The exact text we saved

This is the version exactly as it was saved, character for character. The SHA-256 hash below is a 64-character code worked out from that text. Run the same text through SHA-256 and you get the same code; change a single letter and the code comes out completely different. Each version also stores the previous version’s hash, which chains them together so no version can be swapped out quietly.

TRUVACE RECORD VERSION
record: TRV-2026-2769
version: 1
kind: certified
reason: Certified into the record
timestamp: 2026-11-12T17:21:18.388126Z
status: published
lens: p_space
sector: health
headline: Large language models in undergraduate medical education: a systematic review of examination performance, knowledge acquisition, and learner perceptions
dek: BACKGROUND: Large language models (LLMs), particularly successive iterations of ChatGPT (GPT-3.5 and GPT-4), have demonstrated increasingly sophisticated performance on standardised medical assessments and are now widely accessible to undergraduate medical students worldwide. Although the evidence base has expanded rapidly since ChatGPT's public release in November 2022, findings remain fragmented across three largely distinct research domains: examination performance, knowledge acquisition, and learner percepti…
gain_title: (none)
problem_title: In a systematic review of 43 studies covering about 12,400 participants, 82% of medical students reported receiving no institutional guidance on using AI chatbots, and 61% raised academic integrity concerns.
trace_subject: (none)
gain_reading: (none)
gain_evidence: (none)
problem_reading: In a systematic review of 43 studies covering about 12,400 participants, 82% of medical students reported receiving no institutional guidance on using AI chatbots, and 61% raised academic integrity concerns.
problem_evidence: 82% reported receiving no institutional guidance | 61% raised academic integrity concerns
quick_read: A systematic review of 43 studies, covering roughly 12,400 participants in 19 countries, examined how ChatGPT-type language models perform on medical exams, affect learning and are viewed by undergraduate medical students. Among students, 82% reported getting no institutional guidance on using them.

GPT-4 passed USMLE Steps 1-3 in every study that tested it, with mean accuracy of 86% against 56% for GPT-3.5. Learning benefits were smaller: six of twelve studies found modest short-term gains, with no durable retention.

The authors call the lack of guidance the most urgent finding. Most of the underlying evidence is weak, with moderate-to-high risk of bias in 86% of the studies, so the learning results are tentative.
limitation: Risk of bias was moderate-to-high in 86% of the included studies, and the learning gains were short-term only, without durable retention, so the benefit findings are tentative.
tag: Evidence-backed problem
key_points: The review screened 1,847 records and included 43 studies with about 12,400 participants across 19 countries, searching January 2022 to March 2026. | GPT-4 passed USMLE Steps 1-3 in all studies that tested it (mean accuracy 86%, interquartile range 81% to 89%), well above GPT-3.5's mean of 56%. | GPT-4 accuracy fell to 61%-71% on multi-step clinical vignettes and 48%-63% on image-dependent questions. | Six of twelve studies reported short-term knowledge gains, with effect sizes of 0.31 to 0.71 (median about 0.53) and no durable retention. | Of students surveyed, 74% viewed LLMs positively as study aids, yet 61% raised academic integrity concerns and 59% cited hallucination risk. | Faculty were less positive than students, at 51% positive.
rundown: The review followed PRISMA 2020 and searched PubMed, Scopus and Web of Science, plus grey literature and citation tracking. Two reviewers independently screened records, extracted data and assessed bias with the Newcastle-Ottawa Scale and ROBINS-I. Results were synthesised narratively rather than pooled in a meta-analysis.

On exams, GPT-4 consistently reached passing thresholds, but the authors report clinically important weaknesses in multi-step reasoning, uncertainty calibration and tasks involving images. Accuracy on multi-step clinical vignettes was 61%-71% and on image-dependent items 48%-63%.

On learning and perceptions, students were broadly positive (74%), but 59% cited hallucination risk and 82% reported receiving no institutional guidance. The authors say longitudinal controlled research with validated outcome measures and standardised AI literacy curricula are the priorities.
sources:
- peer_reviewed | BMC Medical Education | https://doi.org/10.1186/s12909-026-09999-7 | 2026-07-30
prev: 0000000000000000000000000000000000000000000000000000000000000000

SHA-256 hash

a68f1017925da3447ddbe8d5d6c5a3c9af8f70caf1100a3a5854cf9ca565db0e

Previous version’s hash

0000000000000000000000000000000000000000000000000000000000000000