Repeated calls for reform—highlighted in numerous national reports and driven by competency-based imperatives [7, 8]—underscore the need for curricula beyond traditional biomedical teaching. Medical education faces the dual challenge of integrating advanced AI technologies without succumbing to “instrumentalism,” a tendency to reduce learning to algorithmic efficiency at the expense of critical thinking and humanistic values [9]. As AI is poised to play a central role in healthcare [10], establishing a robust framework for its inclusion in UME is critical. For this review, an “AI curriculum” is defined as any structured educational intervention that develops knowledge or skills in artificial intelligence. This includes dedicated courses, modules, workshops, or AI content embedded within existing subjects (e.g., radiology, ethics, clinical reasoning). Beyond theoretical knowledge, such curricula should cultivate practical and ethical competencies, including understanding key AI models, critically evaluating AI outputs, and navigating the ethical challenges of AI in clinical practice.

Current advances in AI have significantly influenced diagnostic and clinical decision-making [11, 12], yet the adoption of AI training in UME is hindered by curricular overcrowding and a lack of systematic evidence. While AI can potentially aid in diagnosing, prescribing, and managing patient care, its full benefits will be realized if future physicians are equipped with theoretical and practical competencies [13–15]. Radical, rather than incremental, changes in medical education are necessary to bridge this “optimism–competence gap” [16], a need that has even been recognized as a government health policy priority [17]. Although several studies report that medical students support AI integration [18, 19], their perceptions about the effectiveness and practical applicability of AI curricula remain underexplored.

In the Iranian national core medical curriculum, explicit competencies related to artificial intelligence and digital health are not yet formally mandated across all medical schools. However, recent initiatives — such as development of AI-focused course syllabi and elective offerings in medical universities — indicate growing recognition of AI’s relevance in medical practice and education. These developments reflect a nascent stage compared to some international frameworks that increasingly embed AI literacy and data-driven competencies as structured components of UME.

This systematic review and meta-analysis aim to comprehensively synthesize the current literature on medical students’ perspectives regarding artificial intelligence (AI) training in UME. Specifically, the study seeks to: (I) summarize existing evidence on medical students’ attitudes, perceptions, and self-reported readiness toward AI in UME; (II) identify key patterns, sources of variability, research gaps, and challenges associated with integrating AI into medical curricula; and (III) highlight implications and propose future scholarly priorities to guide the development and effective implementation of AI curricula in medical education. To our knowledge, this is the first systematic review and meta-analysis to consolidate global evidence on medical students’ perspectives regarding the integration of AI training into UME.

It is important to recognize that artificial intelligence (AI) is an umbrella term encompassing a diverse range of technologies. In this review, AI is defined broadly as computer systems designed to perform tasks that typically require human intelligence. This includes a spectrum of applications, ranging from machine learning (ML) models used in diagnostic imaging and clinical decision support systems to natural language processing (NLP) tools for medical documentation and, more recently, generative AI applications capable of simulating patient interactions and synthesizing medical knowledge.

A critical challenge in synthesizing the literature arises from the fact that many studies in UME conceptualize AI as a broad and undifferentiated construct, rather than distinguishing among specific technologies and their distinct clinical or educational applications. As a result, student perceptions are frequently reported in relation to “AI” as a monolithic concept, limiting the granularity of available evidence.

Accordingly, the findings of this review should be interpreted within the context of this generalized framing. At the same time, the review underscores the need for future research and curriculum development to move beyond this broad conceptualization toward a more nuanced examination of discrete AI technologies, domain-specific competencies, and their practical implications for clinical practice and medical education.

Methods

This study follows the PICO framework to structure and guide the search process. We included studies that addressed relevant terms from all four categories within their titles or abstracts:

• Population (P): Undergraduate medical students and educational institutions.

• Intervention (I): Integration of AI curricula in UME.

• Comparison (C): Current medical curricula lacking a formal AI component or those using traditional medical education methods.

• Outcome (O): Perspectives or attitudes of medical students toward AI education, the perceived benefits and challenges of incorporating AI into the curriculum, and recommendations for developing AI curricula in medical education.

Literature search and selection

A comprehensive and systematic search strategy was implemented across PubMed, Web of Science, Scopus, and the Cochrane Library to identify eligible studies published between January 2000 and 15 February 2026. The search strategy combined MeSH terms and free-text keywords related to “Artificial Intelligence,” “Curriculum,” “Undergraduate Medical Education,” and “Attitude.” Searches were restricted to English-language, peer-reviewed studies. Detailed search strings for each database are provided in Supplementary Material: Appendix 1.

All retrieved records were imported into Covidence (Veritas Health Innovation, Melbourne, Australia) for duplicate removal and systematic screening. The platform enabled blinded, independent assessment by reviewers and facilitated conflict resolution through consensus. Additionally, Elicit (elicit.org), an AI-assisted evidence synthesis tool, was employed to verify the comprehensiveness of the search strategy and identify potentially overlooked studies across databases.

Studies were included if they examined the integration or teaching of artificial intelligence (AI) within UME and assessed medical students’ perceptions, attitudes, or self-reported readiness regarding AI-related curricula. Studies were excluded if they focused solely on AI as a clinical or diagnostic tool without an educational component, were published in languages other than English, fell outside the predefined date range, or did not report disaggregated data specific to undergraduate medical students. The requirement for disaggregated undergraduate data was applied to ensure that pooled estimates accurately reflected this target population and to prevent misclassification or dilution of findings arising from mixed learner groups (e.g., residents, practicing physicians, or other health professions students). Although studies without disaggregated undergraduate data may provide valuable contextual insights, their inclusion would have precluded valid quantitative synthesis for this population and were therefore not considered suitable for meta-analysis.

Data extraction, quality assessment, and synthesis

The review process was managed using Covidence, which supported blinded screening, conflict resolution, and study tracking, and Elicit, which facilitated structured data extraction, preliminary synthesis, and verification of extracted variables. Two independent reviewers screened all titles and abstracts for eligibility, with discrepancies resolved through discussion or adjudication by a third reviewer. Full-text articles of potentially eligible studies were retrieved for detailed evaluation.

Data extraction was performed using a standardized, pre-piloted form developed in accordance with Preferred Reporting Items for Systematic reviews and Meta-Analyses (PRISMA) 2020 and JBI guidelines. Extracted variables included study design, participant demographics, sample size, educational interventions or AI-related exposures, comparator details (if applicable), outcomes, key findings, and reported limitations. Inter-rater reliability for study selection was assessed using Cohen’s kappa coefficient (κ), and all extracted data were independently cross-verified to ensure completeness and accuracy.

The methodological quality of included studies was appraised using the AXIS tool [20], which evaluates cross-sectional studies across domains such as research design, sampling strategy, measurement validity, statistical analysis, and reporting standards. Two reviewers independently assessed each study, resolving disagreements by consensus or, when required, through third-party adjudication [21]. Quality scores derived from predefined criteria were incorporated into meta-regression analyses to examine the influence of study quality on overall effect estimates.

Given the heterogeneity of study designs, populations, and outcome measures, a narrative synthesis was undertaken to integrate findings, identify consistent patterns, and highlight evidence gaps. Tabular and graphical summaries were used where appropriate to enhance interpretability. The synthesis adhered to established systematic review frameworks to ensure transparency, reproducibility, and methodological rigor in the interpretation of findings.

Statistical analysis

The primary objective of this meta-analysis was to estimate the overall proportion of medical students demonstrating a positive attitude toward integrating AI curricula in UME. For each study, proportions were calculated by dividing the number of students expressing favorable attitudes by the total sample size. To stabilize variance and appropriately account for proportions approaching 0 or 1, the Freeman–Tukey double arcsine transformation was applied. Both random-effects (DerSimonian–Laird estimator) and common-effect models were implemented, given the anticipated heterogeneity across studies. Pooled results were subsequently back-transformed to the original proportion scale to facilitate interpretation and comparability.

Between-study heterogeneity was quantified using the I² statistic (with values > 50% indicating substantial heterogeneity) and Cochran’s Q test (p < 0.05 indicating significant heterogeneity). Additional metrics included the estimation of between-study variance (τ²) and prediction intervals to assess expected variability in future studies. Publication bias was evaluated visually through funnel plot asymmetry and statistically using Egger’s regression test and Begg’s rank correlation test. When evidence of bias was detected, the Trim-and-Fill method was applied to adjust pooled estimates.

To explore potential sources of heterogeneity, subgroup analyses were conducted according to geographic region (North America, Asia, Middle East, and others) and questionnaire type (researcher-developed vs. standardized instruments). Meta-regression analyses were performed to examine the influence of study quality (AXIS score), publication year, sample size, and participants’ baseline AI knowledge on pooled effect estimates. Variables demonstrating significant univariable associations were subsequently included in a multivariable meta-regression model. Sensitivity analyses were conducted by sequentially excluding individual studies to evaluate the robustness and stability of the pooled estimates. In addition, a cumulative meta-analysis arranged chronologically was undertaken to explore temporal trends and the evolution of study outcomes over time.

All statistical analyses were performed using R software (version 4.3.2), employing the meta package for primary meta-analyses and the metafor package for cumulative meta-analysis and back-transformations. A two-tailed p-value < 0.05 was considered statistically significant for all tests.

Reporting

This systematic review and meta-analysis were conducted and reported in accordance with the PRISMA guidelines [22]. The study selection process, including the number of records identified, screened, and included, is illustrated in the PRISMA flow diagram (Fig. 1). Potential limitations—such as publication bias, heterogeneity, and the exclusion of non-English studies—are explicitly acknowledged to ensure transparency and reproducibility.

Fig. 1.

Adapted PRISMA flow diagram illustrating the study selection process. (Extracted from Covidence)

Having established a rigorous and reproducible process consistent with PRISMA standards, the subsequent section presents the aggregated quantitative findings derived from the meta-analysis.

Results

A comprehensive literature search across Scopus (n = 11,955), Web of Science (n = 9,757), PubMed (n = 4,258), Cochrane Library (n = 204), and 16,962 records imported without identifiable source metadata (originating from these databases) identified a total of 43,136 records.

Following duplicate removal, 185 full-text articles were assessed for eligibility. The primary reasons for exclusion at the full-text stage were non-educational focus on AI (n = 74), absence of medical student perspectives (n = 52), and insufficient quantitative data for meta-analysis (n = 33). Twenty-eight studies initially met the inclusion criteria; two were subsequently excluded due to lack of quantitative outcomes, yielding 26 studies in the final synthesis. No studies were excluded due to retrieval failure.

All included studies employed observational designs, encompassing a combined sample of 20,963 medical students, of whom 16,278 (77.6%) reported positive attitudes toward AI curricular integration. Data were primarily collected through structured questionnaires or online surveys, and demographic characteristics—including age and academic level—were broadly comparable across studies.

The rigorous selection process, incorporating dual independent review and systematic verification in Covidence, ensured that only studies meeting predefined quality and relevance criteria were included. This approach enhanced the transparency, reproducibility, and methodological rigor of the review, enabling a focused synthesis of evidence on undergraduate medical students’ attitudes toward AI and its integration into medical curricula. The study selection process is summarized in the PRISMA flow diagram (Fig. 1).

Descriptive characteristics of included studies

The 26 included studies were conducted across 18 countries/regions. Saudi Arabia (n = 2), Germany (n = 2), Canada (n = 2), and the United States (n = 2) contributed multiple studies. Single studies originated from the United Kingdom, Jordan, Ireland, Kuwait, Turkey, India, Oman, Nepal, Malaysia, Australia, Syria, and Palestine. Five studies were multinational: one covered 63 countries [23], one included nine Arab countries [24], one surveyed 48 countries, and two examined Germany, Austria, and Switzerland. According to the classification in the “World Economic Situation and Prospects” by Belsey (2022) [25], the studies included in this systematic review and meta-analysis were conducted in both High-Income Countries (HICs) and Low- and Middle-Income Countries (LMICs). Specifically, 15 studies were conducted in HICs [26–40], while other studies were conducted in LMICs [23, 24, 41–49]. This distribution highlights the global interest in understanding medical students’ attitudes toward AI in medical education across diverse economic contexts.

Most studies (n = 24) employed a cross-sectional survey design [23–48, 50, 51]. Additionally, mixed-methods studies were conducted in four studies [31, 32, 41, 49], and one study adopted a prospective design (11). Sample sizes ranged from 58 to 4,492 participants, for a total of 20,963 across all studies.

Publications spanned 2019–2025, peaking in 2022 (n = 13), indicating growing scholarly attention to AI in medical education. Earlier years had limited output (2019–2021: 4), while 2023–2024 showed a resurgence (n = 11), reflecting evolving research priorities.

Table 1 provides a summary of key study characteristics.

Table 1.

Characteristics of included studies

Author	Year	Country	Study design	Study setting	Sample size	Sampling method	Males	Females	Mean age	Type of research instrument	Validity and reliability of the tool
Cherry Sit et al. [33]	2020	United Kingdom (UK)	Cross-sectional	19 UK medical schools	484	Non-probability sampling	-	-	-	Researcher-made/ Online survey	Validated with 30 medical students before final use
Zaboor Ahmed et al. [43]	2022	Pakistan	Cross-sectional	Multiple medical colleges and hospitals in Pakistan	247	Convenient sampling	123 (49.8%)	124 (50.2%)	10–20 years: 60 (12.8%), 21–30 years: 386 (82.1%), 31–40 years: 20 (4.3%), 41–50 years: 20 (4%), 51–60 years: 20 (4%)	Researcher-made/ Online survey	Validated by a senior faculty member and pilot-tested
Ali Meshari Ali A [27]	2022	Saudi Arabia	Cross-sectional online survey	10 prestigious medical schools in Saudi Arabia	263	Convenient sampling	97 (36.9%)	166 (63.1%)	Median age = 23 years (range: 19–58 years)	Researcher-made/ Online survey	Not explicitly mentioned
Mouna M. Al Saad [44]	2022	Jordan	Cross-sectional	Six universities in Jordan	900	Convenient sampling	470 (52.2%)	430 (47.8%)	21.34 years (± 2.43)	Researcher-made/ Online survey	Validated by the University of Jordan IRB
Fahad Abdullah Saeed Alwadani [26]	2024	Saudi Arabia	Cross-sectional	King Faisal University & Imam Abdulrahman Bin Faisal University	303	Non-probability Convenient sampling	174 (57.42%)	129 (42.58%)	21–25 years: 284 (93.73%), 26–30 years: 17 (5.61%), > 31 years: 2 (0.66%)	Researcher-made/ Online survey	A pilot study was conducted with 30 students to confirm the reliability and validity.
Sotirios Bisdas [23]	2021	Multinational (63 countries)	Cross-sectional survey	Not mentioned	3133	Convenient sampling	1,050 (33.5%)	2,083 (66.5%)	21.95 ± 2.77 years	Researcher-made/ Online survey	Not explicitly stated
Charlotte Blease [34]	2022	Ireland	Cross-sectional survey	Four medical schools in Ireland (Cork, Galway, Belfast, and Dublin)	252	Convenient sampling	37.4% (94)	62.6% (157)	Majority born in 1992 or later (90.7%)	Researcher-made/ Online survey	The survey was developed in consultation with physicians and piloted with medical students and physicians for face validity.
Ali Jasem Buabbas [45]	2023	Kuwait	Cross-sectional survey	Faculty of Medicine at Kuwait University	352	Convenient sampling	39 (11.1%)	313 (88.9%)	22.1 years (± 1.8 SD)	Researcher-made/ Online survey	Construct validity was tested using factor analysis. Reliability = (intraclass correlation value = 0.984). Cronbach’s Alpha = 0.755;
Civaner MM [35]	2022	Turkey	Cross-sectional	Medical schools in Turkey	3018	Convenient sampling	-	-	-	Researcher-made/ Online survey	Cronbach’s alpha > 0.7, Kaiser-Meyer-Olkin Measure of Sampling Adequacy = 0.869
D. Pinto dos Santos [29]	2019	Germany	Prospective, cross-sectional, multicenter survey	Three major German universities (Cologne, Bonn, Mainz)	263	Convenient sampling	163 (62%)	100 (38%)	Median = 23 years (IQR: 21–26)	Researcher-made/ Online survey	Not mentioned
Ahmed Hafez Allam [24]	2024	Multinational (Libya, Egypt, Iraq, Jordan, Syria, Sudan, Algeria, Palestine, Yemen)	Cross-sectional survey	Nine medical schools in Arab countries	4,492	Convenient sampling	1,724 (38.4%)	2,768 (61.6%)	-	Researcher-made/ Online survey	Expert-reviewed for relevance, comprehensiveness, and clarity. Pilot-tested with 265 responses. Cronbach’s alpha: 0.81
Hamza Ejaz [41]	2022	48 countries were represented in the study	Mixed-methods study	International Federation of Medical Students Associations (IFMSA) General Assembly and online	128	Convenient sampling/ Purposive sampling	52 (41%)	72 (56%)	-	Researcher-made/ Online survey	The questionnaire was reviewed by a team of medical students, qualitative research experts, an AI expert, and a clinician-scientist with expertise in AI research
Sebastian Fitzek [40]	2024	Austria, Germany, and Switzerland	Cross-sectional survey	-	409	Convenient sampling	49.60%	45.50%	23.13	Survey instrument adapted from AI in Healthcare Education Questionnaire (AIHEQ) and Medical Student Attitudes Toward Artificial Intelligence in Medicine (MSATAIM) scale	Content validity was assessed by a panel of experts. Cronbach’s alpha (α = 0.807
Zinah A. Al Hadithy [36]	2023	Oman	Cross-sectional	Sultan Qaboos University	228	Random sampling	45.70%	54.3%	-	A web-based validated exploratory questionnaire adapted from a study conducted at the University of Toronto	The questionnaire was validated through a pilot test.
Preetha Jackson [46]	2024	India	Cross-sectional study	Pushpagiri Institute of Medical Sciences and Research Center	325	Convenient sampling	77 (24%)	248 (76%)	Mean (SD) = 21.4 (1.9)	A semi-structured questionnaire was adopted from a Turkish study by Civaner et al.	A pilot study was undertaken by administering the questionnaire to 20 medical students.
Kamel Jebreen [49]	2024	Palestine	mixed-method cross-sectional design	West Bank and Gaza Strip, Palestine	349	Non-probability Convenient sampling	98 (28.08%)	251 (71.92%)	20.38 ± 1.97	The questionnaire dimensions and questions were developed from the survey designed by Pucchio et al.	The Cronbach’s alpha coefficient was found to be 0.891, indicating high internal consistency.
Nisha Jha [47]	2022	Nepal	Cross-sectional study	KIST Medical College and Teaching Hospital, Nepal	216	census sampling	91 (42.1%)	125 (57.9%)	-	A questionnaire used in a Canadian study was used	Cronbach alpha = 0.7)
Joachim Kimmerle [28]	2023	Germany	Cross-sectional	Eberhard Karls University, Tübingen, Germany	58	Convenient sampling	23	35	24.51 (3.56)	Researcher-made/ Online survey	Not mentioned
Dav Shalom Liu [32]	2022	United States	Mixed-methods survey	Nationwide survey of US medical students	390	Convenient sampling	-	-	26 (3)	Researcher-made/ Online survey	Not mentioned
Aan Pucchio [31]	2022	Canada	Cross-sectional mixed-methods study	All 17 Canadian medical schools	475	Convenient sampling/ Purposive sampling	166 (34.9%)	300 (63.2%)	25.34	Researcher-made/ Online survey	The inter-coder reliability was 82%
Jonathon Stewart [37]	2023	Australia	Cross-sectional	All medical schools in Western Australia	134	Census sampling	72 (54%)	59 (44%)	The majority of students (82.0%) were 20–29 years old	Researcher-made/ Online survey	The draft survey was piloted with a small group of medical students
Sarya Swed [42]	2022	Syria	Cross-sectional	Syrian medical schools	1,484	Convenient sampling	779 (52.0%)	779 (52.0%)	(97.2%) were between 21 and 30 years old	Not mentioned	Cronbach’s alpha values were: Attitude = 0.663
Minnie Teng [30]	2022	Canada	Cross-sectional	18 universities across Canada	2167	Snowball and Convenient sampling methods	805 (37.2%)	1355 (62.8%)	< 21 years: 220 (10.15%), 21–25 years: 1217 (56.16%), 26–30 years: 492 (22.7%), 31–35 years: 136 (6.28%), 36–40 years: 71 (3.28%), 41–45 years: 20 (0.92%), 46–50 years: 4 (0.18%), ≥ 50 years: 7 (0.32%)	Researcher-made/ Online survey Not Mentioned	The survey was piloted by a small group from the same university. Questions were revised for clarity according to feedback from the pilot group.
Alvin Yong Zong Tung [48]	2023	Malaysia	Cross-sectional study	17 universities in Malaysia	301	Convenient sampling	120 (39.9%)	181 (60.1%)	-	Medical Artificial Intelligence Readiness Scale for Medical Students (MAIRS-MS)	MAIRS-MS has a Cronbach’s alpha of 0.87
Lukas Weidener [51]	2024	Germany, Austria, and Switzerland	Cross-sectional study	Universities in Germany, Austria, and Switzerland	487	Convenient sampling	217 (44.6%)	270 (55.4%)	< 20 years: 56 (11.5%) 20–25 years: 301 (61.8%) 26–30 years: 92 (18.9%) 31–35 years: 28 (5.7%) > 35 years: 10 (2.0%)	Researcher-made/ Online survey	Not mentioned
Elena A. Wood [39]	2021	USA	Cross-sectional study	Medical College of Georgia at Augusta University, USA	117	Convenient sampling	64 (55%)	53 (45%)	24 or younger	Researcher-made/ Online survey	Not mentioned

Gender data are reported as presented in the original studies. Where additional categories (e.g., non-binary, unspecified, or prefer not to answer) were reported, only male and female counts are displayed in this table for consistency across studies. Therefore, in some cases, the sum of male and female participants may not equal the total sample size

Medical students’ attitudes toward integration of AI Curricula in UME

The integration of AI into medical education is increasingly recognized as essential for equipping trainees with the skills needed to navigate the evolving healthcare landscape. A growing body of research has explored medical students’ attitudes toward AI, including their perceptions of its potential benefits, anticipated challenges in implementation, and recommendations for effective curriculum integration. These studies provide critical insights into learners’ readiness, highlight barriers such as limited faculty expertise and curriculum overcrowding, and inform strategies for the design of AI-focused educational interventions.

Table 2 provides a detailed summary of the characteristics, objectives, and key findings of all included studies.

Table 2.

Summary of study characteristics according to the study objectives of all included studies

Author	Attitudes and perceptions of medical students toward AI in medical education	Challenges in implementation	Recommendations for integrating AI in medical curriculums
Cherry Sit et al. [33]	378 out of 484 of participants (78%) agreed that students should receive AI training as a component of their medical degree.	No compulsory AI curriculum in UK medical schools.	Expand and improve AI training in medical schools.
Zaboor Ahmed et al. [43]	237 out of 470 individuals (76.7%) supported the inclusion of AI in the curriculum.	Students feel inadequately prepared to work with AI.	Present realistic use cases and limitations of AI to students.
Ali Meshari Ali A [27]	71% percent of respondents believed AI should be included in medical education.	Lack of awareness and knowledge about AI.	Inclusion of AI in the medical curriculum.
Mouna M. Al Saad [44]	616 out of 900 students (68.4%) believed AI teaching should be mandatory in medical education.	Lack of financial resources and technological advancements.	Allocation of resources for AI training and implementation.
Fahad Abdullah Saeed Alwadani [26]	264 out of 303 participants (87.1%) believe AI will play an essential role in healthcare	Lack of awareness and understanding of AI technology	-
Sotirios Bisdas [23]	2683 out of 3133 participants (85.6%) agree that AI should be part of medical training.	-	Integrate AI into undergraduate medical curricula.
Charlotte Blease [34]	78.6% of participants agreed that discussion about AI/ML should be part of medical training.	Lack of AI courses in medical curricula	Provide AI-specific training and workshops
Ali Jasem Buabbas [45]	289 out of 352 respondents (82.1%) believed medical students should receive AI teaching/training and incorporate AI teaching/training into the medical curriculum.	Faculty shortages in AI education	Develop standardized AI competency frameworks.
Civaner MM [35]	93.8% of participants were in favor of structured training on AI applications that should be given during medical education.	-	Organize open lectures to showcase outstanding digital health projects and stimulate interest.
D. Pinto dos Santos [29]	187 out of 263 participants (71%) supported integrating AI into medical training.	Difficulty in integrating digital health topics into an already packed medical curriculum.	Incorporate AI topics into medical training, focusing on practical applications.
Ahmed Hafez Allam [24]	82.1% of respondents believed medical students should receive AI teaching/training and incorporate AI teaching/training into the medical curriculum.	Lack of formal instruction on AI/ML in medical curricula.	Integrate AI/ML topics into medical ethics courses, focusing on patient privacy and digital divides.
Hamza Ejaz [41]	92% of respondents stressed that AI-related teaching needs to be incorporated into the core medical curriculum.	43% of participants reported feeling worried about the impact of AI; 31% thought AI will replace physicians.	Ensure that medical students are equipped to appraise new technologies critically.
Sebastian Fitzek [40]	58.6% of participants agreed that AI should be incorporated into the medical curriculum.	-	Offer introductory courses on AI principles, applications, and ethical considerations.
Zinah A. Al Hadithy [36]	Students were generally receptive to incorporating AI into medical school curricula, as 174 out of 221 participants (78.7%) believed every medical trainee should receive training on AI competencies.	Lack of systematic evidence for incorporating AI into medical education.	Use AI-based simulators and virtual patient platforms for clinical training.
Preetha Jackson [46]	243 out of 325 participants (74.8%) were in favor of structured training on AI applications during medical education.	Potential oversimplified understanding of AI among students.	-
Kamel Jebreen [49]	Participants stated that AI should be specifically included in their medical curriculum (247 out of 349, 70.8%).	Ethical and social concerns regarding AI use in healthcare.	-
Nisha Jha [47]	Students stated that integration of AI in their medical curriculum is essential (76.2%).	-	Update medical curriculum to equip trainees with knowledge and skills to effectively use AI applications.
Joachim Kimmerle [28]	86.2% agreed that basic AI knowledge should be taught in medical studies.	Limited knowledge and exposure to AI technologies among medical students.	Ensure professional values and rights are protected in AI education.
David Shalom Liu [32]	355 out of 388 respondents (91.5%) agreed that training in AI concepts during medical school would be useful for their future. Most students (89.4%) agreed that they wanted to learn what medical students should know about AI in medicine.	Need for structured AI education in medical training.	Incorporate AI education into undergraduate and postgraduate medical curricula.
Aidan Pucchio [31]	Respondents agreed that AI should be formally taught in medical education (67%).	Concerns about overreliance on AI for diagnosis.	Provide practical exposure to AI applications in medicine.
Jonathon Stewart [37]	Most of the students (58.6%) agreed that AI should be part of medical training, and most (72.7%) wanted more teaching focusing on AI in medicine.	Lack of faculty expertise in AI education.	Address gender and tech-savviness disparities in AI knowledge.
Sarya Swed [42]	1286 out of 1484 participants (85.3%) highly acknowledged the inclusion of the AI curriculum in medical graduate and undergraduate studies.	Limited accreditation requirements for AI in medical curricula.	Develop structured AI modules in medical education.
Minnie Teng [30]	1373 out of 2167 participants (63.4%) believed that gaining basic literacy in AI should be part of their curriculum.	Overcrowded medical curricula, making AI integration difficult.	Provide hands-on AI training in clinical settings.
Alvin Yong Zong Tung [48]	209 out of 301 respondents (69.4%) agreed that all students should receive teaching in AI.	Insufficient AI training among faculty members.	Address misconceptions about AI’s role in medical careers.
Lukas Weidener [51]	385 out of 487 participants (74.9%) agreed on the need for AI and AI ethics instruction in medical education.	Limited AI content in the curriculum.	Integrate AI courses within medical education.
Elena A. Wood [39]	AI should be part of medical education and training. Mean agreement score: 4.0 (SD: 1.0) on a 5-point Likert scale	Concerns about AI replacing physicians, particularly radiologists.	Provide practical AI training in clinical settings.

Overall positive attitudes

The majority of medical students demonstrated positive attitudes toward AI and its incorporation into medical education, acknowledging its growing importance in clinical practice. For example, Bisdas et al. reported that 2,683 out of 3,133 students (85.6%) perceived AI as a partner in medical practice [23], reflecting optimism about AI’s role in medicine [44].

Support for AI training and curriculum integration

Many students expressed strong support for integrating AI training into their medical education [26, 30, 35–48]. Swed et al. found that 1,286 out of 1,484 participants (85%) favored including AI training in the curriculum [42]. Other studies reported similar findings, with 616 out of 900 students (68.4%) considering AI education mandatory [44], 4,267 out of 4,492 participants (95%) showing interest in learning AI fundamentals [24], and 357 out of 390 respondents (91.5%) agreeing that AI training would benefit their future careers [32]. In Jha et al., 76.2% of participants believed that every medical student or trainee should receive training on AI competencies [47]. Overall, students consistently expressed the desire for increased teaching on AI in medicine [39, 42], emphasizing its inclusion as a core component of medical education.

Perceived benefits and interests

Medical students recognize AI’s potential to improve healthcare practice. Positive attitudes toward AI were associated with the belief that it can revolutionize medicine and enhance patient care [40]. Structured AI training was highly valued, with 243 out of 325 respondents (74.8%) highlighting its role in reducing medical errors and 258 out of 325 respondents (79.4%) emphasizing its relevance to ethical decision-making [46]. Similarly, Liu et al. reported that 355 out of 388 participants (91.5%) acknowledged the importance of understanding essential AI concepts within medical education [32].

These findings indicate that medical students not only value AI’s potential to transform medical practice but also strongly advocate for formal AI training within undergraduate curricula to equip them with the competencies required for future clinical practice.

Challenges in the implementation of AI in medical curricula

Integrating AI into medical education faces multiple, interrelated challenges. Curriculum constraints—including overcrowding and time limitations—make it difficult to incorporate AI topics [27, 33, 43, 44]. Resource and expertise gaps are also significant, with limited qualified instructors, insufficient faculty awareness, and students’ lack of technical foundations in programming or computer science [27, 32, 33, 43, 48, 49]. Financial limitations, such as high software costs and restricted funding, further hinder implementation [27, 46, 48, 49].

Ethical and social concerns include data privacy, reduced human interaction, AI reliability, and trust issues arising from algorithmic biases [28, 34, 41, 45, 48]. Additional barriers involve clinical adoption challenges, regional disparities in healthcare infrastructure, and potential oversimplification of AI understanding among students [39, 41, 42, 45].

Other obstacles include a lack of systematic evidence, complex regulatory requirements, and limited awareness among educators of AI’s evolving role in medical practice [45, 47, 51]. Overall, integrating AI into medical curricula requires navigating curricular, resource, ethical, and systemic challenges [27, 34, 39, 45, 48, 49].

Recommendations for future AI curriculum development

Future AI curricula in medical education should be designed to equip trainees with the knowledge and skills to effectively and ethically use AI applications [28, 29]. Key recommendations emphasize several themes (Table 3). AI-related teaching should be incorporated into the core medical curriculum [40, 41]. Integrating AI into pre-existing courses can reduce curricular overload and highlight its clinical applications [46, 47]. The AI curriculum may be delivered as a separate module or integrated into current modules, focusing on AI applications relevant to each specialty [41]. A hybrid formal and flexible format appears most suitable for incorporating AI topics in US medical schools [32].

Table 3.

Key themes and recommendations for AI curricula in medical education

Theme	Recommendations	References
Curriculum Integration	Incorporate AI into the core curriculum, either as separate modules or integrated into existing courses; adopt a hybrid formal and flexible format	[32, 41, 46, 47]
Content and Focus	Cover AI principles, clinical applications, ethical and legal considerations, mathematical foundations, AI-assisted risk analysis, and appropriate clinical use	[30, 32, 46, 49]
Hands-on Experience	Provide practical training through AI simulators, virtual patients, case studies, and realistic application scenarios	[24, 45, 51]
Ethical Considerations	Address patient privacy, digital divides, transparency, bias, equitable access, critical appraisal skills, and human oversight	[28, 34, 40]
Interdisciplinary Learning	Encourage collaboration with computer science, biomedical engineering, and data science professionals for AI development and evaluation	[41, 47]
Training Formats	Offer workshops, lectures, collaborative activities, continuing medical education, and accessible media such as videos or podcasts	[31, 32, 42, 49]
Faculty Development	Prepare educators to teach AI effectively	[39]
Misconceptions and Acceptance	Address fears of AI replacing clinicians and promote AI-assisted decision-making	[24, 30]
Continuous Improvement	Update curricula regularly to reflect advances in AI and implement adaptable educational strategies	[40, 51]
Regional Considerations	Adapt AI curricula to local healthcare needs, as demonstrated in Syria, Malaysia, Oman, and Nepal	[36–48]

The curriculum should focus on the applications of AI in clinical medicine [41]. It should cover fundamental AI concepts, guidance on when to use AI in medicine, and the benefits and limitations of AI applications [32]. Essential content includes knowledge and skills related to AI applications, training to prevent and address ethical challenges, and AI-assisted risk analysis for diseases [46]. Teaching should address mathematical concepts to understand algorithms, the fundamentals of AI such as data science, and potential ethical and legal issues associated with AI use [49]. Basic AI programs should help students identify appropriate clinical contexts for AI, understand ethical implications, and recognize ways AI can improve healthcare quality, accommodating students’ interests across multiple disciplines [30].

Hands-on experience and practical application should be emphasized through AI training in clinical settings [24]. This can include AI-based simulators, virtual patient platforms, and practical content such as case studies and application scenarios [45, 51]. Introductory courses should cover AI principles, applications, and ethical considerations, and realistic use cases should illustrate both capabilities and limitations of AI [33, 45, 46].

Ethical considerations should be integrated into medical education, particularly in courses on medical ethics, focusing on patient privacy, digital divides, and data protection [28, 34]. Instruction should address transparency, explainability, discrimination, and bias in AI systems [28, 40]. Training should promote equitable access to AI knowledge, critical appraisal skills for AI information, and appropriate human oversight to detect errors or biases in AI algorithms while ensuring clinicians remain the ultimate decision-makers [28, 34, 40].

Interdisciplinary learning should be encouraged by engaging students with fields such as computer science and biomedical engineering [41]. Innovative approaches can foster collaboration with software engineers, computer scientists, and other professionals involved in developing and evaluating AI algorithms [41, 47]. Collaboration among data scientists, computer scientists, practicing physicians, and medical education experts is essential [47].

Preferred training formats include short lectures, formal electives, workshops, Q and A panels, collaborative activities, and continuing medical education programs [31, 32, 42, 49]. Learning outside traditional lectures, such as through videos or podcasts, can enhance accessibility and engagement [31]. Faculty development is also critical to ensure educators are prepared to teach various aspects of AI technologies [39].

Concerns and misconceptions about AI should be addressed to foster acceptance and encourage AI-assisted decision-making [24, 30]. Fears of AI replacing healthcare professionals can be mitigated by integrating AI education into medical training [30]. The curriculum should be continuously updated to reflect technological advances and adaptable educational strategies should be implemented throughout medical training [40, 51].

Specific regional considerations highlight the importance of adapting AI curricula to local contexts. Syrian medical authorities and international organizations should consider AI integration in residency and fellowship training [42]. Malaysian universities and medical educators in Oman are encouraged to incorporate AI education into medical curricula to prepare future physicians [36, 48]. Systematic implementation of AI in the Nepalese healthcare system could address challenges related to resource and workforce constraints [47].

By integrating these recommendations, medical institutions can develop comprehensive AI curricula that prepare future physicians to utilize AI effectively and ethically, ultimately improving healthcare outcomes.

Meta-analysis results

Using the Freeman-Tukey double arcsine transformation, the individual study proportions ranged from 0.582 (95% CI: 0.494–0.667) to 0.938 (95% CI: 0.929–0.946). Overall, the proportion with favorable attitudes was 0.779 (95% CI: 0.774–0.783) under the common-effect model and 0.780 (95% CI: 0.730–0.820) under the random-effects model using the DerSimonian-Laird estimator (Fig. 2). The large prediction interval ([0.484; 0.967]) shows immense heterogeneity in the students’ attitudes among the studies, i.e., most of the estimates are clustered around the pooled proportion, but some studies have substantially lower or higher rates.

Fig. 2.

Forest plot of pooled proportion of positive attitudes toward integration of AI curricula in UME

There was substantial heterogeneity (I² = 98.5%, τ² = 0.0214, Q = 1621.50, df = 25, P < 0.001) so that the majority of the variance was due to differences between studies and not due to random chance (Fig. 2). However, the funnel plot (Fig. 3) was asymmetrical to no great extent, suggesting slight publication bias, which was supported by Egger’s test (P = 0.624) and Begg’s test (P = 0.707). However, the use of the Trim-and-Fill method gave the imputation of six studies, shifting the random-effects pooled proportion to 0.818 (95% CI: 0.771–0.861) with a prediction interval of [0.500; 0.991] (see Supplementary Material: Appendix 3, Figure S1). This adjustment indicates that the true pooled estimate would be marginally more significant if publication bias were present than the initially calculated.

Fig. 3.

Funnel plot for assessment of publication bias

Assessment of publication bias yielded mixed findings in our study. Egger’s and Begg’s tests did not indicate significant effects, consistent with the absence of strong statistical evidence of publication bias. Nonetheless, the funnel plot showed some asymmetry, and the Trim-and-Fill procedure imputed six potentially missing studies. Although such imputation adjusted the pooled estimate upward only slightly, this suggests that selective reporting cannot be fully excluded. Importantly, in the context of very high heterogeneity, funnel plot–based methods and Trim-and-Fill may be less reliable, as asymmetry can arise from heterogeneity itself rather than true publication bias. Taken together, these results imply that while the possibility of bias should be acknowledged, our main conclusion about the report of supportive attitudes toward AI integration appears robust to potential unpublished data.

To identify sources of heterogeneity, we conducted subgroup analyses. Subgroups were stratified by region (Asia, the Middle East, North America, and others) and by type of questionnaire (standardized tools and researcher-developed surveys). The pooled proportion with favorable attitudes toward AI was: North America (0.780, 95% CI: 0.660–0.880), Asia (0.780, 95% CI: 0.700–0.860), the Middle East (0.730, 95% CI: 0.660–0.800), and other regions (0.910, 95% CI: 0.830–0.960). Stratified by type of questionnaire, researcher-developed surveys gave a proportion of 0.780 (95% CI: 0.710–0.840), and standardized tools gave 0.780 (95% CI: 0.690–0.850). Heterogeneity was extremely high across all subgroups, ranging from 86.4% to 98.8% (Figs. 4 and 5), indicating that regional or methodological differences alone do not account for the variability in student attitudes. The consistently high level of heterogeneity suggests the presence of additional underlying factors influencing these attitudes.

Fig. 4.

Subgroup analyses of pooled proportion by region

Fig. 5.

Subgroup analyses of the pooled proportion by region

The quality assessment of the 26 included studies, conducted using the 20-item AXIS critical appraisal tool, revealed generally high methodological quality, with total scores ranging from 13 to 18 out of a possible 20. The majority of studies (n = 23) scored 14 or above, indicating robust reporting in core domains such as study aims, design, and ethical considerations. All studies clearly stated their objectives and appropriately described their target populations and sample characteristics. However, the appraisal consistently identified specific areas of weakness across the literature. A critical and near-universal limitation was the failure to account for non-responders, with only one study (Pucchio et al., 2022) addressing non-response bias. Consequently, items 12 and 13, concerning the handling of non-responders and non-response bias, were the most frequently unmet criteria. Other common methodological shortcomings included a lack of justification for sample sizes (items 5 and 6 were often scored 0) and a failure to address or control for potential confounding variables (item 15 was almost universally scored 0). Overall, Bisdas et al. (2021) achieved the highest score (18/20), demonstrating the most comprehensive adherence to the AXIS criteria, while the lowest scores (13/20) were observed in studies by Ejaz et al., Kimmerle et al., and Liu et al., primarily due to issues with sample justification and response rate reporting (see Supplementary Material: Appendix 2, Tables S1 and S2).

Then, we conducted meta-regression to see if factors at the level of the studies explained heterogeneity in students’ attitudes. None of the tested predictors reached statistical significance in univariable analyses: study quality (AXIS score:β = -0.0096, P = 0.707), publication year (β = − 0.0161, P = 0.491), sample size (β = 0.0000, P = 0.727), and basic knowledge (β = 0.0001, P = 0.928). In the multivariable model, with publication year, sample size, basic knowledge about AI, and quality of the studies, none of these factors influenced the overall pooled proportion. The heterogeneity remained high (I² = 98.43%, τ² = 0.0289), and the R² value indicated that these factors explained 0% of the variability. This suggests that factors beyond the level at which we could capture them, such as differences in AI curricula, faculty knowledge, or regional policies for the inclusion of AI in medical education, drive students’ attitudes.

Sensitivity analysis confirmed the robustness of the findings. The pooled estimate remained stable, ranging from 0.782 to 0.804, when each study was sequentially omitted (Fig. 6).

Fig. 6.

Sensitivity analysis of pooled proportion upon study omission

A cumulative meta-analysis examined the trend in the pooled proportion over time (see Supplementary Material: Appendix 3, Figure S2). The earliest study, D. Pinto dos Santos [19], reported an initial pooled proportion of 0.715. The proportion slightly declined with Cherry Sit [33] (0.700) and Elena A. Wood [39] (0.693) before stabilizing around 0.780 by 2022, as seen in studies such as Zaboor Ahmed [43] and Fahad Abdullah Saeed Alwadani [26]. Confidence intervals narrowed with the progression of time, which is an indication of increased precision in estimates. The pooled proportion plateaued, showing that students’ attitudes toward integrating AI curricula in UME have remained consistently positive in the past few years with slight fluctuations.

Overall, medical students strongly supported AI training (pooled agreement: 83.1%), but only 36.4% reported confidence in applying AI clinically, highlighting a persistent optimism-competence gap.

Discussion

This systematic review and meta-analysis demonstrate that approximately 78% of medical students worldwide support the integration of AI into UME, although substantial variability and an optimism–competence gap were observed. The pooled proportion (78%) represents a descriptive summary across studies with extreme heterogeneity (I² = 98.5%). Drawing on 26 studies from 21 countries, the review provides a comprehensive overview of the current landscape of AI in medical education and students’ perspectives on UME. The findings offer critical insights into the opportunities, challenges, and future directions for AI integration, highlighting the necessity of structured educational frameworks to adequately prepare future healthcare professionals.

Three main findings emerged. First, the meta-analysis quantified strong global support, with a pooled proportion of 78.0% (95% CI: 0.730–0.820) of students expressing positive attitudes toward AI in medical education. Second, a marked disparity was observed between this optimism and students’ perceived competence: while 83.1% agreed on the need for AI training, only 36.4% felt confident in applying AI clinically. Third, the review identified common barriers to implementation, including curricular overcrowding (reported in 68% of studies), lack of faculty expertise (52%), and ethical concerns (41%). Notably, the extreme heterogeneity (I² = 98.5%) suggests that unmeasured institutional factors—such as local faculty expertise, curriculum design, or elements of the hidden curriculum—may play an important role in shaping student readiness. This variability raises the possibility that such institutional-level influences could be as substantial as, or potentially more influential than, broad regional or demographic differences; however, this interpretation remains speculative and requires direct investigation in future research.

Assessment of medical students’ attitudes and perceptions

The integration of AI into UME has gained widespread support among medical students worldwide, with 70–95% endorsing its importance for future clinical practice. This enthusiasm is largely driven by AI’s perceived potential to reduce diagnostic errors (57.2% in India) [46] and enhance healthcare accessibility (49% in Turkey) [35]. The meta-analysis further quantifies this global enthusiasm, revealing that 78.0% of students expressed positive attitudes toward AI integration (95% CI: 0.730–0.820). This proportion remained consistent across regions—North America (78.0%), Asia (78.0%), and the Middle East (73.0%)—and across questionnaire designs, suggesting a broad consensus on the relevance of AI in medical training. However, demographic factors introduced some variation: male students and those with STEM (science, technology, engineering, and mathematics) backgrounds reported higher confidence in AI applications, whereas female and non-STEM students expressed greater ethical concerns [34, 51]. These gender- and discipline-based differences mirror broader trends in technology adoption across professional fields.

Despite their optimism, many students feel inadequately prepared to use AI in clinical contexts. Only 31% of respondents in Malaysia and 18.4% in Jordan reported feeling competent to apply AI practically [44, 48]. This pronounced “optimism–competence gap” suggests that current medical curricula emphasize theoretical understanding over practical application—an imbalance previously observed in the implementation of genomics and telemedicine education. In this review, the term “optimism–competence gap” is used to describe the observed discrepancy between high levels of support for AI training (83.1%) and substantially lower self-reported confidence in applying AI in practice (36.4%). The quantified discrepancy (83.1% vs. 36.4%) provides a novel empirical benchmark for assessing the effectiveness of future AI curricular interventions. The pooled proportion of AI readiness (83.1%) represents a descriptive summary across highly heterogeneous studies (I² = 98.5%). Thus, the 83.1% pooled estimate reflects overall trends across diverse contexts rather than a uniform level of readiness. Although student attitudes toward AI have remained consistently positive since 2022, the substantial heterogeneity across studies (I² = 98.5%) indicates marked differences in preparedness levels. This variability may reflect unmeasured institutional factors—such as the presence of formal AI curricula, faculty expertise, or elements of the hidden curriculum—that could influence students’ readiness to apply AI in clinical settings. However, as these factors were not directly assessed in the included studies, this explanation remains hypothetical and warrants empirical examination. Possible explanations for this optimism–competence gap, as suggested by prior literature, may include limited hands-on exposure to AI tools, variability in curricular integration, and the influence of institutional or informal learning environments. However, these factors should be interpreted as hypotheses rather than direct findings of the present meta-analysis. Qualitative insights from Palestine [49] and Ireland [34] further highlight students’ preference for interactive, practice-oriented learning experiences, such as AI-assisted simulations, to enhance hands-on competence.

The present findings both corroborate and extend previous research emphasizing students’ enthusiasm for AI. While earlier studies predominantly examined perspectives from HICs, this review incorporates evidence from LMICs, thereby illuminating context-specific challenges such as the post-war healthcare environment in Syria [42] and curriculum constraints in Nepal [47]. The disparity in AI exposure between students in HICs and LMICs—where formal AI education remains limited or absent [24, 41]—underscores persistent global inequities in access to AI-related educational resources, paralleling challenges seen in engineering and computer science domains. Moreover, the coexistence of optimism about AI’s benefits and apprehension regarding job displacement reflects varying levels of AI literacy, a pattern consistent with broader studies of technology adoption in healthcare. Collectively, these findings emphasize the urgent need to align students’ enthusiasm with practical competence through structured, equitable, and hands-on AI education within medical curricula.

Gaps, challenges, and policy implications for AI integration in medical education

Integrating AI into UME faces three key barriers: curricular overcrowding, limited faculty expertise, and misconceptions about AI’s role in clinical practice. Curricular overload is particularly pronounced in LMICs, while the substantial heterogeneity observed across studies may be attributable, at least in part, to unmeasured institutional factors such as faculty training, institutional commitment, and hidden curricula. Although this interpretation is consistent with the data, it remains inferential and should be tested in future studies designed to isolate these institutional-level effects. Faculty unpreparedness further limits meaningful engagement with AI, forcing reliance on informal learning, whereas HICs have begun addressing these gaps through interdisciplinary collaborations. Misconceptions about AI contribute to career-related anxieties, particularly in “AI-vulnerable” specialties, underscoring the need for curricula that emphasize AI as a supportive tool.

Addressing these challenges requires context-sensitive strategies. In HICs, this includes developing interdisciplinary faculty expertise, advanced AI modules, and industry partnerships. In LMICs, scalable, low-cost interventions—such as webinars, online modules, and embedding core AI concepts in existing courses—can gradually build capacity. Standardized competency frameworks, continuous faculty development, and cross-sector collaboration are essential for sustainable, practice-oriented AI education. Together, these measures can transform student enthusiasm into applied competence, enabling effective and ethical integration of AI into medical training.

Limitations and future research priorities

This systematic review and meta-analysis have several methodological limitations that should be considered when interpreting the findings. First, the review protocol was not prospectively registered in PROSPERO or any other public registry. While we strictly adhered to the PRISMA 2020 guidelines and transparently reported all methodological steps, the absence of prior registration remains a limitation, as it pre-specifies fewer analytical decisions and theoretically increases the risk of selective outcome reporting. Second, the vast majority of the included studies relied on cross-sectional designs and self-reported survey data rather than objective proficiency assessments. This methodology carries inherent risks of social desirability bias and often conflates students’ perceived enthusiasm for technology with actual digital and AI literacy. Third, we observed extreme statistical heterogeneity across the pooled analyses. This variability is likely driven by pronounced geographic, linguistic, and institutional differences, which conventional subgroup analyses could not fully resolve. A related constraint is the conceptual ambiguity surrounding ‘AI’ within the primary literature; most existing studies treat artificial intelligence as a single, monolithic construct rather than distinguishing between specific subtypes, such as machine learning, natural language processing, or generative AI. Consequently, this review could not evaluate student readiness with domain-specific granularity. Furthermore, important institutional-level moderators—such as local curricular space, financial resources, and faculty readiness—remained unmeasured and could not be empirically quantified due to the lack of granular, contextual data in the primary studies. Lastly, our focus on undergraduate medical education limits the direct generalizability of these insights to postgraduate residency programs or continuing medical education.

Future research should prioritize longitudinal and interventional studies to evaluate the impact of AI education on clinical decision-making and patient outcomes. Culturally adaptable frameworks, interdisciplinary approaches addressing ethical dimensions, and refined definitions of AI subtypes are critical. Investigating moderators such as institutional infrastructure and prior exposure will enhance understanding. These efforts will inform evidence-based AI curricula that prepare medical students to engage with AI competently, critically, and ethically.

Conclusion

This review advances the field by demonstrating consistent global support for AI integration in UME while quantifying a critical optimism–competence gap (83.1% vs. 36.4%). The substantial heterogeneity observed further suggests that institutional factors—such as curricular design and faculty expertise—may influence student readiness, potentially outweighing broad regional patterns. This hypothesis requires direct validation in future research. This discrepancy highlights a mismatch between favorable attitudes toward AI education and perceived competence, rather than providing evidence of a specific underlying mechanism. These findings indicate that integrating comprehensive AI education into medical curricula is becoming increasingly critical. As AI technologies begin to reshape clinical workflows, adapting medical education is essential to ensure future physicians can utilize these tools both competently and ethically. The observed gap between student awareness and practical application points to an educational lag in aligning current undergraduate curricula with the rapid evolution of AI-driven healthcare. Addressing this discrepancy is important to prevent future cohorts of clinicians from facing preparedness deficits in their clinical practice.

Challenges such as curricular overcrowding, faculty unpreparedness, and ethical concerns are substantial but surmountable. Strategic investment in faculty development, institutional policy, and context-specific curriculum design is essential. Structured AI competency frameworks can be integrated into existing courses to mitigate overload, while interdisciplinary collaborations can address faculty expertise gaps. High-income countries can leverage advanced partnerships, whereas low- and middle-income countries should prioritize accessible, foundational AI education to prevent widening global inequities.

AI education must move beyond theoretical exposure to include simulation, ethical training, and instruction on critically appraising algorithmic biases. Even when students are not immediate users of AI, they must be informed critics. Policymakers and accrediting bodies should define standardized competencies to ensure baseline AI literacy comparable to foundational subjects like anatomy or physiology.

Future research should aim to close these gaps. Interventional trials evaluating hybrid learning models, longitudinal studies tracking AI’s impact on clinical decision-making, and the development of ethical frameworks addressing societal implications of AI are essential next steps. Medical education must advance in step with AI’s trajectory to lead its development, rather than react to it.

Ultimately, bridging this curriculum gap is closely linked to the long-term quality of patient care. Preparing AI-literate clinicians is expected to support the development of safer and more equitable healthcare delivery. Proactive curriculum adaptation in UME can empower the next generation of physicians, ensuring they remain capable and critical participants in an evolving technological landscape.

Abbreviations

AI

Artificial Intelligence

UME

Undergraduate Medical Education

ML

Machine Learning

DL

Deep Learning

NLP

Natural Language Processing

PRISMA

Preferred Reporting Items for Systematic reviews and Meta-Analyses

LMICs

Low- and Middle-Income Countries

HICs

High-Income Countries

Authors’ contributions

H.M. conceptualized the study and supervised the project. H.M., E.T., S.M., and S.S. contributed to study design and methodology, quality assessment, and manuscript drafting. S.M. conducted the literature search, while H.M. and S.S. performed screening and selection of studies. Data extraction was carried out by S.S., and interpretation of results was performed by H.M., S.S., and E.T. Critical revision of the manuscript was done by H.M., E.T., and S.S. All authors read and approved the final manuscript.

Funding

None.

Data availability

The datasets are available from the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate.

Not applicable.

Consent for publication

Not applicable

Competing interests

The authors declare no competing interests.