On November 8, 2026, Texas A&M University researcher Taner Cokyasar released an arXiv preprint, What Changed in Engineering Writing After Generative AI Chat Assistants: Evidence from 1.8 Million Articles, tracking textual shifts across 256 engineering journals from 2015 to 2025. The study documents an 82% surge in vocabulary characteristic of language models, presenting crucial empirical benchmarks for journal editors and research integrity officers evaluating AI disclosure policies and manuscript screening.
What happened
To measure how academic prose shifted following the November 30, 2022 public release of ChatGPT, the study assembled a corpus of 1,839,962 articles published between January 2015 and December 2025. The sample drew from 256 high-impact engineering journals across 14 Google Scholar Metrics subcategories matched to OpenAlex bibliographic records. To isolate genuine behavioral changes from indexing anomalies, the author established monthly time series across 229 journals with sustained publication histories, projecting pre-2023 baselines while controlling for publication distortions during the acute COVID-19 pandemic period from March 2020 to June 2021.
A key methodological discovery involved publication date hygiene: roughly 8.5% of articles in the core journals carried January 1 publication dates because only the publication year had been indexed. Unless removed from monthly modeling, these default timestamps create artificial January spikes that mimic seasonal writing cycles.
In article abstracts, a tracking list of 31 marker words associated with language model outputs appeared 81.8% more frequently between March 2023 and December 2025 than their projected historical trend. In 2025, marker words reached 55.4 occurrences per 10,000 words against a projected 27.0. By contrast, four formal control words of similar register—holistic, vital, state-of-the-art, and moreover—fell 15.6% below trend, confirming that the departure was not a broad drift toward formal tone. Individual marker words displayed massive surges: “underscore” appeared 24.0 times more often than projected (107.5 per million words versus 4.5 projected), with “intricate” (5.83 times) and “meticulous” (4.05 times) showing steep increases. Meanwhile, “delve” remained uncommon in engineering, rising from 0.7 to 8.1 per million words.
Linguistic hedging also changed. Hedging terms such as “may” fell 13% in abstracts, while emphatic assertions such as “clearly” rose 20%, and words with nine or more letters increased 11%. A random, weighted sample of 12,866 full texts confirmed the pattern: marker words surged 98.6% in article bodies. However, bibliographic indicators—author counts, reference list lengths, page counts, and publication volume—remained within 4% of historical projections.
Methodology and key findings
Unlike biomedicine, where PubMed provides a centralized index of abstracts, engineering research faces fragmented open metadata. In the 1.8-million-article corpus, three publishers—Elsevier, Springer, and the Royal Society of Chemistry—represent 46% of all papers but deposit open abstracts at low rates (35.9% for Elsevier and 21.6% for Springer). Supplementing OpenAlex with Semantic Scholar allowed 68.0% of articles to be evaluated for abstract metrics.
The study also revealed a pronounced gap between stylistic evidence and formal transparency. Explicit statements disclosing generative AI assistance rose from zero before 2023 to 1.1% in 2024 and 3.7% in 2025. Yet 79 of the 98 disclosing papers in the full-text sample came from a single publisher: Elsevier, where 7.0% of 2025 articles contained a declaration, compared to 2.2% at MDPI, 1.4% at IEEE, and 0% at the American Chemical Society. Excluding Elsevier reduced the 2025 disclosure rate to 1.3%. Rather than reflecting author openness, current disclosure metrics primarily reflect institutional compliance with Elsevier’s generative AI authorship policy, contrasting with the varied guidelines seen in the IEEE generative AI policy.
Why it matters for research practice
For research integrity officers and editors, the preprint delivers a sharp warning against naive manuscript screening. Automated keyword filters flagging marker words risk high false-positive rates, disproportionately penalizing non-native English scholars who adopt formal vocabulary and unfairly flagging authors whose natural voice leans formal. Such blunt screening differs fundamentally from identifying systematic fabrication like tortured phrases and paper mill detection. Aggregate lexical surveillance offers valuable macro-level monitoring across journals, but applying it to reject individual manuscripts is statistically unsound and easily gamed.
Furthermore, because voluntary disclosure captures only a fraction of assisted writing, research institutions and publishers need standardized reporting mechanisms. Adopting actionable AI disclosure statement examples and structured taxonomy contributions helps authors transparently acknowledge language polishing without fear of punitive editorial bias.
Finally, the paper stresses that the design is descriptive rather than causal. Lexical preferences are already evolving: several early marker words such as “additionally” and “crucial” declined in late 2025 while terms like “robust” and “paradigm” accelerated. As authors adapt their prompting habits and frontier models evolve, research offices must rely on nuanced editorial evaluation rather than rigid lexical gatekeeping.
