Effects of Artificial Intelligence-Supported Education on Critical Thinking, Problem-Solving, Clinical Reasoning, and Decision-Making in Health Science Education: A Systematic Review of Interventional Studies Authors: Anas Ali Alhur, Nouf Khalid Al-Kahtani, Emad A Mahgoub Published: 2026 ABSTRACT BackgroundArtificial intelligence (AI) is increasingly incorporated into health science education through diagnostic systems, intelligent tutoring systems, adaptive learning platforms, chatbots, virtual patients, and large language models. However, its effects on higher-order cognitive skills remain uncertain.ObjectiveTo synthesize interventional evidence on the effects of AI-supported education on critical thinking, problem-solving, clinical reasoning, and decision-making among health science learners.MethodsThis systematic review followed PRISMA 2020 and a prospectively registered protocol (PROSPERO CRD420251128495). PubMed, Scopus, the Cochrane Library, Google Scholar, and supplementary sources were searched for interventional studies evaluating AI-supported education and at least one of four predefined cognitive outcomes: critical thinking, problem-solving, clinical reasoning, or decision-making. Searches were updated through 24 August 2026. Randomized studies were assessed using RoB 2, non-randomized intervention studies using ROBINS-I, and quasi-experimental studies using JBI critical appraisal criteria. Findings were synthesized narratively and classified as positive, null, mixed, or negative.ResultsTwenty-nine interventional studies involving 2138 participants met the eligibility criteria. The included studies evaluated conventional and generative AI technologies, including diagnostic AI systems, intelligent tutoring systems, AI-enabled chatbots, generative image systems, and large language models. Findings varied across the four cognitive outcomes. Structured AI-supported interventions frequently demonstrated favorable effects on selected clinical reasoning, decision-making, problem-solving, and critical-thinking outcomes; however, null, mixed, and negative findings were also reported, particularly when AI was compared with established educational resources or used without structured instructional guidance.ConclusionAI-supported education may enhance selected higher-order cognitive outcomes among health science learners, but the evidence is inconsistent across technologies, instructional designs, comparators, and cognitive constructs. Structured integration that preserves active learner engagement and independent reasoning appears important. Further rigorous studies using validated outcome measures and assessments of sustained, unaided performance are needed. Introduction Artificial intelligence (AI) has become increasingly integrated into health science education, reflecting broader developments in digital learning, clinical decision support, simulation, and data-driven healthcare. AI is a broad category encompassing computational systems capable of performing tasks commonly associated with human intelligence, including pattern recognition, prediction, classification, natural language processing, adaptive learning, decision support, and content generation.1–4 Within health professions education, these applications include machine-learning and deep-learning systems, intelligent tutoring systems, AI-enabled chatbots, diagnostic decision-support tools, AI-supported simulations, generative image systems, and large language models (LLMs).2–6 Generative artificial intelligence (GenAI), characterized by its capacity to generate new text, images, or other content in response to user inputs, therefore represents a subtype of AI rather than being synonymous with AI as a whole.2,3,6 The application of AI in education predates the rapid expansion of conversational LLMs. Earlier educational interventions incorporated intelligent tutoring systems, adaptive learning technologies, diagnostic classifiers, image-recognition systems, and algorithm-based decision-support tools.4,7,8 More recently, the emergence of ChatGPT and other LLM-based systems has expanded the educational functions of AI to include conversational tutoring, individualized feedback, case generation, simulated patient interactions, explanation generation, and support for problem-based and case-based learning.2,3,6,9 This evolution has resulted in considerable technological and pedagogical heterogeneity. A recent systematic review of randomized controlled trials in undergraduate health professions education identified 13 AI subcategories based on technology type and educational function, including LLM-based personalized learning aids, LLM content generators, natural language processing chatbots, and non-LLM adaptive learning platforms.10 These differences are important because AI systems with different functions and modes of interaction may influence learning through different educational and cognitive mechanisms and should not be considered a homogeneous intervention. Health science education requires learners to progress beyond factual knowledge acquisition toward higher-order cognitive processes necessary for safe and effective professional practice. These processes include interpreting and evaluating information, identifying and resolving problems, generating and testing diagnostic hypotheses, managing uncertainty, and selecting appropriate clinical actions.11–14 AI may support these processes by providing individualized feedback, repeated practice, adaptive cases, diagnostic exercises, interactive simulations, virtual patient encounters, and opportunities to compare alternative explanations or decisions.6,10,15 AI-generated virtual patients, for example, are increasingly being used for communication training, diagnostic reasoning, management reasoning, and simulation-based education and may provide scalable, immediate, and individualized learning experiences.15 A 2026 scoping review identified 97 studies involving AI-generated virtual patients and documented applications across communication, diagnostic reasoning, and management reasoning, while also identifying limitations related to clinical accuracy, conversational authenticity, and excessive agreeableness.15 However, the availability of AI does not necessarily translate into improved higher-order cognitive performance. Recent evidence demonstrates substantial variability in educational effects according to the AI technology, instructional design, comparator, learner characteristics, and outcome assessed.10,16,17 In a 2026 systematic review and meta-analysis of 66 randomized controlled trials involving 4911 undergraduate health professions learners, Lai et al found favorable effects for selected outcomes and AI applications, but substantial heterogeneity, imprecision, and predominantly low-to-very-low certainty of evidence limited the strength and reproducibility of the findings.10 Similarly, a recent systematic review focused specifically on critical thinking and clinical reasoning concluded that AI-based educational interventions may improve these competencies, but emphasized that the evidence remained uncertain because of small samples and substantial intervention heterogeneity.16 These findings suggest that the educational effect of AI depends not simply on whether learners have access to AI, but on how AI is incorporated into the learning process and what cognitive activity learners are required to perform. Four higher-order cognitive outcomes are of particular importance in health science education: critical thinking, problem-solving, clinical reasoning, and decision-making. Although these constructs are closely related, they are conceptually distinct and should not be used interchangeably. Critical thinking involves purposeful, self-regulatory judgment through interpretation, analysis, evaluation, inference, explanation, and reflection.11 It enables learners to evaluate the quality of evidence, identify assumptions, consider competing explanations, and justify conclusions rather than accepting information at face value.11,18 This competency has become particularly relevant in AI-supported learning because learners may need to critically evaluate the reliability, completeness, and contextual appropriateness of AI-generated information.18,19 Problem-solving involves recognizing and representing a problem, identifying relevant information, generating possible solutions, evaluating alternatives, implementing an appropriate response, and revising the approach when necessary.12,20 AI may support this process by providing interactive cases, prompts, feedback, and alternative approaches, particularly when integrated within problem-based learning.20,21 However, AI-generated solutions may also alter the cognitive demands of a task if learners can obtain an answer without independently representing the problem or generating and evaluating alternatives. Consequently, improved performance on an AI-assisted task should not automatically be interpreted as improved independent problem-solving ability. Clinical reasoning is a health-professions-specific process through which learners integrate patient information with biomedical and clinical knowledge, generate and evaluate hypotheses, interpret findings, manage uncertainty, and formulate clinical judgments.13,16 It is central to the translation of theoretical knowledge into patient-centered clinical practice. AI-supported clinical reasoning has received increasing research attention through diagnostic decision-support systems, intelligent tutors, AI-supported PBL, LLM-generated feedback, and virtual or standardized patients.8,16,22–24 Recent systematic reviews indicate potential benefits, but also emphasize substantial heterogeneity and uncertainty regarding whether improvements observed during AI-supported activities transfer to independent clinical reasoning.16,22,23 Decision-making is related to clinical reasoning but represents a distinct process involving the selection of an action from available alternatives after considering evidence, uncertainty, risks, benefits, and contextual factors.13,25 AI can potentially support decision-making by organizing information, identifying patterns, presenting alternatives, and providing recommendations. Nevertheless, learners remain responsible for determining whether AI-generated information is reliable and appropriate for the specific clinical context. This distinction is particularly important because an AI system may improve access to information or diagnostic suggestions without necessarily improving the learner’s underlying capacity to reason independently or select an appropriate action. The relationships among these four constructs are therefore important but should not obscure their conceptual differences. Critical thinking and problem-solving may contribute to clinical reasoning, while decision-making can represent an action-oriented outcome arising from the reasoning process. Improvements in academic achievement, knowledge scores, satisfaction, communication, technical skills, or global clinical competence do not necessarily demonstrate improvement in any of these four cognitive constructs unless the relevant process is directly or separately assessed. This distinction is especially important in AI-education research, where heterogeneous outcome definitions and measurement approaches have complicated comparisons across interventions.10,16 Recent reviews have similarly identified variation in the methods used to operationalize clinical reasoning and higher-order cognitive outcomes.16,22,23 The potential effects of AI on these cognitive outcomes may also be understood through established theories of learning. Experiential learning emphasizes the development of knowledge through experience, reflection, conceptualization, and subsequent application.26 AI-supported simulations, virtual patients, and interactive cases can potentially provide repeated experiential learning opportunities when learners are required to interpret information, make decisions, receive feedback, and reconsider their reasoning.15,26 Constructivist approaches similarly emphasize active knowledge construction rather than passive receipt of information, suggesting that AI may be most educationally useful when it functions as a scaffold for inquiry, comparison, explanation, and reflection rather than as a substitute for learner reasoning.2,27 Cognitive load theory provides an additional perspective. Working memory has limited capacity, and instructional design can facilitate learning by reducing unnecessary cognitive demands while directing cognitive resources toward processes relevant to schema acquisition and problem solving.28,29 AI may reduce avoidable demands associated with searching, organizing information, or generating practice opportunities; however, if it performs central reasoning processes on behalf of learners, it may also reduce opportunities for the cognitive activity required to develop independent expertise. These competing possibilities reinforce the importance of examining not only whether AI is used, but how it is integrated into educational activities and whether cognitive performance persists beyond AI-assisted conditions. This issue is increasingly important as empirical findings remain mixed. Individual interventional studies have reported improvements in problem-solving, clinical reasoning, decision-making, and higher-order thinking following structured AI-supported learning.20,21,24,30–32 Conversely, other studies have reported null or unfavorable effects when AI was compared with expert feedback, conventional instruction, or other established educational resources.25,33,34 For example, ChatGPT-generated feedback did not demonstrate an overall advantage over expert feedback for clinical reasoning questions in one randomized trial,33 while Shin et al reported poorer ethical decision-making and clinical reasoning outcomes in an AI-assisted group compared with textbook-based instruction.25 Such findings indicate that the relationship between AI and higher-order cognition is unlikely to be uniformly positive and may depend on the instructional context, comparator, and degree of learner engagement. The need for a focused synthesis remains despite the rapid publication of broader reviews of AI in health professions education. Lai et al10 synthesized 66 randomized controlled trials involving 4911 participants across multiple AI technologies and educational outcomes; however, their review addressed a broad range of outcomes rather than focusing specifically on the four cognitive constructs examined here. Tay et al16 more specifically examined critical thinking and clinical reasoning across 21 randomized or controlled clinical trials and found small-to-medium favorable effects for selected surrogate clinical-reasoning outcomes, while emphasizing low-to-very-low certainty of evidence and considerable intervention heterogeneity. Other recent reviews have focused specifically on AI-supported clinical reasoning in nursing,22,23 AI-generated virtual patients,15 or GenAI-based approaches to clinical reasoning assessment.35 Despite this growing evidence base, an important gap remains. Existing syntheses have generally examined broad educational outcomes, focused on individual professions or specific AI applications, restricted eligibility to selected study designs or learner groups, or concentrated primarily on critical thinking and clinical reasoning.10,15,16,22,23,35 No synthesis has comprehensively characterized interventional evidence across the combined constructs of critical thinking, problem-solving, clinical reasoning, and decision-making while simultaneously considering both conventional and generative AI technologies and their educational functions. Examining these constructs together while preserving their conceptual distinctions may provide a more precise understanding of where AI-supported education demonstrates benefit, where effects are null or unfavorable, and which instructional approaches warrant further evaluation. Accordingly, this systematic review aimed to synthesize interventional evidence concerning the effects of AI-supported education on critical thinking, problem-solving, clinical reasoning, and decision-making among health science learners. Specifically, the review sought to characterize the AI technologies and educational functions used; examine instructional designs, comparators, and assessment approaches; evaluate the methodological quality of the evidence; compare the direction and consistency of findings across the four cognitive outcomes; and identify implications for educational practice and future research. Methods Design, Protocol Registration, and Reporting This systematic review was conducted and reported in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 statement. A protocol specifying the review objectives, eligibility criteria, information sources, outcomes, and planned methods was prospectively registered in the International Prospective Register of Systematic Reviews (PROSPERO; CRD420251128495). The review examined interventional evidence concerning the effects of artificial intelligence (AI)-supported education on four predefined higher-order cognitive outcomes among health science learners: critical thinking, problem-solving, clinical reasoning, and decision-making. Both conventional and generative AI technologies were eligible. Statistical pooling was not undertaken because substantial clinical and methodological heterogeneity was identified across AI technologies, educational functions, learner populations, intervention designs, comparators, outcome definitions, measurement instruments, and assessment time points. Accordingly, findings were synthesized using a structured narrative synthesis and study-outcome direction-of-evidence classification. Eligibility Criteria Eligibility criteria were defined according to the population, intervention, comparator, outcome, and study-design components of the review question. Population Eligible participants were undergraduate or postgraduate learners enrolled in a health science or health professions discipline, including medicine, nursing, dentistry, physiotherapy, allied health disciplines, and clinical residency or professional training programs. Intervention Studies were required to evaluate an educational or training intervention incorporating AI. AI was operationally defined as a computational system using techniques such as machine learning, deep learning, natural language processing, intelligent tutoring, algorithmic diagnostic or decision support, or generative modelling to perform or support an educational function. Eligible technologies included machine-learning and deep-learning systems, intelligent tutoring systems, AI-enabled chatbots, diagnostic decision-support systems, generative image systems, and large language model (LLM)-based applications. Generative AI was considered a subtype of the broader AI category. Comparator Eligible comparators included standard or conventional instruction, expert or educator feedback, textbook-based learning, conventional information-search resources, alternative educational interventions, other AI-supported approaches, or pre-intervention performance in eligible quasi-experimental designs. Outcomes Studies were required to directly assess at least one of four predefined cognitive outcomes: Critical thinking: purposeful interpretation, analysis, evaluation, inference, explanation, or self-regulatory judgment.Problem-solving: identification or representation of a problem, generation and evaluation of potential solutions, implementation of a response, or revision of an approach.Clinical reasoning: integration and interpretation of clinical information and relevant knowledge to generate, evaluate, or refine hypotheses and clinical judgments.Decision-making: selection of an action or course of action after consideration of available evidence, alternatives, uncertainty, risks, benefits, or contextual factors. Critical thinking: purposeful interpretation, analysis, evaluation, inference, explanation, or self-regulatory judgment. Critical thinking: purposeful interpretation, analysis, evaluation, inference, explanation, or self-regulatory judgment. Problem-solving: identification or representation of a problem, generation and evaluation of potential solutions, implementation of a response, or revision of an approach. Problem-solving: identification or representation of a problem, generation and evaluation of potential solutions, implementation of a response, or revision of an approach. Clinical reasoning: integration and interpretation of clinical information and relevant knowledge to generate, evaluate, or refine hypotheses and clinical judgments. Clinical reasoning: integration and interpretation of clinical information and relevant knowledge to generate, evaluate, or refine hypotheses and clinical judgments. Decision-making: selection of an action or course of action after consideration of available evidence, alternatives, uncertainty, risks, benefits, or contextual factors. Decision-making: selection of an action or course of action after consideration of available evidence, alternatives, uncertainty, risks, benefits, or contextual factors. General knowledge, academic achievement, satisfaction, attitudes, communication, technical skills, confidence, self-efficacy, or global clinical competence were not considered eligible cognitive outcomes unless the study reported a distinct and separable measure corresponding to at least one of the four predefined constructs. Study Design Peer-reviewed quantitative and mixed-methods interventional studies were eligible, including randomized controlled trials, randomized crossover trials, non-randomized controlled intervention studies, and quasi-experimental studies. Only studies published in English were included. Studies were excluded if they were observational, cross-sectional, descriptive, purely qualitative, reviews, editorials, commentaries, protocols, conference abstracts without sufficient outcome data, or studies evaluating the performance of an AI system without a learner-focused educational intervention. Studies for which sufficient methodological or outcome information could not be obtained for reliable eligibility assessment and data extraction were also excluded. Information Sources The primary electronic databases were PubMed, Scopus, and the Cochrane Library. Searches were supplemented by Google Scholar and backward reference-list screening of included studies and relevant systematic reviews. The initial database searches were conducted between September and November 2025. An updated search using the same conceptual eligibility framework was conducted to identify subsequently published studies, with the final search completed on 24 August 2026. For each information source, the exact search date, complete search string, search fields, filters or limits applied, and number of records retrieved are reported in Supplementary Table 1. This information was retained to permit replication of the search process. Search Strategy The search strategy combined three principal concept groups: (1) AI technologies, (2) higher-order cognitive outcomes, and (3) health science learners and education. AI-related search terms included combinations and variations of artificial intelligence, machine learning, deep learning, intelligent tutoring system, chatbot, large language model, generative artificial intelligence, generative AI, ChatGPT, GPT, Gemini, Claude, DeepSeek, and clinical decision support. Cognitive-outcome terms included critical thinking, problem solving, problem-solving, clinical reasoning, diagnostic reasoning, clinical judgment, clinical decision making, and decision-making. These terms were combined with database-specific terms identifying health professions, health science education, students, trainees, residents, and other eligible learners. Controlled vocabulary, where available, was combined with free-text keywords. Search syntax was adapted to the indexing structure and search functionality of each database. Boolean operators, truncation, phrase searching, spelling variants, and database-specific subject headings were applied where appropriate. The complete, database-specific search strategies are provided in Supplementary Table 1 rather than reconstructed from abbreviated search terms, enabling the searches to be independently reproduced. Google Scholar was used as a supplementary search source. For each search combination, the first 200 results ranked by relevance were screened for potentially eligible studies. Searches were conducted using combinations of AI-related terms, the four predefined cognitive outcomes, and health professions education terms. Potentially eligible records were assessed using the same eligibility criteria applied to the primary database searches. Reference lists of all included studies and relevant systematic reviews were examined to identify additional potentially eligible studies. Records identified through supplementary searching were subjected to the same eligibility criteria and study-selection procedures as records identified through the primary database searches. Study Selection All records identified through database and supplementary searching were compiled and duplicates were removed before screening zotore. Study selection was conducted in two stages. First, two reviewers independently screened titles and abstracts against the predefined eligibility criteria. Records considered potentially eligible by either reviewer proceeded to full-text assessment. Second, the same reviewers independently assessed the full texts against the eligibility criteria. At full-text assessment, particular attention was given to two requirements: (1) whether AI constituted part of a learner-focused educational or training intervention and (2) whether the study directly and separately assessed at least one of the four predefined cognitive outcomes. Studies reporting only general academic achievement, knowledge, satisfaction, communication, confidence, technical performance, diagnostic accuracy without an eligible reasoning or decision-making construct, or broad clinical competence were excluded unless a distinct eligible cognitive outcome could be extracted. Disagreements at either screening stage were resolved through discussion and consensus. When consensus could not be reached, a third reviewer was consulted. The number of records identified, duplicates removed, records screened, full-text reports assessed, full-text reports excluded with reasons, and studies included in the final synthesis are reported in the PRISMA 2020 flow diagram (Figure 1). Figure 1PRISMA 2020 Flow Diagram of the Study Selection Process. The flow diagram summarizes the identification, screening, eligibility assessment, and final inclusion of studies in the systematic review. A total of 29 interventional studies were included in the final synthesis.The flowchart outlines the study selection process for a systematic review. It begins with the identification of 1,276 records from databases: PubMed (412), Scopus (643), Cochrane Library (166) and grey literature and other sources (55). After removing duplicates, 812 records remain. These records are screened by title and abstract, resulting in 592 exclusions. The remaining 220 reports are sought for retrieval, with none not retrieved. Full-text reports are assessed for eligibility, with 191 excluded for reasons summarized in Supplementary Table S3. Finally, 29 studies are included in the review, with a total of 2,138 participants across these studies.A flowchart of study selection process for a systematic review, detailing identification to inclusion. PRISMA 2020 Flow Diagram of the Study Selection Process. The flow diagram summarizes the identification, screening, eligibility assessment, and final inclusion of studies in the systematic review. A total of 29 interventional studies were included in the final synthesis. Data Extraction A standardized data-extraction form was developed before extraction to ensure consistent collection of study characteristics and outcomes. Two reviewers independently extracted data from each eligible study and subsequently compared their entries. Discrepancies were resolved by re-examination of the primary publication and consensus. The following information was extracted: Author and publication year;Country;Study design;Health science discipline and learner level;Sample size;AI technology and subtype;Educational function of the AI intervention;Intervention characteristics and duration;Comparator;Cognitive outcome assessed;Outcome definition and measurement instrument;Assessment time point;Principal quantitative findings; andDirection of evidence. Author and publication year; Author and publication year; Country; Country; Study design; Study design; Health science discipline and learner level; Health science discipline and learner level; Sample size; Sample size; AI technology and subtype; AI technology and subtype; Educational function of the AI intervention; Educational function of the AI intervention; Intervention characteristics and duration; Intervention characteristics and duration; Comparator; Comparator; Cognitive outcome assessed; Cognitive outcome assessed; Outcome definition and measurement instrument; Outcome definition and measurement instrument; Assessment time point; Assessment time point; Principal quantitative findings; and Principal quantitative findings; and Direction of evidence. Direction of evidence. Where a study reported more than one eligible cognitive outcome using distinct measures, each study-outcome pair was extracted and synthesized separately. This prevented findings for one cognitive construct from being generalized to another construct assessed within the same study. Sample sizes were extracted as the number of participants included in the analysis relevant to the eligible intervention–comparator comparison. Where studies reported both enrolled or randomized and analyzed samples, the analyzed sample was used to reflect the population contributing data to the synthesized cognitive outcome. When multiple reports described the same participant cohort, they were treated as a single study unless they reported distinct eligible interventions or outcomes. Risk-of-Bias Assessment Risk of bias was evaluated independently by two reviewers using appraisal instruments appropriate to each study design. Disagreements were resolved through discussion and consensus, with consultation of a third reviewer when required. Randomized controlled trials were assessed using the Cochrane Risk of Bias 2 (RoB 2) tool, covering bias arising from the randomization process, deviations from intended interventions, missing outcome data, measurement of the outcome, and selection of the reported result. Overall judgments were classified according to RoB 2 guidance as low risk of bias, some concerns, or high risk of bias. Randomized crossover studies were evaluated using the RoB 2 approach appropriate to crossover trials, including consideration of potential period and carryover effects where applicable. Non-randomized intervention studies were evaluated using ROBINS-I, including bias due to confounding, participant selection, classification of interventions, deviations from intended interventions, missing data, outcome measurement, and selection of the reported result. Quasi-experimental studies not appropriately evaluated using RoB 2 or ROBINS-I were assessed using the relevant Joanna Briggs Institute (JBI) critical appraisal checklist. JBI findings were reported at the item level rather than converted into RoB 2-style overall categories. Methodological characteristics such as small sample size, single-center recruitment, and short follow-up were considered when interpreting the evidence but were not automatically classified as risk-of-bias domains unless specified by the corresponding appraisal instrument. Data Synthesis A structured narrative synthesis was undertaken because statistical pooling was considered inappropriate given the heterogeneity of interventions, technologies, populations, comparators, outcome constructs, measurement instruments, and assessment time points. Studies were initially organized according to the cognitive outcome directly assessed: critical thinking, problem-solving, clinical reasoning, or decision-making. Studies reporting multiple eligible outcomes could contribute to more than one domain, but each eligible outcome was considered separately. For each study-outcome pair, the direction of evidence was classified as positive, null, mixed, or negative according to prespecified operational criteria: Positive: a statistically significant finding favoring the AI-supported intervention for the eligible cognitive outcome.Null: no statistically significant difference between the AI-supported intervention and the relevant comparator, or no significant change where a within-group comparison was the appropriate primary contrast.Negative: a statistically significant finding favoring the comparator or demonstrating poorer performance associated with the AI-supported intervention.Mixed: findings differed across eligible measures, assessment time points, or relevant comparisons within the same cognitive domain, preventing classification as uniformly positive, null, or negative. Positive: a statistically significant finding favoring the AI-supported intervention for the eligible cognitive outcome. Positive: a statistically significant finding favoring the AI-supported intervention for the eligible cognitive outcome. Null: no statistically significant difference between the AI-supported intervention and the relevant comparator, or no significant change where a within-group comparison was the appropriate primary contrast. Null: no statistically significant difference between the AI-supported intervention and the relevant comparator, or no significant change where a within-group comparison was the appropriate primary contrast. Negative: a statistically significant finding favoring the comparator or demonstrating poorer performance associated with the AI-supported intervention. Negative: a statistically significant finding favoring the comparator or demonstrating poorer performance associated with the AI-supported intervention. Mixed: findings differed across eligible measures, assessment time points, or relevant comparisons within the same cognitive domain, preventing classification as uniformly positive, null, or negative. Mixed: findings differed across eligible measures, assessment time points, or relevant comparisons within the same cognitive domain, preventing classification as uniformly positive, null, or negative. Direction-of-evidence classification was based on the eligible cognitive outcome rather than general educational outcomes reported by the study. For example, improvement in knowledge, satisfaction, confidence, or general academic performance was not used to classify the direction of evidence for critical thinking, problem-solving, clinical reasoning, or decision-making. The synthesis additionally considered AI technology, educational function, instructional design, comparator, learner population, measurement approach, follow-up duration, and risk of bias when interpreting differences across studies. Because direction-of-effect classification does not incorporate effect magnitude or precision, these categories were used to organize the narrative evidence rather than to estimate an overall treatment effect. Study characteristics are reported in Table 1, risk-of-bias assessments in Table 2, and study-outcome direction of evidence in Table 3. Table 1Characteristics of Included StudiesStudyYearCountryDisciplineDesignnAI InterventionCognitive Outcome(s)Arkan et al122025TürkiyeNursingRCT96ChatGPT-integrated nursing-process trainingProblem-solvingBrügge et al322024GermanyMedicineRCT21ChatGPT-3.5 patient simulation with structured feedbackClinical reasoning; decision-makingCheng et al82020TaiwanMedicine/RadiologyRCT30Deep-learning hip-fracture diagnostic AIClinical reasoning; diagnostic decision-makingÇiçek et al332025TürkiyeMedicineRCT115ChatGPT-generated feedbackClinical reasoningHan et al362022South KoreaNursingQuasi-experimental61AI chatbot (LandBot)Clinical reasoningHaupt et al372025GermanyDentistryRCT59GPT-4o educational decision supportDecision-makingHou et al242025ChinaHematology residencyNon-randomized controlled60DeepSeek-assisted PBLClinical reasoning; decision-makingJiang et al382024ChinaMedicineRCT61AI chatbot standardized patientClinical reasoningKejingyun & Mingjun212025ChinaNursingRCT100LLM-assisted PBLCritical thinkingPark & Kim392025South KoreaNursingQuasi-experimental72AI tutor-supported simulationCritical thinkingErgezen Sahin et al312025Türkiye/SpainPhysiotherapyRCT40ChatGPT-4-supported educationClinical reasoning; decision-makingShin et al252024South KoreaNursingQuasi-experimental99ChatGPT-assisted learningClinical reasoning; decision-makingSong et al302025ChinaNursingQuasi-experimental132Doubao GenAI-assisted interactive teachingCritical/higher-order thinkingTabuchi et al402024Japan/UKOptometry/Vision ScienceRCT55Stable Diffusion XL-generated educational imagesDiagnostic decision-makingWu et al202020ChinaOphthalmologyRCT38CC-Cruiser AI-tutoring PBLProblem-solving; diagnostic reasoningNissen et al342025GermanyMedicineRandomized crossover154GPT-4 individualized feedbackClinical reasoningWang et al412025ChinaMedicineRCT56Custom GPT simulated patientClinical reasoningRoemer et al422026GermanyMedicineRCT68ChatGPT-4o diagnostic supportClinical reasoning; diagnostic decision-makingGustafsson et al432026FinlandMedicineRandomized pilot41ChatGPT-4o mini during trauma simulationDecision-makingNartthanarung et al442026ThailandMedicineRandomized crossover46ChatGPT versus conventional searchClinical reasoningGu et al452026ChinaOphthalmologyRandomized pilot74DeepSeek-assisted PBLClinical reasoningSutthiviriwan et al462026ThailandNursingQuasi-experimental60ChatGPT-assisted concept mappingCritical thinkingShen et al472026ChinaMedicineRCT80LLM/NLP virtual standardized patientClinical reasoningJi et al482026ChinaOncology residencyRCT124AI-empowered intelligent training systemClinical reasoning; decision-makingLai et al492026ChinaSurgical residencyRCT72DeepSeek with structured reasoning frameworkClinical reasoning; decision-makingSalem et al502026EgyptNursingMixed-methods RCT64AI-powered adaptive ECG simulationClinical decision-makingTang et al512026ChinaMedicineRCT60GenAI diagnostic explanationsDiagnostic reasoning; decision-makingWei et al522026ChinaGastroenterology residencyRCT90DeepSeek-enhanced scenario-based MDT teachingCritical thinkingLi & Zhao532026ChinaMedical imagingRCT110PACS-embedded AI cognitive-support frameworkClinical decision-makingTotal29 studies2138Abbreviations: AI, artificial intelligence; GenAI, generative artificial intelligence; LLM, large language model; MDT, multidisciplinary team; NLP, natural language processing; PACS, picture archiving and communication system; PBL, problem-based learning; RCT, randomized controlled trial. Table 2Risk-of-Bias Assessment of Included StudiesStudyDesignToolMain Appraisal FindingsOverall JudgmentArkan et al12RCTRoB 2Randomized design; participant blinding infeasible; self-reported problem-solving outcomeSome concernsBrügge et al32RCTRoB 2Randomized design; open educational intervention; performance-based outcomesSome concernsCheng et al8RCTRoB 2Randomized comparison; limited reporting of allocation concealment/maskingSome concernsÇiçek et al33RCTRoB 2Randomized design; attrition and open feedback interventionSome concernsHan et al36Quasi-experimentalJBINon-random allocation; potential baseline/cohort confoundingItem-level concernsHaupt et al37RCTRoB 2Randomized multi-arm comparison; intervention/resources visible to participantsSome concernsHou et al24Non-randomized controlledROBINS-IPotential confounding and selection bias arising from non-random allocationModerate riskJiang et al38RCTRoB 2Randomized educational comparison; masking limitationsSome concernsKejingyun & Mingjun21RCTRoB 2Randomized design; intervention awareness and outcome-measurement concernsSome concernsPark & Kim39Quasi-experimentalJBINon-random allocation; possible cohort/instructor effectsItem-level concernsErgezen Sahin et al31RCTRoB 2Randomized design; educational blinding infeasible; attrition considerationsSome concernsShin et al25Quasi-experimentalJBINon-random allocation; potential confoundingItem-level concernsSong et al30Quasi-experimentalJBIControlled pre-post design; non-random allocationItem-level concernsTabuchi et al40RCTRoB 2Randomized comparison; limited masking/concealment reportingSome concernsWu et al20RCTRoB 2Randomized design; limited concealment/masking informationSome concernsNissen et al34Randomized crossoverRoB 2Randomized crossover; potential period/carryover effectsSome concernsWang et al41RCTRoB 2Randomized design; open educational interventionSome concernsRoemer et al42RCTRoB 2Randomized design; objective diagnostic outcomes but intervention visibleSome concernsGustafsson et al43Randomized pilotRoB 2Randomized pilot; unblinded intervention; exploratory outcomesSome concernsNartthanarung et al44Randomized crossoverRoB 2Randomized crossover; possible sequence/carryover effectsSome concernsGu et al45Randomized pilotRoB 2Randomized pilot; masking/concealment reporting limitationsSome concernsSutthiviriwan et al46Quasi-experimentalJBIControlled pre-post design; non-random allocation and short follow-upItem-level concernsShen et al47RCTRoB 2Randomized design; masking limitations inherent to educational interventionSome concernsJi et al48RCTRoB 2Computer-generated randomization; allocation concealment and blinded assessment reportedLow riskLai et al49RCTRoB 2Randomized trial; blinded evaluator and structured assessmentLow riskSalem et al50RCTRoB 2Randomization/concealment reported; retrospective registration raises reporting concernSome concernsTang et al51RCTRoB 2Randomized three-arm design; concealment/masking insufficiently reportedSome concernsWei et al52RCTRoB 2Random allocation; validated outcome; concealment/masking incompletely reportedSome concernsLi & Zhao53RCTRoB 2Randomized controlled design with longitudinal follow-up; masking/concealment require domain-level considerationSome concernsNotes: These are study-level appraisal judgments based on the published methodological reporting, not simply assumptions based on sample size. For JBI studies, I do not force an overall RoB 2-style category because JBI does not prescribe one. Table 3Study-Level Direction of Evidence Across Cognitive OutcomesStudyAI Type/FunctionCognitive Outcome(s)DirectionInterpretationArkan et al12ChatGPT-integrated nursing trainingProblem-solvingPositiveAI-integrated nursing-process training improved problem-solving scores.Brügge et al32LLM patient simulation + structured feedbackClinical reasoning; decision-makingPositiveStructured LLM simulation and feedback improved clinical decision performance.Cheng et al8Deep-learning diagnostic AIClinical reasoning; diagnostic decision-makingPositiveAI-assisted learning improved diagnostic interpretation relative to conventional teaching.Çiçek et al33ChatGPT-generated feedbackClinical reasoningMixedOverall performance was similar, with expert feedback superior for more complex tasks/subgroups.Han et al36AI chatbotClinical reasoningNullNo significant between-group improvement in clinical reasoning competency.Haupt et al37GPT-4o decision supportDecision-makingMixedGPT-4o improved accuracy versus no support but was not clearly superior to alternative resources.Hou et al24DeepSeek-assisted PBLClinical reasoning; decision-makingPositiveAI-supported PBL improved case analysis and diagnostic decision performance.Jiang et al38AI standardized patientClinical reasoningPositiveAI-standardized-patient learning improved clinical reasoning-related performance.Kejingyun & Mingjun21LLM-assisted PBLCritical thinkingPositiveLLM-assisted PBL significantly improved critical-thinking scores.Park & Kim39AI tutor-supported simulationCritical thinkingNullNo significant between-group difference in critical-thinking disposition.Ergezen Sahin et al31ChatGPT-supported physiotherapy educationClinical reasoning; decision-makingNullClinical performance did not significantly differ between groups.Shin et al25ChatGPT-assisted nursing educationClinical reasoning; decision-makingNegativeComparator instruction produced better ethical decision-making and clinical-reasoning performance.Song et al30Doubao GenAI interactive teachingCritical/higher-order thinkingPositiveAI-assisted teaching improved higher-order-thinking outcomes.Tabuchi et al40AI-generated educational imagesDiagnostic decision-makingMixedBoth groups improved without significant superiority of AI-generated-image education overall.Wu et al20AI-tutoring PBLProblem-solving; diagnostic reasoningMixedGreater improvement occurred for signs/diagnosis but not treatment planning.Nissen et al34GPT-4 individualized feedbackClinical reasoningNullAI-generated individualized feedback did not outperform generic expert feedback.Wang et al41GPT simulated patientClinical reasoningPositiveGPT-based simulated-patient training improved structured clinical performance.Roemer et al42ChatGPT-4o diagnostic supportClinical reasoning; diagnostic decision-makingPositiveAI-supported learners demonstrated greater diagnostic performance.Gustafsson et al43Unstructured LLM trauma supportDecision-makingMixedAI did not improve accuracy and slowed responses in some scenarios.Nartthanarung et al44ChatGPT vs conventional searchClinical reasoningMixedBoth approaches improved reasoning without a clear ChatGPT-specific advantage.Gu et al45DeepSeek-assisted PBLClinical reasoningPositiveDeepSeek-assisted PBL improved diagnostic reasoning and clinical proficiency.Sutthiviriwan et al46ChatGPT-assisted concept mappingCritical thinkingPositiveAI-assisted concept mapping produced greater critical-thinking improvement.Shen et al47AI virtual standardized patientClinical reasoningPositiveAI-VSP combined with SP training improved clinical reasoning relative to SP training alone.Ji et al48AI-empowered oncology trainingClinical reasoning; decision-makingPositiveAI-supported training improved diagnostic reasoning and treatment-decision performance.Lai et al49Structured DeepSeek reasoningClinical reasoning; decision-makingPositiveStructured AI interaction improved clinical reasoning/decision quality compared with unguided use.Salem et al50Adaptive AI ECG simulationDecision-makingPositiveAI-supported simulation improved clinical decision-making scores.Tang et al51GenAI diagnostic explanationsDiagnostic reasoning; decision-makingPositiveGenAI-supported diagnostic explanations improved diagnostic performance, with effects varying by explanation style.Wei et al52DeepSeek-enhanced MDT teachingCritical thinkingPositiveIntervention participants achieved higher critical-thinking scores.Li & Zhao53PACS-embedded AI cognitive supportClinical decision-makingPositiveAI-supported instruction improved clinical decision-making competencies with longitudinal assessment. Table 1Characteristics of Included StudiesStudyYearCountryDisciplineDesignnAI InterventionCognitive Outcome(s)Arkan et al122025TürkiyeNursingRCT96ChatGPT-integrated nursing-process trainingProblem-solvingBrügge et al322024GermanyMedicineRCT21ChatGPT-3.5 patient simulation with structured feedbackClinical reasoning; decision-makingCheng et al82020TaiwanMedicine/RadiologyRCT30Deep-learning hip-fracture diagnostic AIClinical reasoning; diagnostic decision-makingÇiçek et al332025TürkiyeMedicineRCT115ChatGPT-generated feedbackClinical reasoningHan et al362022South KoreaNursingQuasi-experimental61AI chatbot (LandBot)Clinical reasoningHaupt et al372025GermanyDentistryRCT59GPT-4o educational decision supportDecision-makingHou et al242025ChinaHematology residencyNon-randomized controlled60DeepSeek-assisted PBLClinical reasoning; decision-makingJiang et al382024ChinaMedicineRCT61AI chatbot standardized patientClinical reasoningKejingyun & Mingjun212025ChinaNursingRCT100LLM-assisted PBLCritical thinkingPark & Kim392025South KoreaNursingQuasi-experimental72AI tutor-supported simulationCritical thinkingErgezen Sahin et al312025Türkiye/SpainPhysiotherapyRCT40ChatGPT-4-supported educationClinical reasoning; decision-makingShin et al252024South KoreaNursingQuasi-experimental99ChatGPT-assisted learningClinical reasoning; decision-makingSong et al302025ChinaNursingQuasi-experimental132Doubao GenAI-assisted interactive teachingCritical/higher-order thinkingTabuchi et al402024Japan/UKOptometry/Vision ScienceRCT55Stable Diffusion XL-generated educational imagesDiagnostic decision-makingWu et al202020ChinaOphthalmologyRCT38CC-Cruiser AI-tutoring PBLProblem-solving; diagnostic reasoningNissen et al342025GermanyMedicineRandomized crossover154GPT-4 individualized feedbackClinical reasoningWang et al412025ChinaMedicineRCT56Custom GPT simulated patientClinical reasoningRoemer et al422026GermanyMedicineRCT68ChatGPT-4o diagnostic supportClinical reasoning; diagnostic decision-makingGustafsson et al432026FinlandMedicineRandomized pilot41ChatGPT-4o mini during trauma simulationDecision-makingNartthanarung et al442026ThailandMedicineRandomized crossover46ChatGPT versus conventional searchClinical reasoningGu et al452026ChinaOphthalmologyRandomized pilot74DeepSeek-assisted PBLClinical reasoningSutthiviriwan et al462026ThailandNursingQuasi-experimental60ChatGPT-assisted concept mappingCritical thinkingShen et al472026ChinaMedicineRCT80LLM/NLP virtual standardized patientClinical reasoningJi et al482026ChinaOncology residencyRCT124AI-empowered intelligent training systemClinical reasoning; decision-makingLai et al492026ChinaSurgical residencyRCT72DeepSeek with structured reasoning frameworkClinical reasoning; decision-makingSalem et al502026EgyptNursingMixed-methods RCT64AI-powered adaptive ECG simulationClinical decision-makingTang et al512026ChinaMedicineRCT60GenAI diagnostic explanationsDiagnostic reasoning; decision-makingWei et al522026ChinaGastroenterology residencyRCT90DeepSeek-enhanced scenario-based MDT teachingCritical thinkingLi & Zhao532026ChinaMedical imagingRCT110PACS-embedded AI cognitive-support frameworkClinical decision-makingTotal29 studies2138Abbreviations: AI, artificial intelligence; GenAI, generative artificial intelligence; LLM, large language model; MDT, multidisciplinary team; NLP, natural language processing; PACS, picture archiving and communication system; PBL, problem-based learning; RCT, randomized controlled trial. Characteristics of Included Studies Abbreviations: AI, artificial intelligence; GenAI, generative artificial intelligence; LLM, large language model; MDT, multidisciplinary team; NLP, natural language processing; PACS, picture archiving and communication system; PBL, problem-based learning; RCT, randomized controlled trial. Table 2Risk-of-Bias Assessment of Included StudiesStudyDesignToolMain Appraisal FindingsOverall JudgmentArkan et al12RCTRoB 2Randomized design; participant blinding infeasible; self-reported problem-solving outcomeSome concernsBrügge et al32RCTRoB 2Randomized design; open educational intervention; performance-based outcomesSome concernsCheng et al8RCTRoB 2Randomized comparison; limited reporting of allocation concealment/maskingSome concernsÇiçek et al33RCTRoB 2Randomized design; attrition and open feedback interventionSome concernsHan et al36Quasi-experimentalJBINon-random allocation; potential baseline/cohort confoundingItem-level concernsHaupt et al37RCTRoB 2Randomized multi-arm comparison; intervention/resources visible to participantsSome concernsHou et al24Non-randomized controlledROBINS-IPotential confounding and selection bias arising from non-random allocationModerate riskJiang et al38RCTRoB 2Randomized educational comparison; masking limitationsSome concernsKejingyun & Mingjun21RCTRoB 2Randomized design; intervention awareness and outcome-measurement concernsSome concernsPark & Kim39Quasi-experimentalJBINon-random allocation; possible cohort/instructor effectsItem-level concernsErgezen Sahin et al31RCTRoB 2Randomized design; educational blinding infeasible; attrition considerationsSome concernsShin et al25Quasi-experimentalJBINon-random allocation; potential confoundingItem-level concernsSong et al30Quasi-experimentalJBIControlled pre-post design; non-random allocationItem-level concernsTabuchi et al40RCTRoB 2Randomized comparison; limited masking/concealment reportingSome concernsWu et al20RCTRoB 2Randomized design; limited concealment/masking informationSome concernsNissen et al34Randomized crossoverRoB 2Randomized crossover; potential period/carryover effectsSome concernsWang et al41RCTRoB 2Randomized design; open educational interventionSome concernsRoemer et al42RCTRoB 2Randomized design; objective diagnostic outcomes but intervention visibleSome concernsGustafsson et al43Randomized pilotRoB 2Randomized pilot; unblinded intervention; exploratory outcomesSome concernsNartthanarung et al44Randomized crossoverRoB 2Randomized crossover; possible sequence/carryover effectsSome concernsGu et al45Randomized pilotRoB 2Randomized pilot; masking/concealment reporting limitationsSome concernsSutthiviriwan et al46Quasi-experimentalJBIControlled pre-post design; non-random allocation and short follow-upItem-level concernsShen et al47RCTRoB 2Randomized design; masking limitations inherent to educational interventionSome concernsJi et al48RCTRoB 2Computer-generated randomization; allocation concealment and blinded assessment reportedLow riskLai et al49RCTRoB 2Randomized trial; blinded evaluator and structured assessmentLow riskSalem et al50RCTRoB 2Randomization/concealment reported; retrospective registration raises reporting concernSome concernsTang et al51RCTRoB 2Randomized three-arm design; concealment/masking insufficiently reportedSome concernsWei et al52RCTRoB 2Random allocation; validated outcome; concealment/masking incompletely reportedSome concernsLi & Zhao53RCTRoB 2Randomized controlled design with longitudinal follow-up; masking/concealment require domain-level considerationSome concernsNotes: These are study-level appraisal judgments based on the published methodological reporting, not simply assumptions based on sample size. For JBI studies, I do not force an overall RoB 2-style category because JBI does not prescribe one. Risk-of-Bias Assessment of Included Studies Notes: These are study-level appraisal judgments based on the published methodological reporting, not simply assumptions based on sample size. For JBI studies, I do not force an overall RoB 2-style category because JBI does not prescribe one. Table 3Study-Level Direction of Evidence Across Cognitive OutcomesStudyAI Type/FunctionCognitive Outcome(s)DirectionInterpretationArkan et al12ChatGPT-integrated nursing trainingProblem-solvingPositiveAI-integrated nursing-process training improved problem-solving scores.Brügge et al32LLM patient simulation + structured feedbackClinical reasoning; decision-makingPositiveStructured LLM simulation and feedback improved clinical decision performance.Cheng et al8Deep-learning diagnostic AIClinical reasoning; diagnostic decision-makingPositiveAI-assisted learning improved diagnostic interpretation relative to conventional teaching.Çiçek et al33ChatGPT-generated feedbackClinical reasoningMixedOverall performance was similar, with expert feedback superior for more complex tasks/subgroups.Han et al36AI chatbotClinical reasoningNullNo significant between-group improvement in clinical reasoning competency.Haupt et al37GPT-4o decision supportDecision-makingMixedGPT-4o improved accuracy versus no support but was not clearly superior to alternative resources.Hou et al24DeepSeek-assisted PBLClinical reasoning; decision-makingPositiveAI-supported PBL improved case analysis and diagnostic decision performance.Jiang et al38AI standardized patientClinical reasoningPositiveAI-standardized-patient learning improved clinical reasoning-related performance.Kejingyun & Mingjun21LLM-assisted PBLCritical thinkingPositiveLLM-assisted PBL significantly improved critical-thinking scores.Park & Kim39AI tutor-supported simulationCritical thinkingNullNo significant between-group difference in critical-thinking disposition.Ergezen Sahin et al31ChatGPT-supported physiotherapy educationClinical reasoning; decision-makingNullClinical performance did not significantly differ between groups.Shin et al25ChatGPT-assisted nursing educationClinical reasoning; decision-makingNegativeComparator instruction produced better ethical decision-making and clinical-reasoning performance.Song et al30Doubao GenAI interactive teachingCritical/higher-order thinkingPositiveAI-assisted teaching improved higher-order-thinking outcomes.Tabuchi et al40AI-generated educational imagesDiagnostic decision-makingMixedBoth groups improved without significant superiority of AI-generated-image education overall.Wu et al20AI-tutoring PBLProblem-solving; diagnostic reasoningMixedGreater improvement occurred for signs/diagnosis but not treatment planning.Nissen et al34GPT-4 individualized feedbackClinical reasoningNullAI-generated individualized feedback did not outperform generic expert feedback.Wang et al41GPT simulated patientClinical reasoningPositiveGPT-based simulated-patient training improved structured clinical performance.Roemer et al42ChatGPT-4o diagnostic supportClinical reasoning; diagnostic decision-makingPositiveAI-supported learners demonstrated greater diagnostic performance.Gustafsson et al43Unstructured LLM trauma supportDecision-makingMixedAI did not improve accuracy and slowed responses in some scenarios.Nartthanarung et al44ChatGPT vs conventional searchClinical reasoningMixedBoth approaches improved reasoning without a clear ChatGPT-specific advantage.Gu et al45DeepSeek-assisted PBLClinical reasoningPositiveDeepSeek-assisted PBL improved diagnostic reasoning and clinical proficiency.Sutthiviriwan et al46ChatGPT-assisted concept mappingCritical thinkingPositiveAI-assisted concept mapping produced greater critical-thinking improvement.Shen et al47AI virtual standardized patientClinical reasoningPositiveAI-VSP combined with SP training improved clinical reasoning relative to SP training alone.Ji et al48AI-empowered oncology trainingClinical reasoning; decision-makingPositiveAI-supported training improved diagnostic reasoning and treatment-decision performance.Lai et al49Structured DeepSeek reasoningClinical reasoning; decision-makingPositiveStructured AI interaction improved clinical reasoning/decision quality compared with unguided use.Salem et al50Adaptive AI ECG simulationDecision-makingPositiveAI-supported simulation improved clinical decision-making scores.Tang et al51GenAI diagnostic explanationsDiagnostic reasoning; decision-makingPositiveGenAI-supported diagnostic explanations improved diagnostic performance, with effects varying by explanation style.Wei et al52DeepSeek-enhanced MDT teachingCritical thinkingPositiveIntervention participants achieved higher critical-thinking scores.Li & Zhao53PACS-embedded AI cognitive supportClinical decision-makingPositiveAI-supported instruction improved clinical decision-making competencies with longitudinal assessment. Study-Level Direction of Evidence Across Cognitive Outcomes Results Study Selection A total of 29 interventional studies involving 2138 participants met the eligibility criteria and were included in the final synthesis. The included studies evaluated conventional and generative AI-supported educational interventions and assessed at least one of the four predefined cognitive outcomes: critical thinking, problem-solving, clinical reasoning, or decision-making. The study-selection process is presented in Figure 1. Characteristics of Included Studies The 29 included studies, comprising 2138 participants, were published between 2020 and 2026 and represented a range of health science disciplines and training contexts, including medicine, nursing, dentistry, physiotherapy, ophthalmology, radiology, optometry/vision science, and postgraduate clinical training. AI technologies varied considerably and included image-classification and diagnostic-support systems, intelligent tutoring systems, fixed or natural language processing-enabled chatbots, generative image systems, LLM-based simulated patients, and conversational LLMs. The educational functions of these technologies included problem-based learning (PBL), case-based learning, diagnostic interpretation, virtual-patient interaction, individualized feedback, simulation, and decision support. Most studies assessed outcomes immediately or within a short period following the intervention, whereas relatively few included delayed assessments. Detailed study characteristics, including study design, learner population, AI technology, educational function, comparator, cognitive outcome, and assessment approach, are presented in Table 1. Risk of Bias Risk-of-bias judgments varied across the included studies and study designs. Randomized and randomized crossover trials were assessed using RoB 2, non-randomized intervention studies using ROBINS-I, and quasi-experimental studies using the relevant JBI critical appraisal checklist. Detailed study-level assessments and domain-specific judgments are presented in Table 2. Evidence by Cognitive Outcome Decision-Making Decision-making findings were mixed. Brügge et al32 reported better clinical decision performance when structured feedback was incorporated into LLM-based patient simulation. Haupt et al37 found that GPT-4o support resulted in greater dental-trauma decision accuracy than no support; however, performance was similar to that achieved using other information resources. Hou et al24 reported favorable diagnostic decision performance with DeepSeek-assisted PBL. In contrast, Shin et al25 reported poorer ethical decision-making and nursing-process reasoning in the ChatGPT-assisted group than in the textbook-based group. Gustafsson et al43 found that unstructured LLM access during trauma simulation did not improve decision accuracy and was associated with slower responses in some scenarios. Tabuchi et al40 reported improvement following AI-generated-image training; however, the overall improvement in diagnostic accuracy did not significantly differ from that achieved with traditional educational approaches. Collectively, decision-making outcomes included favorable, null, mixed, and unfavorable findings across the included interventions. Clinical Reasoning Clinical reasoning was the most frequently represented cognitive outcome and showed substantial variation across studies. Favorable findings were reported for interventions involving AI-assisted image interpretation, structured LLM-based patient simulation, AI-supported PBL, AI-standardized-patient learning, GPT-supported history-taking, rheumatology diagnostic support, DeepSeek-assisted PBL, and combined AI virtual-standardized-patient and standardized-patient training.8,20,24,32 However, several studies reported null or unfavorable findings. Han et al36 found no significant improvement in clinical reasoning competency following a chatbot-based educational intervention. Çiçek et al33 found no overall advantage of ChatGPT-generated feedback over expert feedback on clinical reasoning questions, while Nissen et al34 reported no significant superiority of AI-generated individualized feedback. Ergezen Sahin et al31 similarly reported no significant between-group difference in Mini-CEX performance. Nartthanarung et al44 observed improvement following both ChatGPT and conventional search, without evidence of a significant advantage attributable specifically to ChatGPT. Shin et al25 reported poorer clinical-reasoning performance in the AI-assisted group than in the textbook-based group. Overall, clinical-reasoning findings were favorable in several structured AI-supported interventions but were accompanied by null and unfavorable findings. Problem-Solving Evidence specifically addressing problem-solving was comparatively limited. Arkan et al12 reported significantly improved problem-solving scores following ChatGPT-integrated nursing-process training. Wu et al20 found that AI-tutoring PBL produced greater improvement in learning related to disease signs and diagnosis than traditional lecture-based instruction; however, improvement in treatment planning did not significantly differ between groups. Several other AI-supported simulated-patient and case-based interventions involved activities relevant to problem formulation and case resolution. However, where the reported measures primarily assessed clinical reasoning rather than a distinct problem-solving construct, these studies were synthesized under clinical reasoning rather than problem-solving. Critical Thinking Critical-thinking findings included both favorable and null results. Kejingyun and Mingjun21 reported significantly improved critical-thinking scores following LLM-assisted PBL. Song et al30 reported improvement in higher-order thinking following AI-assisted interactive teaching. Sutthiviriwan et al46 found significantly greater improvement in critical thinking following AI-assisted concept mapping among orthopaedic nursing students compared with standard instruction. In contrast, Park and Kim39 found no significant between-group difference in critical-thinking disposition following AI tutor-supported labor nursing simulation. Overall, favorable effects were observed in several structured AI-supported interventions, although improvement in critical thinking was not consistent across all studies. Overall Direction of Evidence Across the 29 included studies, the direction of evidence varied according to the cognitive outcome assessed, AI technology and educational function, instructional design, and comparator. Favorable findings were identified across the four cognitive domains; however, null, mixed, and unfavorable findings were also observed. Overall, the findings did not demonstrate a uniform effect of AI-supported education across higher-order cognitive outcomes. The study-level direction of evidence, including the AI intervention, cognitive outcome, comparator, principal finding, and classification as positive, null, mixed, or negative, is summarized in Table 3. Discussion This systematic review synthesized evidence from 29 interventional studies involving 2138 health science learners examining AI-supported education in relation to critical thinking, problem-solving, clinical reasoning, and decision-making. Overall, the evidence does not indicate a uniform educational effect of AI. Favorable findings were observed across all four cognitive domains, particularly when AI was incorporated into structured learning activities such as problem-based learning (PBL), simulated-patient encounters, diagnostic exercises, concept mapping, adaptive simulation, and guided feedback. However, null, mixed, and unfavorable findings were also identified. The findings therefore suggest that the educational effects of AI are conditional on the technology and educational function, instructional design, comparator, learner population, assessment approach, and cognitive construct being evaluated. Clinical reasoning was the most frequently represented outcome and demonstrated substantial variation across studies. Favorable findings were reported with AI-assisted diagnostic interpretation, structured LLM-based patient simulation, AI-supported PBL, GPT-based history-taking, rheumatology diagnostic support, and combined virtual and conventional standardized-patient training.8,20,24,32,41,42,47 For example, Cheng et al8 reported improved interpretation of hip fractures following AI-supported education, while Wang et al41 found that GPT-based simulated-patient training improved structured clinical performance compared with conventional role-playing. Roemer et al42 similarly reported improved diagnostic performance when medical students used ChatGPT-4o alongside traditional resources in rheumatology. In the 2026 study by Shen et al,47 combining AI-driven virtual standardized patients with conventional standardized-patient training produced greater improvement in clinical-reasoning scores than standardized-patient training alone. These findings indicate potential value when AI is incorporated into educational activities requiring learners to gather information, interpret cases, generate hypotheses, and apply knowledge in clinically relevant contexts. Nevertheless, favorable clinical-reasoning effects were not consistent. Çiçek et al33 found no overall advantage of ChatGPT-generated feedback over expert feedback, while Nissen et al34 reported no significant superiority of AI-generated individualized feedback. Ergezen Sahin et al31 similarly found no significant between-group difference in Mini-CEX performance. Nartthanarung et al44 reported improvement following both ChatGPT and conventional search without a clear ChatGPT-specific advantage, whereas Shin et al25 reported poorer clinical-reasoning performance in the AI-assisted group than in the textbook-based group. Jiang et al38 also illustrates the importance of distinguishing within-group improvement from comparative effectiveness: clinical-thinking scores improved within the chatbot-supported group, but the between-group difference was not statistically significant. These findings caution against interpreting improvement from baseline as evidence that AI itself produced superior cognitive development. This variability is consistent with recent evidence syntheses. Tay et al16 identified potential benefits of AI-supported interventions for critical thinking and clinical reasoning but emphasized substantial heterogeneity and low-to-very-low certainty of evidence. Reviews focused on AI-supported clinical reasoning have reached similarly cautious conclusions.22,23 Lai et al,10 in a broader synthesis of 66 randomized controlled trials involving 4911 undergraduate health professions learners, likewise reported considerable variation across AI technologies, educational functions, and outcomes, with much of the evidence characterized by limited certainty. Collectively, the current and previous findings indicate that AI should not be conceptualized as a single educational intervention with an inherently positive effect; its educational value depends substantially on how it is incorporated into learning. Decision-making findings were similarly heterogeneous. Structured AI support produced favorable findings in several contexts, including LLM-based patient simulation with feedback, DeepSeek-assisted PBL, adaptive ECG simulation, and workflow-embedded diagnostic support.24,32,50,53 Li and Zhao,53 for example, evaluated a PACS-integrated AI framework that incorporated automated lesion annotation, structured diagnostic prompting, and contextualized feedback and reported superior clinical decision-making competencies with effects maintained during longitudinal assessment. Conversely, Shin et al25 reported poorer ethical decision-making with AI-assisted learning, while Haupt et al37 demonstrated an important comparator-dependent finding: GPT-4o improved performance compared with no support but was not clearly superior to other information resources. Gustafsson et al43 likewise found no consistent improvement in decision accuracy during unstructured LLM-supported trauma simulation. These findings emphasize an important distinction between AI-assisted performance and development of independent decision-making competence. Decision-making requires learners not only to obtain information or recommendations but also to evaluate evidence, uncertainty, risks, benefits, contextual constraints, and the reliability of the information informing an action. Consequently, improved performance while AI assistance is available does not necessarily demonstrate improvement in the learner’s underlying capacity to make decisions independently. This distinction should be considered both when interpreting existing evidence and when designing future assessments. Evidence concerning critical thinking was less extensive but generally favorable in several structured interventions. Kejingyun and Mingjun21 reported improved critical-thinking scores following LLM-assisted PBL, Song et al30 reported improved higher-order thinking with GenAI-assisted interactive teaching, and Sutthiviriwan et al46 found greater improvement in critical thinking following AI-assisted concept mapping. In contrast, Park and Kim39 found no significant between-group difference in critical-thinking disposition following AI tutor-supported labor simulation. These findings suggest that AI may be more likely to support critical thinking when the instructional activity requires learners to organize information, compare alternatives, evaluate evidence, and justify conclusions rather than simply obtain generated responses. This interpretation is consistent with the conceptualization of critical thinking as purposeful and self-regulatory judgment involving interpretation, analysis, evaluation, inference, and explanation.11 AI-supported learning therefore should not itself be considered evidence of critical-thinking development. Learners may need to actively interrogate AI outputs, identify unsupported assumptions, compare competing explanations, evaluate source credibility, and justify whether an AI-generated conclusion should be accepted or rejected.18,19 The relevant educational mechanism may therefore be the cognitive activity required around the AI output rather than exposure to the AI system itself. Problem-solving was comparatively underrepresented. Arkan et al12 reported improved problem-solving scores following ChatGPT-integrated nursing-process training, while Wu et al20 found that AI-tutoring PBL produced greater improvement in learning related to disease signs and diagnosis but not in treatment planning. The limited number of studies directly assessing problem-solving is itself important. Several interventions involved activities that could reasonably be described as problem solving, yet measured diagnostic performance, clinical reasoning, knowledge, or general clinical performance rather than a distinct problem-solving construct. The present review therefore avoided automatically reclassifying these outcomes as problem-solving. This construct-specific approach is important across all four domains. Critical thinking, problem-solving, clinical reasoning, and decision-making are related but not interchangeable. Likewise, diagnostic accuracy, examination performance, knowledge acquisition, communication competence, and technical performance should not automatically be interpreted as evidence of higher-order cognitive development. Where diagnostic performance is used as an indicator of clinical reasoning or decision-making, it should be interpreted as a performance-based proxy unless the study directly measures the underlying reasoning process. This distinction may partly explain apparent inconsistencies across the literature because studies using different assessments may be measuring different aspects of learner performance despite applying similar cognitive labels. Another pattern emerging from the evidence concerns the role of instructional structure. Favorable results were frequently observed when AI was embedded within a defined educational process rather than provided as unrestricted assistance. Examples included structured feedback during LLM patient simulation,32 PBL supported by AI,20,21,24 AI-assisted concept mapping,46 virtual standardized-patient training combined with conventional standardized patients,47 and workflow-integrated diagnostic scaffolding.53 These interventions generally required learners to interact with cases, information, feedback, or reasoning prompts rather than simply receive a final AI-generated answer. This pattern should be interpreted cautiously because the studies were heterogeneous and were not designed to directly compare “structured” versus “unstructured” AI use, but it provides a useful hypothesis for future experimental research. The evidence also highlights the importance of comparator selection. AI may appear effective when compared with no support or passive instruction yet show little or no advantage when compared with expert feedback, conventional information resources, standardized-patient training, or other active educational approaches. Haupt et al,37 Çiçek et al,33 Nartthanarung et al,44 and Jiang et al38 illustrate this issue. Future trials should therefore use educationally credible active comparators whenever the research question concerns whether AI provides added value rather than merely whether learners improve after exposure to it. Educational Implications The findings support the use of structured human–AI interaction in health science education. AI may be most appropriate as a source of simulation, feedback, comparison, prompting, or decision support while learners remain responsible for generating and defending their reasoning. One possible instructional sequence is to require learners to formulate an initial response independently, compare their reasoning with AI-generated output, identify disagreements or unsupported claims, verify clinically important information using authoritative sources, and then provide a justified final response. This approach preserves learner responsibility while using AI to expand opportunities for practice and feedback. Assessment design is equally important. Educators should evaluate the reasoning process in addition to the correctness of the final answer. Written rationales, structured clinical-reasoning tasks, case-based explanations, justification of decisions, and identification of uncertainty may provide stronger evidence of cognitive performance than accuracy scores alone. Where the educational objective is development of an independent cognitive skill, assessment should also include performance without AI assistance. This distinction between assisted and unaided performance is particularly important. If improved performance occurs only while AI is available, it remains uncertain whether the intervention has produced transferable learning or primarily enhanced performance through external cognitive support. Delayed and unaided assessments are therefore needed to determine whether learners retain the relevant reasoning processes after AI assistance is withdrawn. This is especially important given that most studies included in the present review evaluated immediate or short-term outcomes. AI literacy should accompany educational implementation. Learners need to understand that fluent and confident AI-generated responses may still contain inaccurate, incomplete, or contextually inappropriate information.2,3,18 Training should therefore incorporate verification, evaluation of evidence and sources, recognition of uncertainty, responsible use, and appropriate human oversight. The educational objective should not simply be effective use of AI, but the ability to determine when its output should be accepted, questioned, modified, or rejected. Strengths and Limitations A principal strength of this review is its construct-specific approach. Rather than treating general educational performance as evidence of cognitive development, the review distinguished critical thinking, problem-solving, clinical reasoning, and decision-making and synthesized findings according to the cognitive outcome directly assessed or represented by the study measure. Inclusion of both conventional and generative AI also reflects the broader development of AI-supported health science education rather than limiting the evidence base to recent conversational LLMs. Several limitations should be considered. First, substantial heterogeneity existed across AI technologies, educational functions, disciplines, learner populations, comparators, intervention durations, outcome definitions, measurement instruments, and assessment time points. This heterogeneity precluded meaningful statistical pooling and limits direct comparisons across studies. Similar methodological variation has been reported in previous systematic reviews.10,16,22,23 Second, many studies were conducted at single institutions with relatively small samples, and most evaluated immediate or short-term outcomes. Consequently, the evidence regarding retention, transfer, and generalizability remains limited. The rapid evolution of AI technologies introduces an additional challenge because findings obtained with a particular model, version, interface, or prompting approach may not generalize to subsequent systems. Third, measurement of the four cognitive constructs was inconsistent. Some studies employed explicit construct-specific measures, whereas others relied on diagnostic accuracy, structured clinical performance, or related performance indicators as proxies for reasoning or decision-making. Although these measures may capture relevant components of cognition, they do not necessarily provide direct evidence of the underlying cognitive process. This measurement heterogeneity should be considered when interpreting the direction of evidence. Fourth, the direction-of-evidence classification used in this review facilitates structured narrative comparison but does not incorporate effect magnitude, precision, or relative study weight. A positive classification therefore should not be interpreted as equivalent evidence across studies, nor should the number of positive studies be interpreted as a pooled estimate of effectiveness. Risk of bias, comparator quality, measurement validity, and the magnitude and precision of individual findings remain important when interpreting the overall evidence. Finally, evidence concerning sustained independent performance remains particularly limited. Improvement observed while learners use AI cannot establish that the corresponding cognitive skill has been internalized. Conclusions regarding long-term development of critical thinking, problem-solving, clinical reasoning, or decision-making should therefore remain cautious until studies incorporate delayed transfer assessments conducted without AI assistance. Future Research Future studies should define the cognitive construct being investigated before intervention implementation and align the assessment directly with that construct. Validated or well-supported measures should be prioritized, and investigators should distinguish knowledge acquisition, diagnostic accuracy, and immediate assisted task performance from changes in critical thinking, problem-solving, clinical reasoning, or decision-making. Greater reporting transparency is also needed. Studies should specify the AI system, model and version, educational function, prompting strategy, interaction protocol, degree of educator supervision, duration and frequency of exposure, and whether learners were permitted unrestricted interaction. Such information is essential for reproducibility and interpretation as AI systems continue to evolve rapidly. Future trials should use stronger active comparators and examine potential moderators including learner expertise, prior knowledge, AI literacy, previous AI experience, and task complexity. Larger multicenter studies and longer follow-up periods are needed. Most importantly, studies should assess both AI-assisted performance and subsequent unaided performance, preferably through delayed and novel transfer tasks. Such designs would help determine whether AI merely improves performance while assistance is available or contributes to sustained development of independent higher-order cognitive competence. Conclusion AI-supported education may enhance selected higher-order cognitive outcomes in health science learners, particularly when incorporated into structured learning activities that preserve active learner engagement. However, effects are not consistent across technologies, educational designs, comparators, or cognitive outcomes. AI should therefore be implemented as a structured educational support rather than a replacement for independent reasoning or educator guidance. Further rigorous studies with validated outcome measures and longer-term assessments are needed to determine whether these benefits transfer to independent clinical performance.