Most biomedical publications show signs of LLM-assisted writing

This study introduces an assumption-light estimator for LLM-assisted writing that tracks excess frequencies of vocabulary associated with language-model editing. It fits linear regressions to word frequencies during 2018–2022, before OpenAI released ChatGPT, extrapolates a human-writing counterfactual through 2025, and optimizes across sets of 379 marker words rather than relying on prompts, synthetic reference text, or unreliable AI detectors. Applied to 1.194287M English biomedical papers from PubMed Central, the method estimates that 89% of full papers showed signs of LLM assistance by December 2025; this indicates detectable editing or drafting, not necessarily wholesale machine authorship. Section-level analysis using comparable 255-word crops found the highest prevalence in Discussion paragraphs, at 68%, versus 32% in Methods paragraphs, although Methods still exceeded 50% when entire sections were analyzed. A simulation with known ground truth recovered the true usage rate with absolute error below 0.02 across the tested range. The authors argue that earlier frequency-gap methods systematically underestimated prevalence, while the observed convergence toward LLM-associated vocabulary—especially among non-native English-speaking authors—raises policy issues involving disclosure, hallucinated citations, scientific homogenization, and research integrity.