Reporting Gaps and Challenges in Randomized Trials of Generative AI for Medical Use: Insights From a Systematic Review
Journal of Medical Internet Research ·
Background: Generative AI (GenAI) is being deployed rapidly in medicine, and evidence of its clinical safety and efficacy depends on transparent reporting of randomized controlled trials (RCTs). The CONSORT-AI (Consolidated Standards of Reporting Trials–AI) extension is a benchmark for AI trial reporting, but it was drafted for conventional AI, and its applicability to GenAI has not been examined. Objective: This study aimed to assess item-level adherence to the CONSORT-AI among RCTs evaluating GenAI for medical purposes and to distinguish deficiencies attributable to nonadherence from those attributable to items that do not correspond to how GenAI interventions are specified or delivered. Methods: PubMed, CENTRAL (Cochrane Central Register of Controlled Trials), and Embase were searched for peer-reviewed, English-language RCTs published from 2022 to April 3, 2026, supplemented by backward and forward citation searching, in which patients, health care professionals, or laypeople received a GenAI intervention for a medical purpose; non-GenAI, nonmedical, nonrandomized, and non-peer-reviewed reports were excluded. Two reviewers independently screened studies, extracted data, and rated CONSORT-AI items as reported, partially reported, not reported, or not applicable using signaling questions, with disagreements resolved by a third reviewer (Cohen κ=0.81 and prevalence-adjusted, bias-adjusted κ=0.84). Risk of bias was not assessed, as the target was reporting adherence. Adherence was the proportion of reports rated “reported” per item; item-level adherence was compared across 9 characteristics using Fisher exact and Freeman-Halton tests (2-sided α=.05). Results: From 16,485 retrieved records, 74 reports from 73 RCTs (24,150 participants; median 100, IQR 50.0‐247.0) were included. Large language models (LLMs) accounted for 98.6% (73/74) of interventions, with 60.8% (45/74) using GPT-family models. Item-level reporting adherence ranged from 0% (0/72) to 98.6% (73/74). High adherence occurred in general metadata and output specification items (range 59-73, 79.7%‐98.6%). Conversely, reporting was poor for input and technical execution domains, input data eligibility (3/73, 4.1%), input acquisition/selection (9/74, 12.2%), algorithm versioning (17/74, 23.0%), trial setting integration (21/74, 28.4%), human-AI interaction (25/74, 33.8%), performance error (27/74, 36.5%), and open access/code (31/74, 41.9%). Input quality handling (Item 5[iii]) was completely unreported (0/72). Only 9 of 74 (12.2%) reports explicitly cited CONSORT-AI. While citing reports showed higher compliance for interaction and data sharing, major technical gaps persisted regardless of citation. Conclusions: This review provides the first empirical audit distinguishing general nonadherence from structural mismatches where deterministic CONSORT-AI items fail to capture GenAI mechanics. Differing from traditional AI evaluations, it reveals that key reproducibility drivers—system prompts, inference settings, and access modalities—are omitted from current standards. These empirical findings highlight critical reporting gaps in medical GenAI trials and identify domains for a tailored extension to current reporting standards, although the evidence is limited by the use of nonvalidated signaling questions. Establishing GenAI-specific trial guidelines will ensure that clinical decisions, regulatory approvals, and patient safety safeguards are grounded in reproducible and trustworthy evidence. Trial Registration: PROSPERO CRD420251034690; https://www.crd.york.ac.uk/PROSPERO/view/CRD420251034690