Why do medical journals record AI use so differently?
A study of 97,916 medical articles found that AI transparency statements rose sharply, but publisher workflow explained much of the pattern. Missing declarations cannot be treated as proof that authors used no AI.
The short version
3 key points
- 1.Researchers analysed full text from 97,916 English-language medical articles published from 2015 to April 2026 across 14 clinical fields.
- 2.Any AI transparency statement reached 22.4% in January–April 2026, but fell to 9.9% after Frontiers journals were excluded; one publisher supplied 71.1% of all negative declarations.
- 3.The classifier performed strongly in blinded human review, yet the study can measure only explicit statements—not undisclosed AI use, author intent or whether a declaration was accurate.
The answer: forms and workflows shape disclosure as much as author choice
Medical journals are recording AI use very differently because they do not ask authors in the same way. In a retrospective analysis of 97,916 articles, any explicit AI transparency statement rose from 1.4% in 2023 to 22.4% in January–April 2026. But the 2026 rate was 9.9% after journals from Frontiers were removed. Frontiers supplied only 6.1% of analysed articles yet accounted for 71.1% of statements saying AI had not been used. The result suggests that a required declaration field can make disclosure routine, while free-text policies leave many articles silent.
Silence is not the same as non-use. The study searched for positive disclosures, explicit declarations of non-use, named tools, use cases and statement locations. An article without such text may have involved no AI, may have used AI outside the policy's scope, or may simply lack a captured declaration. The research therefore measures publication governance—the visible record created by journal systems—not the true prevalence of AI assistance in medical research and writing.[1]
What was analysed
The authors assembled 97,916 English-language medical and surgical journal articles published from 2015 through April 2026 across 14 fields. They retrieved full text from PubMed Central XML or publisher-hosted article files and used a large language model to classify positive AI-use disclosures, explicit non-use statements, named tools, disclosed tasks and the location of each statement. The retrospective cross-sectional design describes the published record at different times; it does not follow individual authors or journals prospectively.
The sample is large, but it is not the whole medical literature. It depends on full text that could be retrieved and classified, excludes non-English publications, and reflects the selected fields and journals. Publisher policies, article templates and indexing routes may differ systematically between included and missing material. Rates should therefore be read as estimates for this corpus, not universal percentages for every clinical paper worldwide.[1]
How the automated classification was checked
The authors used three validation routes. First, pre-ChatGPT articles provided a specificity check against implausible modern generative-AI declarations. Second, a different model reclassified 10,979 articles. Agreement between the two models was high: Cohen's kappa was 0.96 for positive disclosures and 0.98 for negative disclosures. Third, two independent reviewers conducted blinded assessment of 300 stratified cases, with adjudication of disagreements.
Human review confirmed 93.3% of positive classifications and 97.3% of negative classifications. That supports scalable surveillance, but it does not make every label correct. A stratified validation set is designed to test different classes rather than reproduce their natural prevalence, and confirmation of statement type is distinct from verifying whether the authors' declaration was truthful or complete. Automated monitoring can find what papers say; it cannot inspect every unrecorded use of a writing, coding or analysis tool.[1]
Positive and negative declarations tell different stories
Positive disclosure—authors saying they used AI—reached 7.3% in January–April 2026, or 6.2% without Frontiers journals. Negative disclosure reached 15.1% overall but only 3.7% after Frontiers was excluded. Health Informatics and Digital Medicine had the highest positive rate during 2023–2026 at 11.1%, or 10.5% without Frontiers. Among 91 journals with at least 20 classified articles in 2025–2026, statement rates ranged from zero to 100%; all three journals at 100% were Frontiers titles.
Language refinement was the most common disclosed use, appearing in 72.8% of positive statements. That category can cover very different practices, from grammar polishing to extensive rewriting, and a declaration alone does not reveal how much text changed. Named tools and task descriptions improve interpretability, but readers also need to know whether AI touched data extraction, statistical analysis, images, references or substantive interpretation—uses with different risks for error and accountability.[1]
Why this matters to clinicians and readers
Clinicians rely on journals to make methods, authorship and conflicts visible. An AI declaration does not determine whether a paper is trustworthy, but it can tell reviewers where additional checks are warranted. Language assistance may require verification that meaning and citations were preserved. Automated data extraction needs validation against source records. Generated code or statistical analysis needs reproducible inputs, tests and human responsibility. Image generation or alteration requires especially explicit provenance because it can change what appears to be evidence.
For patients and the public, consistent disclosure helps separate the evidence from the production process. It should not become a warning label that discredits research merely because a tool was used, nor a badge implying that declared use was safe. The useful question is whether the authors explain the task, tool, version, human checks and accountability. A structured field can make that information searchable without forcing readers to hunt through acknowledgements, methods, supplements or publisher notes.[1]
The study also discloses its own AI use
The paper reports using Claude Opus 4.6 and 4.8 for code development, data-pipeline construction, manuscript drafting and language refinement. Gemma 4 31B and Nemotron-3-Super-120B were used for automated disclosure classification, while Qwen3-235B-A22B-Instruct-2507 extracted verbatim statements to support human validation. The authors state that they accept responsibility for the resulting content. That level of task-specific detail makes the work itself an example of the structured transparency it advocates.
No external funding was reported. One author disclosed research support from the National Foundation of Emergency Medicine and unrelated consulting on a GE Healthcare medical advisory board; the remaining authors reported no additional competing interests. The article is an accepted, citable early version that may receive copy-editing before the final version of record, so readers should expect presentation changes without treating the peer-reviewed findings as a press release.[1]
What would change the assessment
The clearest next test is prospective policy evaluation. Journals could introduce a mandatory structured declaration with fields for tool, version, task and human verification, then measure completeness and error rates before and after implementation. Random or risk-based audits of manuscripts, code and revision histories could test whether recorded declarations match observable workflows. A comparison with otherwise similar journals that do not change their systems would help separate policy effects from the general rise in AI use over time.
Coverage should expand to non-English journals, preprints, conference proceedings and disciplines outside medicine. Researchers should also distinguish authoring assistance from AI used as a scientific method and record whether policies capture peer reviewers and editors, not only authors. Until those studies exist, this paper supports a practical governance conclusion: publication systems can make declarations far more visible, but a count of disclosure statements is not a count of actual AI use.[1]
What this means for people
- Clinicians and patients need to know when AI affected analysis, evidence presentation or interpretation—not merely whether a generic declaration exists.
- Authors face less ambiguity when journals collect the tool, task and human checks through a consistent structured field.
- Editors can use automated monitoring to find missing or inconsistent statements, but human review remains necessary for consequential uses and disputed cases.
Global context
The corpus covers English-language medical literature and is influenced by publisher-specific workflows. Journals in other languages, smaller regional publishers and disciplines outside medicine may have different policies and technical capacity, so the measured rates should not be assumed globally representative.
What the evidence does not yet show
- The analysis detects explicit statements; it cannot measure undisclosed AI use or verify that a declaration is complete and accurate.
- The 97,916-article corpus covers retrievable English-language medical literature rather than every journal, language or discipline.
- Publisher workflow is part of the observed signal, so article-level rates do not directly measure author willingness to disclose.
- Model and human validation support classification accuracy but do not eliminate errors or establish the prevalence of hidden use.
- The cross-sectional design describes change over time without proving which policy caused it.
What to watch next
- Mandatory structured fields recording tool, version, task and human verification
- Prospective before-and-after studies when journals change disclosure workflows
- Audits comparing declarations with manuscript histories, code and research records
- Comparable monitoring in non-English publishing and outside medicine
Living evidence record
Impact record IAI-0J861HO
Evidence stage
Studied
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Present
Record status
Monitoring
Last checked
13 October 2026
Source trail
1 direct source across 1 source type.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Single-source reporting disclosure
This record analyses one direct source. It can establish what Research Integrity and Peer Review published or reported, but it is not independent corroboration of every performance claim or predicted outcome. The confidence label will change only when broader evidence is added.
Evidence trail
Sources used for this report
Links checked 13 October 2026
Research question, topics and regions
How often medical articles contain explicit AI-use or non-use statements, and how publisher submission systems shape what becomes visible
Topics: Research integrity · Medical publishing · AI disclosure · Peer review · Publication governance · Large language models
Regions: North America · Global
This report is labelled source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
The Impact Brief · Free
Keep following the stories that matter to you.
Get our weekday briefing on what changed and why it matters. Choose your topics below.
Continue the story
There’s more to this story
Science & Research
Can an LLM spot what a clinical-trial paper leaves out?
A peer-reviewed benchmark found that a retrieval-plus-LLM pipeline reached F1 0.822 across 119 reporting questions on 40 held-out trial articles. Better evidence retrieval lifted F1 to 0.893, showing both the promise and the main bottleneck.
9 min · 2 sources
Science & Research
Can AI design better optimisation rules?
A Nature Machine Intelligence paper reports that a structured LLM framework outperformed prior LLM approaches across 36 combinatorial-optimisation benchmarks and produced feasible algorithms for four new port-logistics problems. It remains an offline benchmark, not a live operations result.
9 min · 2 sources
Science & Research
Can a general-purpose LLM search crystal compositions?
In a closed computational benchmark, GPT-5.4 recovered 95.65% of 3,740 low-energy Elpasolite targets within 5,000 proposals. Iterative evaluator feedback drove the result; no new material was synthesised, and the approximate energy labels are not proof of stability or usefulness.
8 min · 2 sources
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.