Researchers who resist using artificial intelligence in their scientific writing exhibit citation habits that predate generative AI, citing older literature and repeatedly returning to familiar sources, according to an extensive study of 71,133 first-authors released on November 12, 2026.
Tracking Resistance Across 71,133 Authors in OpenAlex
The preprint, titled ‘Earlier citation practices are linked to signs of artificial intelligence use in scientific writing’ (arXiv:2611.16833) by Hongkan Chen, Yongyuan He, Qingshan Zhou, and Yi Bu of Peking University, provides one of the largest longitudinal examinations to date of how individual researchers adopt generative text tools. Drawing on publication records from the OpenAlex scholarly database, the investigators tracked 71,133 active first-authors across 565,602 peer-reviewed articles, reviews, and conference papers published between 2023 and late September 2026, pairing them with complete pre-2023 publication histories.
To measure AI uptake, the team estimated abstract-level writing similarity scores using maximum likelihood estimation against reference corpora generated by large language model (LLM) systems. They defined an author’s ‘resistance score’ as the predicted share minus the observed share of abstracts displaying AI-assisted lexical markers. Authors with positive scores produced fewer AI-assisted abstracts than expected given their discipline, career stage, prior productivity, and institutional context. While early resistance strongly predicted later resistance, the proportion of resistant authors steadily shrank across the four-year window, declining from 52.9% in 2023–2024 to 44.1% in 2025 and 37.2% by September 2026.
Environmental Drivers and Collaborator Influence
Decomposing the predictive factors among initially resistant researchers revealed that environmental and institutional environments exerted far greater influence over subsequent AI adoption than individual career metrics. In out-of-sample models, disciplinary affiliation, institutional country language category, and co-author network characteristics contributed two to three times more predictive gain than an author’s academic age or past publication volume.
Disciplinary cultures showed marked divergence in adoption dynamics. Over half of initially resistant authors in mathematics and physics maintained continuous resistance through 2026. Conversely, fewer than a quarter of researchers in computer science, engineering, and business remained resistant over the same period. Furthermore, authors based at institutions in English-speaking countries sustained higher resistance rates than peers in non-English-speaking regions.
The study also uncovered a measurable social contagion effect within co-authorship networks. Among resistant authors who published with measurable collaborators, partnering with even a single colleague whose prior writing showed AI markers reduced the focal author’s subsequent resistance score by an estimated 0.0595 in 2025 and 0.0515 in 2026 (p < 0.001). The effect remained pronounced when restricted specifically to corresponding-author collaborators.
Pre-Existing Citation Habits and the Widening Reuse Gap
The central contribution of the Peking University study lies in comparing reference practices between consistently resistant authors (positive resistance scores in both 2023–2024 and 2025–2026) and consistently non-resistant authors. Rather than emerging alongside generative chatbots, the bibliometric divergence was established years before the public release of generative AI tools.
Consistently resistant authors routinely drew upon older foundational literature and concentrated their citations within fewer scientific subfields. In contrast, researchers who readily integrated AI tools cited more recent work across wider disciplinary boundaries. Furthermore, resistant authors exhibited a pronounced tendency to reuse references they had cited in prior first-authored publications, a gap that expanded substantially over time.
| Bibliometric Metric | Pre-2023 Baseline | 2023–2024 Early Period | 2025–2026 Late Period |
|---|---|---|---|
| Adjusted mean reference age difference (years) | +0.75 [0.66, 0.84] | +1.47 [1.36, 1.58] | +1.59 [1.48, 1.70] |
| Adjusted median reference age difference (years) | +0.52 [0.44, 0.60] | +1.16 [1.07, 1.26] | +1.29 [1.19, 1.38] |
| Adjusted reference reuse gap (percentage points) | +3.24 pp [2.91, 3.57] | +6.69 pp [6.22, 7.16] | +7.95 pp [7.43, 8.47] |
| Subfield variety difference (variety index V) | -0.285 [-0.360, -0.211] | -0.612 [-0.711, -0.512] | -0.713 [-0.814, -0.612] |
| Share of references aged 20+ years (percentage points) | +1.99 pp [1.75, 2.22] | +3.57 pp [3.30, 3.84] | +3.75 pp [3.47, 4.04] |
Implications for Scientometrics and Integrity Screening
These findings provide essential context for scientometricians and research administrators evaluating the systemic impacts of artificial intelligence on scholarly communication. Because researchers who already cited newer, interdisciplinary literature adopted AI tools faster, shifting citation patterns in published literature reflect differential adoption rates across distinct author groups rather than generative models arbitrarily distorting citation practices.
For journal editors and Research Integrity Officers (RIOs), the results underline the hazard of relying on blunt lexical screens. Automated AI detection tools across research offices capture surface phrasing and vocabulary conformity rather than evidentiary rigor. Authors working in fast-evolving fields or writing in second languages may adopt generative assistance for language polish while preserving sound citation practices, whereas scholars with deep, established citation routines remain tethered to traditional workflows.
Sources
- Chen, H., He, Y., Zhou, Q., and Bu, Y. (2026). Earlier citation practices are linked to signs of artificial intelligence use in scientific writing. arXiv preprint arXiv:2611.16833v1. https://arxiv.org/abs/2611.16833 (2026-11-15).
- OpenAlex Documentation and API, OurResearch. https://openalex.org (2026-11-15).
