54% of 2025 Astronomy Papers Carry LLM Vocabulary Traces, Preprint Estimates

The study applies a hierarchical Bayesian mixture model to word counts normalized by paper length across the full astro-ph corpus from 2015 through mid-2026. Pre-2020 papers establish baseline unassisted rates, while 392 self-disclosing papers calibrate the assisted component; three background-rate assumptions extend the model past 2020. This yields the 54% central estimate with statistical and systematic uncertainties that remain above 36% under varied calibrations and word lists drawn from subfield-consistent rises.

The disclosure gap and temporal fading of markers point to broader reproducibility risks. Authors appear to adapt phrasing to evade detection, halving excess signal between 2023 and 2026, which complicates attribution of ideas, methods, and even data interpretations. Parallel patterns appear in biomedical preprints where LLM-generated sections have introduced factual hallucinations that later required corrections, suggesting astronomy may face similar hidden error propagation if detection methods lag.

Policy responses remain underdeveloped. Journals have issued vague AI-use guidelines without enforcement mechanisms, while grant agencies have not updated authorship standards. A coordinated requirement for structured disclosure tied to model version and prompt logs could close the gap, but adoption would depend on community norms shifting faster than the current one-in-sixty-six disclosure rate.

Future detection will require moving beyond static word lists toward syntactic and citation-pattern signals that evolve more slowly than vocabulary.