AI Learning Evidence

Not guidance. A record of published research on AI-assisted learning, and our assessments of it. The limits

Effects of Generative AI–Supported Virtual Reality Dental Skill Training on Learning Performance, Visual Behavior, and Cortical Activation Among Dental Students: Stratified Randomized Controlled Trial

Design
Randomized trial · randomized at student level · 123 participants
Subject and task
Other
Seven daily one-hour sessions of tooth-preparation practice on a desktop VR simulator with haptic joysticks (Zhonghui). Day 7 sessions also carried eye-tracking (15-minute exploratory task) and an fNIRS block paradigm (6 x 50-second operative blocks). The AI arm put questions to ChatGPT on a separate computer using standardized prompt templates and an operation guide; the control arm asked on-site professional instructors.
Exposure: 7 days, one hour per day (about 7 hours of training); calendar dates not reported
Population match
Secondary students: Different · University students: Direct
Attrition
129 recruited; 6 excluded for failing eye-tracking calibration (stated in Results; the Methods places calibration before the experiment, so this appears to be pre-randomization); 123 randomized and analysed (62/61). No post-randomization dropouts or exclusions are reported in the text; the CONSORT diagram (Figure 3) is an image not in the bundle, so the eligible and randomized counts cannot be checked against it. Subgroup ns per stratum and arm are not in the text.
Study tier
Study tier 3For the primary outcome only (F1). In favour: individual stratified randomization, 123 analysed with no reported post-randomization loss, a declared primary outcome with a cited registration, two raters stated to be masked, and summary statistics for F1 that recompute exactly (Welch t 3.88, df about 110.6, d 0.70, 95% CI 4.81 to 14.91 from the reported means, SDs and ns). Against: (1) randomization method and concealment unreported; (2) the comparison is ChatGPT-plus-unmasked research assistants who audited outputs versus instructors of unreported number, qualification and identity, with the amount of guidance received unrecorded in both arms, so "AI vs human" is not a clean contrast and differential attention cannot be excluded; (3) conditions of the operative test are not described: the paper never says whether ChatGPT or instructors were accessible during it (the only relevant sentence is a general ban on AI use outside the learning tasks), nor whether it was on the VR simulator or another medium, nor exactly when it was taken; (4) single site, 7 hours of training, one basic task, immediate test only, from a group with a prior positive ChatGPT-instructor trial; (5) registration timing unverified. The secondary tables contain reporting errors serious enough to lower trust in the analysis overall: the SRL and engagement means/SDs imply near-total separation of the arms (d about 2.4 and 3.1 for average- and low-achieving SRL; d 1.3-2.0 for engagement) yet the reported U statistics, r and P values correspond to modest effects; the control-arm Dunn effect sizes in Table 1 (0.18, 0.23, 0.41) do not follow from their z values (expected about 0.32, 0.40, 0.72); the anger CI is printed with the wrong sign; and the Discussion and Conclusions describe timed visual-spatial tests as “self-rated”. None of these touches the F1 arithmetic directly, but together they make a single-trial benefit of d 0.70 on a rater-scored, same-task test LOW, not MODERATE. Secondary outcomes (process, physiological and self-report measures, and post-only visual-spatial tests) are VERY_LOW as learning evidence.
Assessment
Version 1 · AI: two passes agreed · Nov 11, 2026

Study facts come from the paper. Population match and study tier are our judgments, made separately for the two populations this collection covers.

At one Chinese dental school, 123 fourth- and fifth-year students were randomly assigned (within three pretest-score bands) to practise tooth preparation on a VR simulator for one hour a day for seven days with either ChatGPT answering their questions or on-site instructors doing so. Research assistants watched the ChatGPT answers throughout. On a 30-minute practical test scored by two experts said to be unaware of group, the ChatGPT group scored 77.3 vs 67.4 out of 100 (difference 9.9 points, 95% CI 4.8 to 14.9; d 0.70). The paper does not say whether ChatGPT was available during that test, how randomization was done, or how much help either group actually used. Eye-tracking, brain-imaging and questionnaire results also favoured the ChatGPT group, but these are process or self-report measures, and several of the secondary tables contain numbers that do not agree with each other. One small, short, single-site trial: low-certainty evidence for an immediate skill benefit.

01Findings

What the study measured

Arms

ArmWhat it got
A · GenAI-assisted: ChatGPT guidance during VR skill training, no human tutoring (research assistants auditing outputs)
n = 62
Intervention · teacher supervised
ChatGPT (OpenAI) · GPT-4-class (stated as ChatGPT-4.0) · Prompted pedagogy · chat interface · in class · used not_reported
B · Teacher-led control: on-site professional instructor guidance during the same VR skill training
n = 61
Active comparison · teacher led
Identical VR tasks, equipment and one hour a day for 7 days; unlimited opportunity to ask on-site professional instructors. Number, identity and qualifications of instructors, instructor-student ratio and number of consultations not reported.

Findings

A finding is one outcome contrast: two arms, one measure, one set of conditions. A paper is never "positive" or "negative" here; each finding carries its own direction and its own conditions. The badge that matters most is the first one: whether the AI was still available when the outcome was measured. Why that gates everything.

Finding and conditionsResult
F1 · 30-minute practical operation test (basic tooth preparation), scored 0-100 by two experts using rubrics adapted from the Chinese Dental Licensing Examination (primary outcome)
GenAI-assisted: ChatGPT guidance during VR skill training, no human tutoring (research assistants auditing outputs) vs Teacher-led control: on-site professional instructor guidance during the same VR skill trainingAI availability at assessment not reported — cannot carry a learning claimImmediate post-testSame contentGraded artifactPrespecification unclear
Benefit
Mean 77.25 (SD 11.86) AI vs 67.39 (SD 16) control; difference +9.86 points on 0-100; Cohen d 0.70
4.81 to 14.91 (95% CI, mean difference)
Welch t(110.56) = 3.88, P < .001 (unadjusted, as stated for the primary outcome)
Licensing-exam-derived rubric scored by two raters (interrater r=0.89, a correlation, not an agreement index); raters stated masked, method not described; the authors concede possible subjective scoring bias. Test medium, conditions and AI access not reported. No baseline skill score.
F10 · Self-regulated learning (adapted Pintrich and De Groot SRLS, 8 items, 5-point), posttest, high-achieving stratum
GenAI-assisted: ChatGPT guidance during VR skill training, no human tutoring (research assistants auditing outputs) vs Teacher-led control: on-site professional instructor guidance during the same VR skill trainingAI availability at assessment not reported — cannot carry a learning claimImmediate post-testTransfer not assessedSelf-reportPrespecification unclear
Null
32.45 (SD 2.12) AI vs 31.21 (SD 1.95) control; U = 145, r = 0.29; baseline 28.98 vs 29.01
–0.27 to 8.75 (95% CI, median difference)
Holm-adjusted P = .06
F11 · Self-regulated learning (SRLS), posttest, average-achieving stratum
GenAI-assisted: ChatGPT guidance during VR skill training, no human tutoring (research assistants auditing outputs) vs Teacher-led control: on-site professional instructor guidance during the same VR skill trainingAI availability at assessment not reported — cannot carry a learning claimImmediate post-testTransfer not assessedSelf-reportPrespecification unclear
Benefit
30.36 (SD 1.89) AI vs 25.73 (SD 2.03) control; U = 129, r = 0.36; baseline 23.10 vs 23.45
0.36 to 8.90 (95% CI, median difference)
Holm-adjusted P = .04
F12 · Self-regulated learning (SRLS), posttest, low-achieving stratum
GenAI-assisted: ChatGPT guidance during VR skill training, no human tutoring (research assistants auditing outputs) vs Teacher-led control: on-site professional instructor guidance during the same VR skill trainingAI availability at assessment not reported — cannot carry a learning claimImmediate post-testTransfer not assessedSelf-reportPrespecification unclear
Benefit
26.52 (SD 1.99) AI vs 20.39 (SD 1.92) control; U = 113, r = 0.42; baseline 18.51 vs 17.92
2.14 to 10.12 (95% CI, median difference)
Holm-adjusted P = .02
F13 · Enjoyment, single 6-point item, postintervention
GenAI-assisted: ChatGPT guidance during VR skill training, no human tutoring (research assistants auditing outputs) vs Teacher-led control: on-site professional instructor guidance during the same VR skill trainingAI availability at assessment not reported — cannot carry a learning claimImmediate post-testTransfer not assessedSelf-reportPrespecification unclear
Benefit
4.78 (SD 0.85) AI vs 3.98 (SD 0.79) control; Cohen d 0.98
0.51 to 1.09 (95% CI, mean difference)
Welch t(119.20) = 5.41; Holm-adjusted P < .001
F14 · Anger, single 6-point item, postintervention
GenAI-assisted: ChatGPT guidance during VR skill training, no human tutoring (research assistants auditing outputs) vs Teacher-led control: on-site professional instructor guidance during the same VR skill trainingAI availability at assessment not reported — cannot carry a learning claimImmediate post-testTransfer not assessedSelf-reportPrespecification unclear
Benefit
2.34 (SD 0.93) AI vs 3.30 (SD 0.91) control; difference –0.96; Cohen d 1.05
printed as –1.29 to 0.63; from the reported summaries it should be about –1.29 to –0.63 (sign typo)
Welch t(120.97) = –5.80; Holm-adjusted P < .001
F15 · Frustration, single 6-point item, postintervention
GenAI-assisted: ChatGPT guidance during VR skill training, no human tutoring (research assistants auditing outputs) vs Teacher-led control: on-site professional instructor guidance during the same VR skill trainingAI availability at assessment not reported — cannot carry a learning claimImmediate post-testTransfer not assessedSelf-reportPrespecification unclear
Null
1.89 (SD 1.13) AI vs 2.22 (SD 0.89) control; Cohen d 0.32
–0.69 to 0.03 (95% CI, mean difference)
Welch t(120.73) = –1.80; Holm-adjusted P = .07
F16 · Behavioral engagement (Reeve and Tseng engagement scale subscale), post
GenAI-assisted: ChatGPT guidance during VR skill training, no human tutoring (research assistants auditing outputs) vs Teacher-led control: on-site professional instructor guidance during the same VR skill trainingAI availability at assessment not reported — cannot carry a learning claimImmediate post-testTransfer not assessedSelf-reportPrespecification unclear
Benefit
21.29 (SD 1.95) AI vs 18.54 (SD 2.33) control; U = 2344, r = 0.207
0.40 to 5.65 (95% CI, median difference)
Holm-adjusted P = .03
F17 · Emotional engagement (Reeve and Tseng subscale), post
GenAI-assisted: ChatGPT guidance during VR skill training, no human tutoring (research assistants auditing outputs) vs Teacher-led control: on-site professional instructor guidance during the same VR skill trainingAI availability at assessment not reported — cannot carry a learning claimImmediate post-testTransfer not assessedSelf-reportPrespecification unclear
Benefit
17.42 (SD 1.55) AI vs 14.78 (SD 1.29) control; U = 2371, r = 0.219
0.54 to 5.78 (95% CI, median difference)
Holm-adjusted P = .03
F18 · Cognitive engagement (Reeve and Tseng subscale), post
GenAI-assisted: ChatGPT guidance during VR skill training, no human tutoring (research assistants auditing outputs) vs Teacher-led control: on-site professional instructor guidance during the same VR skill trainingAI availability at assessment not reported — cannot carry a learning claimImmediate post-testTransfer not assessedSelf-reportPrespecification unclear
Benefit
28.67 (SD 1.92) AI vs 24.39 (SD 2.40) control; U = 2457, r = 0.258
1.51 to 7.62 (95% CI, median difference)
Holm-adjusted P = .01
F2 · Change in pupil diameter from session-start median during learning on day 7 (eye-tracker, mm), used by the authors as a cognitive-load proxy
GenAI-assisted: ChatGPT guidance during VR skill training, no human tutoring (research assistants auditing outputs) vs Teacher-led control: on-site professional instructor guidance during the same VR skill trainingWith AI at assessment — assisted performance, not learning evidenceDuring interventionTransfer not assessedPhysiologicalPrespecification unclear
Unclear
Mean change 0.180 (SD 0.136) AI vs 0.346 (SD 0.150) control; difference 0.166 mm smaller in AI arm; Cohen d 1.16
0.115 to 0.217 mm (95% CI, mean difference)
Welch t(119.45) = 6.43, P < .001 (unadjusted)
F3 · Percentage of fixation duration in the operative area (AOI1) during day-7 learning
GenAI-assisted: ChatGPT guidance during VR skill training, no human tutoring (research assistants auditing outputs) vs Teacher-led control: on-site professional instructor guidance during the same VR skill trainingWith AI at assessment — assisted performance, not learning evidenceDuring interventionTransfer not assessedProcess / behaviorPrespecification unclear
Unclear
Mann-Whitney U = 862, r = 0.47 (AI higher)
9.22 to 16.74 (95% CI, median difference)
P < .001
F4 · Percentage of time spent in the operative area (AOI1) during day-7 learning
GenAI-assisted: ChatGPT guidance during VR skill training, no human tutoring (research assistants auditing outputs) vs Teacher-led control: on-site professional instructor guidance during the same VR skill trainingWith AI at assessment — assisted performance, not learning evidenceDuring interventionTransfer not assessedProcess / behaviorPrespecification unclear
Unclear
Mann-Whitney U = 1047, r = 0.38 (AI higher)
4.16 to 10.83 (95% CI, median difference)
P < .001
F5 · Fixation rate in the operative area (AOI1) during day-7 learning
GenAI-assisted: ChatGPT guidance during VR skill training, no human tutoring (research assistants auditing outputs) vs Teacher-led control: on-site professional instructor guidance during the same VR skill trainingWith AI at assessment — assisted performance, not learning evidenceDuring interventionTransfer not assessedProcess / behaviorPrespecification unclear
Unclear
Mann-Whitney U = 1268, r = 0.28 (AI higher)
0.17 to 0.48 (95% CI, median difference)
reported P < .001; U 1268 gives z about 3.15, two-sided P about .002 (reviewer recomputation)
F6 · fNIRS HbO GLM β (task activation) in PFC, MC, VAC and TPJ, both hemispheres, during a day-7 block-design operative task
GenAI-assisted: ChatGPT guidance during VR skill training, no human tutoring (research assistants auditing outputs) vs Teacher-led control: on-site professional instructor guidance during the same VR skill trainingWith AI at assessment — assisted performance, not learning evidenceDuring interventionTransfer not assessedPhysiologicalPrespecification unclear
Unclear
not_reported in text (β descriptives in Table S3, not in bundle); AI higher in all regions bilaterally
not_reported
cluster-based permutation t tests; Figure 8 marks P < .01
F7 · Card Rotation Test (low-level visual-spatial ability), 8 tasks in 3 minutes, standardized 0-100; post only
GenAI-assisted: ChatGPT guidance during VR skill training, no human tutoring (research assistants auditing outputs) vs Teacher-led control: on-site professional instructor guidance during the same VR skill trainingAI availability at assessment not reported — cannot carry a learning claimImmediate post-testFar transferStandardized testPrespecification unclear
Benefit
77.62 (SD 18.81) AI vs 62 (SD 22.42) control; difference +15.62; Cohen d 0.76
8.20 to 23.04 (95% CI, mean difference)
Welch t(116.79) = 4.21; Holm-adjusted P = .003
no pretest, so a post-only difference in a trait-like ability after about 7 hours of practice is hard to attribute to the intervention; the Discussion mislabels these as self-rated IN-029 ruling 6 (pass A's note kept): the group difference is reported as significant in favour of arm A, but no baseline measure exists and an ability test of this kind is not a plausible 7-day training target; read as proxy.
F8 · Cube Comparison Test (intermediate visual-spatial ability), 21 tasks in 3 minutes, standardized 0-100; post only
GenAI-assisted: ChatGPT guidance during VR skill training, no human tutoring (research assistants auditing outputs) vs Teacher-led control: on-site professional instructor guidance during the same VR skill trainingAI availability at assessment not reported — cannot carry a learning claimImmediate post-testFar transferStandardized testPrespecification unclear
Benefit
66.96 (SD 23.76) AI vs 53.90 (SD 18.25) control; difference +13.06; Cohen d 0.62
5.52 to 20.60 (95% CI, mean difference)
Welch t(114.30) = 3.44; Holm-adjusted P = .004
IN-029 ruling 6 (pass A's note kept): the group difference is reported as significant in favour of arm A, but no baseline measure exists and an ability test of this kind is not a plausible 7-day training target; read as proxy.
F9 · Mental Rotation Test (high-level visual-spatial ability), 12 tasks in 3 minutes, standardized 0-100; post only
GenAI-assisted: ChatGPT guidance during VR skill training, no human tutoring (research assistants auditing outputs) vs Teacher-led control: on-site professional instructor guidance during the same VR skill trainingAI availability at assessment not reported — cannot carry a learning claimImmediate post-testFar transferStandardized testPrespecification unclear
Benefit
59.27 (SD 24.22) AI vs 46.42 (SD 19.30) control; difference +12.85; Cohen d 0.59
4.90 to 20.80 (95% CI, mean difference)
Welch t(116.02) = 3.27; Holm-adjusted P = .004
IN-029 ruling 6 (pass A's note kept): the group difference is reported as significant in favour of arm A, but no baseline measure exists and an ability test of this kind is not a plausible 7-day training target; read as proxy.
Limitations
Stated by the authors: only partial blinding; full double-blinding not achievable because of the distinct interaction modes, possible subjective scoring bias; single-item scales for learning emotions; only 1 week of training and an immediate test; no long-term retention or real clinical performance; training restricted to basic tooth preparation; not generalizable to complex procedures; multiple comparisons; residual inflation despite Holm-Bonferroni; detailed ChatGPT usage was not recorded; possible decision offloading and overreliance on AI feedback; research assistants monitored AI outputs, but systematic verification of every response was not feasible; fNIRS and other secondary findings exploratory; activation differences may reflect effort, novelty, difficulty or other factors, not better learning. Observed by review: randomization sequence generation and allocation concealment not described; no trial dates; AI arm was not human-free: unmasked on-site research assistants audited ChatGPT outputs throughout training; whether they edited, blocked or supplemented answers is not reported; their number and qualifications are not reported; control-arm instructors not described (number, who they were, relation to the authors, ratio to students); amount of instructor guidance not recorded, as ChatGPT usage was not; operative-test conditions not reported: no statement whether ChatGPT, instructors or reference material were available during the test, which medium (VR simulator or physical typodont) was used, or how raters were masked to allocation for a practical task; interrater reliability given as a correlation (r=0.89) rather than an agreement index; no baseline measure of operative skill, visual-spatial ability, emotions or engagement; secondary 'improvement' language (e.g. greater SRL improvements, better visual-spatial ability) rests on post-only differences except for SRL; a d of 0.59-0.76 on standard visual-spatial tests (card rotation, cube comparison, mental rotation) after about 7 hours of tooth-preparation practice, with no baseline, is implausibly large for a trait-like ability and is not discussed; Discussion and Conclusions misdescribe these timed tests as self-rated; SRL Table 3: posttest means/SDs (e.g. average-achieving 30.36 (1.89) vs 25.73 (2.03); low-achieving 26.52 (1.99) vs 20.39 (1.92)) imply near-complete separation, but U 129 and 113 with r 0.36 and 0.42 correspond to modest overlap; the 95% CIs for the median difference (e.g. 2.14 to 10.12) are far wider than SDs near 2 allow; subgroup ns not given in text (reviewer recomputation); engagement Table 5: means/SDs imply d about 1.3 (behavioral), 1.9 (emotional), 2.0 (cognitive), which would give U near 3,100-3,500 of a possible 3,782, but reported U are 2,344-2,457 (r 0.21-0.26); U, r and P agree with each other but not with the descriptives (reviewer recomputation); Table 1: control-arm Dunn r values (0.18, 0.23, 0.41) are inconsistent with the reported z (2.50, 3.10, 5.60) and n 61 (expected about 0.32, 0.40, 0.72); experimental-arm r values are consistent (reviewer recomputation); Table 4 anger CI printed as –1.29 to 0.63 though the difference is about –0.96 (should be –1.29 to –0.63); frustration Welch df 120.73 recomputes to about 115.5; FR U=1268 gives P about .002, reported <.001; visual-spatial Holm-adjusted P (.003, .004, .004) are larger than the reported t and df imply (about <.001, .001, .001), a conservative discrepancy; pupil-diameter change during learning is the cognitive-load measure, but the AI arm read ChatGPT on a separate computer screen and the control arm talked to instructors, so luminance and gaze-target differences between arms are uncontrolled; pupil, gaze and fNIRS measures were all taken while the assigned guidance was available and are process, not learning, outcomes; AOI cluster analysis: clusters derived from gaze metrics and then related to scores within arms is correlational; the abstract's “more efficient visual attention allocation” is an interpretation; model given as ChatGPT-4.0 “with simple contextual training” (undefined); no version string or dates of use; prompt templates are in an appendix not in the bundle; ages 17-23 for 4th-5th-year students, and 4th-5th-year students with no preclinical skills training, are unusual and unexplained; same author group published the earlier ChatGPT-as-instructor dental RCT cited as supporting evidence (allegiance); data available only on request. Integrity note (IN-029, carried into limitations per the owner's ruling): the primary result (F1) is internally consistent; both passes recomputed its t, df, CI and d from the reported means, SDs and ns. The secondary tables are not: in the SRL and engagement tables the means and SDs imply standardized differences of roughly 1.3 to 3.1 while the reported Mann-Whitney U and r values imply far smaller separation, so one of the two sets is wrong and the source cannot say which. Pass A found a Table 3 CI that does not centre on its mean difference; pass B found Table 1 control-arm Dunn r values that do not follow from their z values and a sign error in the anger CI. Neither pass found anything showing fabrication. The secondaries are carried as reported and should not inform a claim. The same author group earlier reported a positive RCT of ChatGPT as a dental skills instructor (the paper's reference 9); data are available on request only.

Who was studied

Level
university
Years
4th- or 5th-year students of a 5-year dental program, with no prior preclinical skills training or VR simulator experience
Ages
17-23 years (mean 21.86, SD 0.97)
Country
CN
Setting
School and Hospital of Stomatology, Wuhan University (described as a top-tier Chinese university); single site; training in a digital classroom under controlled lighting. 61 male / 62 female across arms (30/32 AI, 31/30 control).
Prior knowledge
Theoretical knowledge measured by a 10-question open-ended pretest (3 expert raters, r=0.87) and used for stratification; by-arm pretest scores are in Table S1 (not in bundle). No baseline operative skill measure (students had no preclinical skills training), and no baseline visual-spatial, emotion or engagement measure; baseline SRL was measured and reported as not differing by arm within strata.

Methodological notes

Extracted from: https://doi.org/10.2196/90938 Access: full text (JMIR, CC BY 4.0), read from data/fulltext-cache/bundles/wen-2026.txt (PMC XML full text only). Figures (including the Figure 3 CONSORT flow diagram) are images and are not in the bundle; Multimedia Appendices 1-4 (scoring rubric, prompt templates and ChatGPT samples, Tables S1-S5 with demographics, subgroup ns and fNIRS β descriptives) are not in the bundle, so their content is not coded. The ChiCTR2500110063 registry entry was not reachable; registration is coded only from what the paper states. Allocation: Stratified (three pretest tertiles) individual allocation. The method of generating the sequence (computer, envelopes, lottery) and any concealment are not reported. Calibration exclusions (6 of 129) appear to have preceded randomization; no post-randomization losses are reported, and 123 were analysed (62/61). Analysis was student-level, matching the allocation unit. IN-029 rulings: arm A teacher_role supervised because unmasked research assistants were on site throughout and audited every ChatGPT output; per pass A they did not teach, and the paper does not say whether they ever corrected or withheld an output (ruling 2). Arm A model_version is the paper's own string, recorded as printed; it is not an OpenAI model name (ruling 3). prespecified is unclear throughout: the paper names the primary outcome and cites ChiCTR2500110063, but the registration, its date and its outcome wording were not readable from this machine, and the secondaries are labelled exploratory (ruling 4). F1 ai_available_at_assessment not_reported: the paper never says whether ChatGPT or instructors were available during the operative test, so F1 fails the Rule 2 gate.

02Funding

Funding and conflicts

Funding
Academic and public (Wuhan University education and research projects; Natural Science Foundation of Hubei Province 2021CFB466); no industry funding reported
Vendor funded
No
Vendor
None identified
Notes
Authors declare no competing interests and no ties to the VR, sensor or AI suppliers. Not a financial COI, but an allegiance note: the same group (Huang S, Wen C, ... Yang D) published an earlier RCT on ChatGPT as an instructor for dental skills (reference 9, JMIR 2025), and the authors are evaluating an approach they have previously promoted.

Funding is shown on every study and never used to score it.

03Claims

Claims this study bears on

  • Unrestricted use of a general-purpose LLM chatbot during practice improves university students' subsequent performance on assessments taken without AI.No direct evidence · whole-study judgment

    Right population (university dental students, direct) and a randomized ChatGPT-versus-instructor result a reader would expect to find here (F1, operative test +9.86 of 100, 95% CI 4.81 to 14.91, d 0.70), but it cannot test the claim, and the reason comes before the gate. The claim names unrestricted use of a general-purpose chatbot; arm A is prompted_pedagogy: students were given “standardized prompt templates and a unified operation guide”, the model had undefined “simple contextual training”, and “Research assistants audited AI outputs throughout training to ensure all content complied with teaching standards.” Templated prompting with human-filtered output is not unrestricted use, so the spec 11.3 lint refuses every testing stance, consistent_but_does_not_test included - the treatment given records 3, 6 and 19 on this claim. Had the class matched, two further barriers would each stop a testing stance: F1's AI availability at the test is not_reported (Rule 2), and the contrast is ChatGPT-plus-on-site-assistants against undescribed instructors with guidance received unrecorded in both arms. LOW (IN-029); the secondary tables are internally inconsistent. Moves nothing.

04Source

The source, as retrieved

Abstract

Background: The rapid advancement of AI, particularly generative AI (GenAI), is reshaping educational paradigms. However, its utility in preclinical digital medical skills education remains insufficiently explored. Objective: This study aims to investigate whether GenAI‑assisted training yields superior performance relative to traditional instruction in virtual reality (VR)–supported preclinical dental skill education. Methods: A total of 123 eligible students from a top‑tier Chinese university completed the trial. Participants were stratified by baseline theoretical scores and randomly assigned within each stratum into two arms: the GenAI-assisted group (n=62) received ChatGPT guidance without human tutoring, with research assistants monitoring AI outputs solely for safety and error checking; the teacher-led control group (n=61) received on-site instructor guidance. Both groups underwent standardized VR-based dental skill training for 7 days. Assessments included dental operative skill examinations, functional near-infrared spectroscopy recording, continuous eye-tracking monitoring, visual-spatial ability tests, and questionnaires measuring self-regulated learning, learning emotions, and engagement. Results: As the primary outcome, there was a statistically significant intergroup difference in operative test scores between the control group (mean 67.39, SD 16) and the GenAI‑assisted group (mean 77.25, SD 11.86), with the GenAI‑assisted group exhibiting higher operative test scores relative to the control group (t110.56=3.88, Cohen d=0.70; P<.001). For secondary outcomes, the GenAI‑assisted group also demonstrated significantly lower cognitive‑load indicators (mean 0.180, SD 0.136) than the control group (mean 0.346, SD 0.150; P<.001), accompanied by more efficient visual attention allocation patterns and stronger activation in prefrontal cortex, motor cortex, visual association cortex, and temporoparietal junction brain regions. The GenAI‑assisted group showed better performance across low (P=.003), medium (P=.004), and high (P=.004) level visual-spatial ability tests. Among low-achieving participants, the largest between-group difference in self-regulated learning was found in favor of the GenAI-assisted group (P=.02). Better scores were also observed for the GenAI‑assisted group in learning enjoyment, as well as behavioral, emotional, and cognitive engagement. Conclusions: Students randomized to GenAI-assisted training had higher immediate operative scores than students receiving teacher-led instruction. Secondary cognitive, behavioral, and neurophysiological findings were exploratory and pointed toward a link between GenAI exposure and variations in skill‑acquisition metrics among dental students receiving VR-based training. The results offer preliminary insights regarding the potential integration of GenAI into preclinical dental skill education. Trial Registration: Chinese Clinical Trial Registry ChiCTR2500110063; https://tinyurl.com/2hkv5wt9

Where this record came from

SourceRetrievedIdentifier
biweekly-sweepNov 11, 2026doi:10.2196/90938
Screened in 2026-11-04 (sweep item 3); cleared the SOURCE-POLICY §6 threshold (peer-reviewed, open full text); metadata verified 2026-11-11 against OpenAlex, Crossref and the PMC JATS front matter.