Active Wake record · Pilot
LLM-assisted writing in biomedical papers
A corpus-level estimate says language-model assistance is already widespread in biomedical publishing; revision and replication will show how much confidence that number deserves.
Current reading
Where the record stands
The public evidence remains a single preprint and reporting derived from it. The 89 percent figure is a corpus-level estimate of LLM-associated vocabulary, not a detector for individual papers, and no independent replication was found during admission review.
- Last updated
- Sep 6, 2026, 7:24 AM
- Next dated review
- Oct 1, 2026
- Event trigger
- Review immediately for a revised preprint, peer-reviewed publication, independent replication, methodological critique, or a major publishing-policy change explicitly tied to prevalence evidence.
Point zero
What was known when we noticed it
A preprint estimated that 89 percent of sampled PubMed Central papers published in December 2025 contained signs of LLM-assisted writing or editing somewhere in their full text.
Supported at the time
- The analysis included 1,194,287 English-language PubMed Central papers dated 2017 through 2025 with Introduction, Methods, Results, and Discussion sections.
- The estimator compared post-2022 frequencies of 379 LLM-associated marker words with trends projected from 2018 through 2022.
- For random 255-word crops in December 2025, the estimated usage was 68 percent in Discussion sections and 32 percent in Methods sections.
Still uncertain
- The work was a preprint and had not completed journal peer review at capture time.
- The estimate depends on the projected human-only vocabulary baseline remaining valid after 2022.
- The method estimates prevalence across a corpus and cannot reliably classify an individual paper or distinguish translation, editing, and substantive generation.
First recorded in Morning Soundings: The Number Is Not the Thing. The packet was captured as contemporaneous evidence.
Append-only history
Observation timeline
Point-zero observation
A preprint estimated that 89 percent of sampled PubMed Central papers published in December 2025 contained signs of LLM-assisted writing or editing somewhere in their full text, while emphasizing that the method cannot classify an individual paper.
Lifecycle: EmergingEvidence: PreliminaryAttention: Unmeasured2 sources attached
- arXiv: Most biomedical publications show signs of LLM-assisted writing (primary-research; lineage holzwarth-biomedical-llm-writing-preprint)
- Nature: Staggering 90% of biomedical papers now show signs of AI help (independent-reporting; lineage holzwarth-biomedical-llm-writing-preprint)
Admitted to The Wake
Admitted because a prominent quantitative estimate is still provisional and has clear paths to revision, replication, and policy consequence; the admission review found the arXiv record still at its initial public version.
Lifecycle: EmergingEvidence: PreliminaryAttention: Unmeasured
Unresolved
Open questions
- Will peer review or independent replication materially revise the estimated prevalence?
- How much of the detected assistance is translation or editing rather than generation of scientific claims?
- Can future estimators separate direct model use from human adoption or avoidance of model-associated vocabulary?
Future evidence
What would move the record
Peer-reviewed publication or revised preprint
The authors release a revised version, journal article, correction, or methodological response.
No defensible date; event-triggeredIndependent replication on scholarly corpora
An independent group tests the estimator on PubMed Central or another large scholarly corpus.
No defensible date; event-triggeredPolicy response tied to population-level evidence
A major publisher or research body cites prevalence evidence when changing LLM-use disclosure or authorship policy.
No defensible date; event-triggeredProvenance
Source trail
A lineage identifies sources that ultimately descend from the same paper, announcement, interview, wire report, or other evidentiary ancestor. Repetition within one lineage is not counted as independent corroboration.
Primary Research
Most biomedical publications show signs of LLM-assisted writing
arXiv
Lineage: holzwarth-biomedical-llm-writing-preprint · Accessed Aug 31, 2026Independent Reporting
Staggering 90% of biomedical papers now show signs of AI help
Nature
Lineage: holzwarth-biomedical-llm-writing-preprint · Accessed Aug 31, 2026