Reply to Topaz and Bahl: Interpreting a post-2022 lexical shift in academic prose Authors: Kyle Siler Published: 2026 The article (1) documents a large, socially stratified post-2022 change in academic vocabulary. Topaz and Bahl (2) confirm the shift and constructively raise three points. Each is well taken, and together they qualify the prevalence and subgroup interpretations while leaving the central descriptive finding intact. On Prevalence The excess-mass statistic equals the total variation distance between each year’s focal-rate density and the 2020-2022 baseline, so it is distributional displacement, not an article count of large language model (LLM) use. By the coupling characterization of that distance, 57% is the minimum probability mass that must differ to reconcile the 2025 and baseline distributions, with no mixture model required. Read as prevalence, this bounds from below only under two assumptions: that the 2020-2022 baseline approximates the 2025 no-LLM counterfactual, and that post-2022 departures reflect LLM influence rather than other change. Under those assumptions, the bound is conservative: AI-influenced writing that resembles ordinary prose blends into the baseline and goes uncounted, as with the osmosis channel the article describes. The true affected share is therefore at least this large. The 1.5 to 2.6% distances among baseline years are smaller placebo benchmarks, and the rapid, large shift is unchanged. On the Lexical Measure The corpus was selected for sharp post-2022 growth, so figure 1’s internal contrast with other vocabulary is descriptive by construction, not an independent validation. The LLM interpretation rests on evidence beyond that selection. These 228 words are generic stylistic terms, not topical ones; they rose in concert when ChatGPT launched, and a subset receded in 2025, consistent with changing model outputs (SI Appendix, figure S4). That timing is hard to square with topical or editorial drift, and the selection rule, which compares only pre- and postperiod averages, cannot produce a launch-synced rise or the later partial reversal. The focal words also overlap with markers reported independently (3–5). These features support corpus-level construct validity but do not establish document-level sensitivity, specificity, or provenance. On Identification The Letter is correct. Since the post indicator is the sum of the 2023, 2024, and 2025 indicators, each post-by-field interaction equals a sum of field-by-year columns already present, and likewise for publishers. Thus, those coefficients are not identified and the plotted field and publisher values are the minimum-norm solution. The remaining interactions carry no year-interacted blocks; they are numerically unchanged when these redundant terms are dropped, so the region, journal-tier, rank, institution-type, and gender results stand. Re-estimating field and publisher as identified main-effect contrasts preserves the ordering: Computer Science highest (+23.0) and Mathematics lowest (−16.7), MDPI (+13.0), and Frontiers (+11.5) above Elsevier. Figure 4B’s descriptive trajectories show the same publisher ordering. On Causality Because no group is untreated after 2022, the interactions are adjusted differential changes relative to reference categories, not causal effects or determinants of LLM adoption; the article’s wording is clarified accordingly. These points refine the prevalence metric, the status of the lexical signal, and the identification of two coefficient sets. The documented phenomenon, a large, rapid, and socially stratified post-2022 shift in published academic language, stands.