This systematic review and meta-analysis was conducted in accordance with the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) 2020 expanded checklist [19] and the PRISMA-S (Preferred Reporting Items for Systematic Reviews and Meta-Analyses Literature Search Extension) extension for search reporting [20]. Completed reporting checklists are provided as Checklist 1 (PRISMA 2020 for Abstracts), Checklist 2 (PRISMA 2020 expanded checklist), and Checklist 3 (PRISMA-S extension), with each item mapped to the corresponding manuscript location. The protocol was prospectively registered with the International Prospective Register of Systematic Reviews (PROSPERO; CRD420261284298, registered on January 13, 2026). Patients or the public were not involved in the design, conduct, reporting, or dissemination plans of our research.

Eligibility was determined using the population, intervention, comparator, and outcomes framework [21]. Eligible populations included health care professionals actively engaged in clinical practice or clinical research, whereas medical students and nonclinical administrative staff were excluded. Eligible interventions consisted of AI-powered clinical tools, such as ambient AI documentation systems, CDSS, diagnostic AI, large language models (LLMs), predictive AI, triage systems, and AI-based burnout intervention applications. The comparator included traditional workflows without AI assistance, preimplementation periods, or no comparator for single-arm studies.

Primary outcomes focused on cognitive workload, mental fatigue, and burnout, provided they were measured with validated instruments. Acceptable cognitive workload measures included the NASA-TLX [22], Subjective Workload Assessment Technique [23], or Paas Cognitive Load Scale [24]. Burnout had to be evaluated using established tools such as the Maslach Burnout Inventory (MBI) [25], Oldenburg Burnout Inventory (OLBI) [26], Stanford PFI [27], Copenhagen Burnout Inventory (CBI) [28], or Mini-Z [29]. Secondary outcomes included alert fatigue, automation bias, trust calibration, verification burden, and System Usability Scale [30].

Eligible study designs included randomized controlled trials (RCTs), non-RCTs, prospective and retrospective cohort studies, cross-sectional studies, and pre- and postimplementation studies. While all eligible studies entered the qualitative narrative synthesis, meta-analytic pooling was deliberately restricted to within-AI-category subsets to avoid combining functionally heterogeneous tools. Consequently, all 6 prespecified meta-analytic pools were drawn from the ambient AI documentation subset. Studies evaluating diagnostic imaging AI, CDSS, LLM inbox tools, and AI-based burnout interventions were reported in the narrative synthesis only and were not pooled together. Qualitative-only studies, editorials, commentaries, conference abstracts without full text, and systematic reviews were excluded. Studies measuring only time-based efficiency outcomes without validated workload instruments, studies assessing only diagnostic accuracy, and those evaluating nonclinical AI applications were also excluded.

Our systematic search covered 4 electronic databases spanning January 2015 to January 2026: MEDLINE (via PubMed), Embase (via OVID), Cochrane CENTRAL, and Web of Science Core Collection. Additional sources included manual searching of reference lists, forward citation tracking, and screening of preprint servers (medRxiv and arXiv) for recent studies not yet indexed in bibliographic databases.

The search strategy combined 3 concepts using Boolean operators: AI and clinical AI tools, health care professionals, and cognitive workload or burnout. The complete database-specific search strategies are provided in Multimedia Appendix 1.

Search results were imported into EndNote (version 21; Clarivate Analytics) and deduplicated. Two reviewers (CSB and EJG) independently screened titles and abstracts against eligibility criteria, followed by an independent full-text assessment. Disagreements were resolved through discussion or by consulting a third reviewer (JJL) if consensus could not be reached. The selection process was documented using a PRISMA 2020 flow diagram.

Data were extracted independently by 2 reviewers using a standardized extraction form. For studies with multiple intervention groups or time points, data from all relevant arms were extracted. Authors were contacted for clarification or additional data when necessary. The extracted information included study characteristics (authors, year, country, journal, and study design), population characteristics (sample size, participant type, clinical specialty, and experience level), intervention details (AI system type, clinical domain, and implementation setting), outcome measures (validated instruments used, outcome definitions, and measurement time points), and results (effect estimates, CIs, P values, and pre-post comparisons).

Risk of bias was assessed independently by 2 reviewers using domain-appropriate tools. For RCTs, we used the Cochrane Risk of Bias tool version 2.0 (RoB 2.0) [31], which evaluates bias arising from the randomization process, deviations from intended interventions, missing outcome data, measurement of the outcome, and selection of reported results. For non-RCTs, we used the Risk of Bias in Non-randomized Studies of Interventions (ROBINS-I) tool [32], which evaluates bias due to confounding, selection of participants, classification of interventions, deviations from intended interventions, missing data, measurement of outcomes, and selection of reported results. Each domain was rated as “low risk,” “some concerns,” or “high risk” for RoB 2.0, and “low,” “moderate,” “serious,” or “critical” risk for ROBINS-I. Disagreements were resolved by discussion.

We calculated standardized mean differences (SMDs and Hedges g) for continuous outcomes measured with different scales (eg, NASA-TLX subscales using 0‐10 vs 0‐20 ranges). For continuous outcomes measured with the same instrument, mean differences (MDs) were used. Odds ratios (ORs) were calculated for binary outcomes (eg, burnout prevalence). For pre-post studies, a within-individual correlation of r=0.7 was assumed between baseline and follow-up, accompanied by sensitivity analyses were performed at r=0.5 and r=0.9.

We adopted a narrative synthesis as the primary integrative approach due to anticipated heterogeneity across AI applications, clinical domains, outcome measures, and study designs [33]. Where 2 or more studies reported the same outcome using the same validated instrument in comparable contexts, random-effects meta-analysis was performed. No participant overlap exists across the 6 prespecified meta-analytic pools. NASA-TLX data were pooled on a subscale-by-subscale basis (mental demand, temporal demand, and effort) using Hedges g to standardize different scale ranges (0‐10 and 0‐20). We explicitly avoided pooling aggregate scores from modified versions of the instrument, combining only studies that reported identical subscales on comparable measurement structures. This pooling strategy introduces measurement heterogeneity, which is addressed in the Limitations section. Statistical heterogeneity was assessed using I² (<25%, 25%‐75%, >75% for low, moderate, and high, respectively), complemented by τ². Subgroup analyses were planned by study design, health care professional type, AI system type, and geographic region; however, the small number of studies per outcome precluded these analyses.

Random-effects meta-analyses applied the Hartung-Knapp-Sidik-Jonkman (HKSJ) adjustment [34] uniformly as the primary inferential method for all 6 prespecified pools, irrespective of the number of contributing studies. The between-study variance τ² was estimated by restricted maximum likelihood. We constructed Knapp-Hartung CIs using the 2-tailed t distribution. The variance-inflation factor q* was truncated to a minimum of 1 to ensure the adjusted interval remained at least as wide as the unadjusted random-effects CIs [34]. Prediction intervals (PIs) were calculated for pools containing 3 or more studies [35,36]. For pools with only 2 studies, we reported the pooled point estimate with its 95% Knapp-Hartung CI without a PI. This conservative approach intentionally trades narrow-band precision for protection against type I error inflation.

CIs are used to quantify the precision of the average pooled effect, whereas PIs estimate the expected distribution of true effects across new settings. These metrics are reported alongside each other where estimable. We confirmed that no participants were shared between studies in any meta-analytic pool. The NASA-TLX subscale pools [4,10], the PFI work exhaustion pool [7,8,37], the burnout prevalence pool [4,9,38], and the documentation time pool [10,39] all combine nonoverlapping samples; therefore, no correlation adjustment for shared participants was required. Where pre-post designs contributed continuous outcomes, we calculated effect sizes using change-score SDs derived from pooled pre and post SDs assuming a pre-post correlation of r=0.7 with sensitivity analyses at r=0.5 and r=0.9. For burnout prevalence, included studies [4,9,38] reported paired pre-post counts within the same cohort, and paired ORs were computed (or derived from pooled pre or post 2×2 tables where individual-level data were unavailable). This paired approach is more conservative than treating proportions as independent because it accounts for within-individual correlation.

We did not perform formal assessments of reporting bias using funnel plots or Egger test because all pools contained fewer than the recommended threshold of 10 studies [40]. Instead, the potential for small-study effects was considered qualitatively within the GRADE (Grading of Recommendations Assessment, Development and Evaluation) assessment.

The certainty of evidence for each outcome was assessed using the GRADE approach [41]. Evidence was rated as high, moderate, low, or very low based on 5 domains: risk of bias, inconsistency, indirectness, imprecision, and publication bias. GRADE-Confidence in the Evidence from Reviews of Qualitative Research for qualitative findings was prespecified in the protocol but could not be performed because no qualitative-only study was identified.

Reporting standards and the corresponding completed checklists are described in Checklists 1-3. The protocol deviations described in this section reflect refinements made during the conduct of the review and are reported transparently in keeping with PRISMA 2020 item 5. The systematic review was prospectively registered in PROSPERO prior to data extraction (CRD420261284298, registered on January 13, 2026).

We acknowledge the following protocol deviations. First, a lower-bound search date restriction of January 2015 was applied to focus on the contemporary clinical AI era. Second, medical students were excluded because their clinical exposure to AI tools differs substantially from that of practicing clinicians. Third, qualitative-only studies were excluded, which consequently omitted the planned thematic synthesis and GRADE-Confidence in the Evidence from Reviews of Qualitative Research assessments. Fourth, we additionally searched Cochrane CENTRAL to capture trials potentially unindexed elsewhere. Fifth, burnout was elevated to a coprimary outcome due to its prevalence in the identified literature. Sixth, several prespecified secondary outcomes were omitted because no included studies measured them with validated instruments. Seventh, AI-based burnout intervention applications were retained, as they directly addressed the research question. Finally, the HKSJ adjustment was uniformly adopted as the primary inferential method for all meta-analytic pools, replacing the DerSimonian-Laird estimator.

The systematic search identified 8848 records across 4 databases: PubMed or MEDLINE (n=2055), Embase or Ovid (n=6024), Cochrane CENTRAL (n=65), and Web of Science (n=704). After deduplication in EndNote, 1520 duplicate records were removed. This left 7328 unique records that underwent title or abstract screening, of which 7019 records were excluded. The remaining 309 reports were retrieved and assessed for full-text eligibility. We excluded 288 papers for specific reasons, including narrative review (n=10), study with incomplete data (n=277), and systematic review (n=1). Furthermore, we excluded 2 large ambient AI scribe studies because they did not meet the validated instrument criterion. Specifically, Tierney et al [5] (n=7260) used custom satisfaction surveys without NASA-TLX or standardized burnout measures, and Albrecht et al [6] (n=181) relied on a custom quality improvement survey without validated workload assessments. In total, 21 studies met all inclusion criteria and were included in the qualitative synthesis. The study selection process is presented in Figure 1.

The 21 included studies [4,7-10,16,17,37-39,42-52] were published between 2019 and 2025. In total, 19 (90.5%) of these studies [4,7-10,16,17,39-45,48-52] were published in 2024‐2025, indicating recent expansion in this research area. Studies were conducted across 7 countries: the United States (15/21, 71.4%) [4,7-10,17,37-39,42-45,47,49], Germany (1/21, 4.8%) [16], Australia (1/21, 4.8%) [50], South Korea (1/21, 4.8%) [51], the United Kingdom (1/21, 4.8%) [46], Zambia (1/21, 4.8%) [48], and the United Arab Emirates (1/21, 4.8%) [52]. The total number of health care professionals across all studies was 2885, with sample sizes ranging from 7 to 1430 participants.

Regarding study design, 3 studies were RCTs [7,8,51], which consisted of 1 parallel 3-arm RCT [7], 1 stepped-wedge RCT [8], and 1 single-blind 3-group RCT [51]. In total, 3 further studies used randomized crossover designs [43,50,52], 2 were randomized simulation studies [46,49], and 13 were observational studies [4,9,10,16,17,37-39,42,44,45,47,48]. The observational group included pre-post designs (n=9) [9,10,16,17,37-39,42,45], cross-sectional studies (n=2) [44,47], quality improvement studies (n=1) [4], and mixed methods study (n=1) [48].

Participants included physicians (n=19) [4,7-10,16,17,37-39,42-47,49,50,52], advanced practice providers (APPs; n=9) [4,8-10,38,39,43-45], nurses (n=3) [17,50,51], and radiology trainees (n=2) [47,48]. The evaluated clinical domains covered outpatient documentation (n=12) [4,7-10,37-39,42-45], diagnostic imaging (n=3) [16,47,48], pediatric clinical care (n=4) [17,38,47,49], trauma and transfusion (n=1) [50], inpatient documentation (n=1) [46], nursing practice (n=1) [51], and psychiatric practice (n=1) [52].

AI interventions were categorized into 5 types: ambient AI documentation systems (n=13) [4,7-10,38,39,42-46,52], CDSS (n=3) [17,49,50], radiology AI (n=3) [16,47,48], LLM-based inbox management (n=1) [37], and AI-based burnout intervention (n=1) [51]. The characteristics of included studies are summarized in Table 1.

The NASA-TLX or its derivative instruments were used in 16 (76%) studies [4,7,10,16,17,37-39,42,43,46-50,52]. These applications included the full 6-subscale version [16,17,42,46,49,50,52] or abbreviated versions, such as the 4-item Physician Task Load [7,37] and the 3-item NASA-TLX [4,10,39,43]. Pelletier et al [38] used the NASA-TLX without specifying subscale details. Burnout-specific instruments included the Stanford PFI (n=5) [7-9,42,45], Mini-Z (n=3) [7,10,38], OLBI (n=1) [44], CBI (n=1) [51], and single-item burnout measures (n=2) [4,10]. No study used the full MBI despite its status as the most widely validated burnout measure [25]. The System Usability Scale was used as a secondary outcome in 5 studies [37,39,42,49,50], and the Sheffield Assessment Instrument for Letters was used in 1 study to assess documentation quality [52].

Risk of bias varied substantially across studies. Among the RCTs, the 3-arm RCT by Lukac et al [7] was rated as having “low risk” across all domains. The stepped-wedge RCT by Afshar et al [8] was rated as having “some concerns” due to potential period effects and lack of blinding inherent to the intervention. The 3-group RCT by Baek and Cha [51] was rated as “low risk” with zero dropouts and appropriate randomization.

Among non-RCTs assessed using ROBINS-I [32], 2 studies were rated as “low” risk of bias [49,50], 6 were rated as “moderate” risk of bias [16,17,37,38,43,47], 8 were rated as “serious” risk primarily due to confounding and selection bias [4,9,10,39,42,44,45,48], and 2 simulation studies were rated as “critical” risk due to small sample sizes (n=7 and n=8, respectively) [46,52]. Common methodological limitations included lack of control groups, short follow-up periods, voluntary participation introducing selection bias, and use of abbreviated or modified versions of validated instruments without separate validation. The risk of bias assessment is presented in Figure 2 and Tables 2 and 3.

The certainty of evidence was assessed using the GRADE approach for each outcome-intervention combination (Table 4). The certainty of evidence for cognitive workload reduction with ambient AI documentation was rated as moderate (⊕⊕⊕◯). Evidence from 13 studies (2 RCTs and 11 observational or quasi-experimental) [4,7-10,38,39,42-46,52] consistently demonstrated NASA-TLX reductions of 24-40 points. The evidence was not downgraded for risk of bias, as the 2 RCTs showed low risk and observational studies showed consistent direction of effect. No serious inconsistency, indirectness, or imprecision was identified. Because every pool contained fewer than 10 studies, formal small-study-effect testing was not performed. Consequently, the possibility of small-study effects was considered qualitatively and judged not to warrant downgrading for this outcome.

The certainty for burnout reduction with ambient AI documentation was rated as low (⊕⊕◯◯). In total, 10 studies (2 RCTs and 8 observational) [4,7-10,38,39,42,44,45] reported burnout prevalence reductions of 7‐26 percentage points. Evidence was not downgraded for risk of bias, given RCT support. However, we noted possible small-study effects, given a small literature concentrated on 2 commercial products in which negative or null studies may be underrepresented [40].

The certainty for documentation time reduction with ambient AI documentation was rated as low (⊕⊕◯◯). In total, 8 studies (2 RCTs and 6 observational or quasi-experimental) [7,8,10,38,39,42,43,46] reported time savings of 0.9‐3 minutes per appointment. Evidence was downgraded for serious risk of bias (predominantly observational designs) and serious inconsistency, as effect sizes varied substantially across studies and settings.

The certainty regarding the effects of diagnostic imaging AI on cognitive workload was rated as very low (⊕◯◯◯). In total, 3 observational studies [16,47,48] showed inconsistent results ranging from no significant reduction to increased workload. Evidence was downgraded for serious risk of bias (all observational), serious inconsistency (conflicting directions of effect), and serious imprecision due to small sample sizes.

Similarly, the certainty for CDSS effects on cognitive workload was rated as very low (⊕◯◯◯). In total, 3 observational studies [17,49,50] showed highly variable results, with workload reductions of 7%‐50% in some studies but increases of 14 points in others (sepsis AI). Evidence was downgraded for very serious inconsistency (contradictory findings), serious indirectness (heterogeneous CDSS types and clinical contexts), and serious imprecision.

The certainty for AI-based burnout intervention was rated as low (⊕⊕◯◯). One RCT [51] demonstrated CBI reductions of 20‐23 points, but evidence was downgraded for serious imprecision due to reliance on a single study.

Taken together, the certainty of evidence for ambient AI documentation was limited both because the literature is commercially concentrated (Abridge and DAX Copilot dominate) and because the 95% PI for both burnout pools crosses the null effect line. Evidence for diagnostic imaging AI and CDSS remains very low certainty, with unintended workload increases observed in single studies of CADe imaging [16] and alert-based sepsis CDSS [17]. Documentation time benefits among ambient AI adopters carry low certainty (SMD −0.24, 95% CI −1.10 to 0.61).

This review set out to answer 3 questions, and our main findings map onto each. First, we asked how much AI changes the mental effort and strain that clinicians experience when it is built into their work, measuring this with well-established questionnaires. Combining the available studies, we found that AI used to automatically draft clinical notes (“ambient AI documentation”) was generally linked to less mental strain and less burnout. Of the 6 outcomes we evaluated, 4 showed a clear benefit. Two additional outcomes, mental demand and documentation time, also indicated potential benefits but lacked statistical conclusiveness due to the limited number of studies available for pooling. Second, we asked whether the effect depends on the type of AI, and it clearly did. Note-drafting AI tended to reduce workload and burnout, whereas AI systems designed to interpret medical images or issue decision-support alerts produced mixed results and, in some cases, unintended increases in clinician workload [16,17]. Third, we asked about subtler costs, including the effort required to double-check AI output and the tendency to overtrust it. None of the 21 studies measured these constructs directly with validated tools. We therefore discuss them as ideas to guide future research rather than as measured results.

This review brings validated workload questionnaires together with a conservative pooling method and PIs, which estimate how the effect might vary in new settings rather than only how precise the average is. Its main value lies in mapping the evidence across 21 studies (2885 health care workers in 7 countries, covering 5 kinds of AI: note-drafting tools, image-interpretation AI, on-screen decision-support alerts, LLM tools that draft inbox replies, and AI aimed at reducing burnout), and in grading how trustworthy the evidence is for each type of AI. The risk-of-bias assessments and the certainty-of-evidence summary are presented in Tables 2-4. Because only 2 to 3 studies could be combined for any single outcome, the pooled numbers are best read as support for this broader map of the evidence, not as conclusions on their own.

Our findings should be interpreted within 3 constraints that bear on how much weight they can carry [35,36]. First, the benefits we observed for ambient AI documentation were not equally firm across outcomes. When the available studies were combined, the reductions in effort, work exhaustion, and burnout were the most consistent. In contrast, the apparent benefits for mental demand and documentation time rested on an insufficient number of studies to be considered established. Furthermore, the studies evaluating mental demand exhibited sufficient variability to preclude treating this reduction as a definitive finding [34]. Even for the outcomes that could be examined most fully, the average effect favored ambient AI, though its size in a new clinical setting could range from substantial to small.

A further and equally important point is that these results speak only to one kind of AI: every pooled outcome came from ambient AI documentation tools. AI as a category-wide remedy for clinician burnout is therefore not supported. The other technologies in this review, including diagnostic imaging AI, CDSS, the LLM inbox tool, and the AI burnout intervention, are discussed narratively and were not pooled. Second, most of the evidence comes from study designs that are more prone to bias. In total, 18 of the 21 studies were non-RCTs, with only 2 RCTs evaluating ambient AI [7,8]. Some of the most pronounced benefits were derived from very small simulation studies [46,52]. Because participants were largely volunteers and early adopters, the benefits may be overstated. Third, when we formally graded how trustworthy the evidence is using GRADE [41], only the reduction in cognitive workload with ambient AI documentation reached moderate confidence. Reductions in burnout and documentation time were graded as low confidence, and the evidence supporting diagnostic imaging AI, CDSS, and the single-trial AI burnout intervention was rated as very low confidence. Whether clinical AI improves the working life of the health care workforce overall (once accuracy, downstream patient safety, and long-term adaptation are taken into account) is not yet established.

Our findings align with and extend the 2024 systematic review [15], which identified the absence of cognitive workload assessment as a notable gap in health care AI implementation research. While previous reviews focused primarily on AI diagnostic performance and efficiency metrics [15], our review specifically quantified the subjective cognitive experience of clinicians using validated instruments. The observed NASA-TLX reductions of 24‐40 points with ambient AI scribes exceed the minimally important difference threshold of 10 points suggested in human factors literature, indicating clinically meaningful cognitive burden relief [22].

The unintended workload increase observed with sepsis prediction AI [17] (NASA-TLX 43→57) corroborates the seminal work of Bainbridge [12] on the “ironies of automation,” predicting that systems designed to reduce human workload often create new cognitive demands through vigilance and oversight requirements. Similarly, the lack of workload reduction despite increased reading time with prostate MRI AI [16] supports the “out-of-the-loop” performance problem [13], where monitoring an automated system erodes the operator’s situation awareness and capacity to detect the very errors the automation was intended to prevent [14].

The direction of our pooled ambient-documentation findings is concordant with the larger primary literature that did not qualify for meta-analysis. Uncontrolled and pre-post evaluations of commercial ambient scribes have repeatedly reported reduced documentation burden, lower work exhaustion, and improved professional fulfillment [4,9,42,44,45]. Furthermore, a dedicated cognitive-load evaluation of an ambient platform reported reductions in subjective workload that align with the NASA-TLX effort effect observed in our synthesis [43]. These reports, however, share the structural features that constrain our pooled estimate. They are concentrated on a small number of commercial products, conducted in voluntary early-adopter cohorts, and predominantly use uncontrolled pre-post designs carrying a serious risk of bias under ROBINS-I [32]. Their convergence should therefore be read as consistent but low-certainty evidence rather than as confirmation of a robust effect, which is the reason the GRADE certainty for burnout reduction with ambient AI documentation was rated low rather than moderate [41].

The 2 RCTs available in this field [7,8] provide the least biased evidence; yet, both evaluated a single product over a short horizon, and neither measured downstream documentation accuracy or patient-safety end points. This scarcity of randomized data, combined with the small number of studies available for each outcome [34] and the fact that benefit could not be assumed in new settings even where the evidence was strongest [35], is the reason our synthesis stops short of asserting a generalizable benefit. The largest workload reductions originated from very small simulation studies [46,52], whose effect sizes are statistically fragile. This pattern suggests the potential presence of small-study effects [40] and reinforces the cautious interpretation demanded by the GRADE assessment.

In contrast to the ambient-documentation literature, evidence from diagnostic imaging AI and CDSS is heterogeneous and, in several reports, contrary to expectations. CADe for prostate MRI produced no workload reduction despite improved diagnostic performance [16], a pediatric sepsis prediction model increased perceived workload and alert fatigue [17], and a mobile CDSS tool improved guideline adherence without lowering mental workload [49]. Conversely, a knowledge-recommender embedded in radiology reporting reduced cognitive load [47]. This divergence is consistent with the imaging review that first identified the near-absence of workload measurement in this domain [15] and with automation-science predictions that alert- and flag-based systems impose monitoring and verification demands distinct from those of generative tools [12-14]. Because no 2 of these studies measured the same construct with the same instrument in the same setting, they could not be pooled, and the certainty of evidence for imaging AI and CDSS outcomes was correspondingly rated very low [41]. This inconsistency is best interpreted as genuine clinical and methodological heterogeneity rather than as evidence of no effect.

Two further categories of AI could likewise not be pooled and therefore described narratively. An LLM tool that drafts replies to patient inbox messages was evaluated in a single pre-post study, in which perceived task load fell substantially and work exhaustion improved over 5 weeks [37]. Because only one study examined this application, the finding is promising but cannot be generalized. It highlights a specific application of AI that differs fundamentally from both clinical note generation and image interpretation. Finally, one RCT examined AI not as a clinical tool but as a means of delivering a personalized burnout intervention. A smartphone app for nurses produced marked reductions in client-related and personal burnout relative to control over 4 weeks [51]. This study addresses clinician burnout from a different perspective, using AI to support well-being rather than to assist clinical work. Although its single-trial evidence was graded as low certainty, it illustrates that the role of AI in mitigating burnout extends beyond documentation tools.

The AI categories included in this review differ fundamentally in function and in the clinician’s role with the tool. Generative AI tools, such as ambient AI scribes and LLM inbox-message drafts, produce clinician-facing draft text that the user reviews, edits, and approves. Conversely, discriminative AI tools, such as CDSS alerts and CADe diagnostic-imaging flags, push specific alerts that the clinician must verify against ground truth at the point of decision. The AI-based burnout intervention RCT delivers a psychological intervention outside the clinical workflow. We refer to the cognitive cost of reviewing, validating, and reconciling AI-generated outputs with clinical judgment as “verification burden.” As AI systems become ubiquitous, clinicians transition from content generation to content verification, which may explain the unintended workload increases observed with discriminative-AI tools. This framework, drawn from classical automation theory [12-14], is offered as hypothesis-generating. Direct measurement of these constructs represents a critical priority for future research.

A potentially important implication of this review, given the certainty caveats, is that AI implementation success cannot be reduced to diagnostic accuracy or time efficiency alone. Our finding that some AI systems unexpectedly increase cognitive workload despite improving efficiency challenges the prevailing assumption that “more AI equals better outcomes.” This shift toward supervisory review is a role that human cognitive architecture may be ill-suited to sustain [11,12]. This evidence suggests that regulatory bodies such as the Food and Drug Administration and Conformité Européenne marking authorities should consider incorporating human factors evaluation, including validated cognitive workload assessment, into the AI medical device approval process. Current regulatory frameworks focus predominantly on algorithmic performance metrics, potentially overlooking the real-world cognitive demands imposed on end users. Similarly, health care institutions implementing AI tools should routinely assess cognitive workload and burnout outcomes alongside traditional efficiency metrics [53,54].

The rapid adoption of ambient AI documentation reflects an ongoing transformation in clinical practice, as physician AI use doubled from 38% to 66% between 2023 and 2024 [1]. Our pooled estimates favor ambient AI documentation in early-adopter outpatient cohorts, but the size of this benefit may vary across settings, and the certainty of the evidence is low to moderate. Therefore, the present evidence supports cautious, evaluation-accompanied adoption rather than category-wide endorsement. Health care systems are increasingly viewing ambient AI scribes as one possible component of broader strategies for the clinician burnout crisis, although high-certainty evidence on the workforce-level impact of such adoption is not yet available [2].

This review has several strengths. We conducted a comprehensive search across 4 databases and included only studies that measured cognitive workload or burnout with validated instruments. We applied a rigorous risk of bias assessment using domain-appropriate tools, specifically RoB 2.0 for RCTs and ROBINS-I for non-RCTs, as detailed in Tables 2 and 3. Furthermore, we used a deliberately cautious approach to combining results across all outcomes to avoid overstating uncertain findings. We provided a clear grading of evidence certainty for each outcome, presented in Table 4, and included forest plots for every pool in Figures 3-8. Additional checks confirmed that our main conclusions did not depend on the specific analytical assumptions applied in the synthesis.

Several caveats apply to the interpretation of our findings. The positive evidence for ambient AI documentation rests on a narrow base. It is concentrated on 2 commercial products (Abridge and DAX Copilot), evaluated predominantly among volunteers and early adopters in US outpatient primary care settings, and followed for a short time (typically 8‐12 weeks). Because each outcome could be based on only a few studies, these pooled results are best read as preliminary rather than definitive, and the benefits for 2 of the 6 outcomes are no longer clear-cut once this limited evidence is taken into account. For this reason, we judged the certainty of the evidence for burnout reduction to be low rather than moderate, given the reliance on observational studies and a small literature dominated by 2 products with uniformly positive results. Throughout, claims are framed cautiously, with ambient AI documentation described as “associated with reductions in” cognitive workload and burnout rather than as reliably reducing them.

Finally, none of the 21 studies directly measured the subtler costs of working with AI. These unmeasured costs include the effort required to verify its output, the tendency to overtrust the system, and the subsequent loss of vigilance. As a result, the original question behind this review (whether AI genuinely reduces the mental burden of clinicians or simply shifts the demand from documentation to verification) cannot be answered with the present evidence and remains a priority for future research.

The following specific limitations should also be acknowledged. First, each outcome could be based on only a small number of studies, which limits how precisely the effects can be estimated, prevents the planned subgroup and related exploratory analyses [55], and leaves considerable uncertainty. Even for the 2 outcomes that could be examined most fully, work exhaustion and burnout prevalence, the average effect favored ambient AI, although the magnitude of this effect may differ in a new clinical setting. Second, the studies of some outcomes, most notably mental demand, varied considerably among themselves. This variability likely occurred because the studies used different measurement scales and were conducted across distinct clinical environments. Furthermore, combining results from abbreviated or modified versions of the same questionnaire adds further inconsistency [22,56,57]. Third, no study used the full MBI, even though it is the most thoroughly validated measure of burnout [25]. Fourth, we cannot rule out that smaller studies reported more favorable results than larger ones. Among the possible explanations, publication bias remains plausible because negative or null studies may be underrepresented in this rapidly commercializing field [40].

The current evidence base, while suggestive of meaningful benefits from ambient AI documentation, remains preliminary, product-concentrated, and predominantly composed of short-term, single-institution evaluations conducted in early-adopter settings. Although ambient AI scribes appear to reduce documentation time and burnout reports, the present evidence does not allow us to verify documentation accuracy, downstream patient-safety end points, or net benefit once verification burden, automation bias, and long-term cognitive adaptation are accounted for.

Several research priorities must be addressed before strong clinical or policy recommendations can be formulated. First, future studies should use RCTs with preregistered protocols that pair clinician-reported workload and burnout outcomes with downstream patient-safety, documentation-accuracy, and diagnostic-error end points. Currently, only 4 of the 21 included studies used registered protocols. Second, prospective cohorts extending beyond 12 months are necessary to detect deskilling, adaptation, and longer-term cognitive effects. Third, researchers must prioritize the standardized measurement of verification burden, automation bias, automation complacency, and trust calibration using validated instruments. While narrower constructs such as trust, usability, and alert fatigue were assessed in some studies, the broader cognitive-cost framework remains unmeasured. Fourth, head-to-head comparative effectiveness studies across multiple commercial ambient AI scribe products, beyond early market leaders, are needed. Fifth, investigations should expand to non-US health care systems and specialties outside outpatient primary care, where workflow requirements and clinician burnout drivers may differ. Finally, future analyses should decompose the NASA-TLX into its specific subscales rather than relying on total scores, allowing for the precise identification of which cognitive domains AI affects most. Without such evidence, the workforce-level benefit of ambient AI scribes cannot yet be confirmed.

This systematic review and meta-analysis is, to our knowledge, the first to combine validated cognitive-workload instruments with uniformly adjusted random-effects meta-analysis and PIs in human-AI interaction in health care. It is also the first to grade the certainty of evidence separately for 5 distinct AI categories: ambient documentation, diagnostic imaging AI, CDSS, LLM inbox tools, and AI-based burnout interventions. Whereas prior systematic reviews of health care AI have focused on diagnostic accuracy, efficiency, or implementation outcomes, the present synthesis quantifies the subjective cognitive experience and burnout of clinicians using validated instruments and grades the certainty of every quantitative outcome.

The contribution to the field is therefore a structured certainty-graded evidence base for the cognitive and burnout consequences of clinical AI. Ambient AI documentation is associated with reductions in NASA-TLX effort, PFI work exhaustion, and burnout prevalence in early-adopter cohorts; however, the evidence is still limited and uncertain. The observed benefits rest on a small number of studies, the effect sizes may not be uniform across settings, and technologies such as diagnostic imaging AI and CDSS have demonstrated context-dependent or unintended workload increases.

The real-world implications are immediate and concrete. Institutional pilots of ambient AI should be paired with prospective, validated-instrument measurement of cognitive workload and burnout. Regulatory human-factors review must be integrated alongside algorithmic-performance evaluation during AI medical-device approval. Furthermore, AI tool developers should incorporate human-centered design that anticipates verification burden as a primary outcome, a metric that was not directly measured in any included study. Until multiproduct, longer-term, multiregion evidence directly measures verification burden, automation bias, and downstream patient-safety end points, the net benefit of clinical AI on the health care workforce remains an open empirical question.