AI evidence extraction

Smaller, Younger, and More Impactful: How AI-Assisted Writing Transforms Research Teams

Haoyang Wang, Mingze Zhang, Yi Bu, Star Zhao, Meijun Liu · ArXiv.org · 2026

AI-generated evidence extraction, verified across multiple analytical personas. Not a substitute for the peer-reviewed original.

9/10

Relevance

2/4

Quality (LMQS)

Evidence

Citations

0.00

FWCI

Methodology & findings

Study design

Large-scale observational study using multiple statistical methods including ordinary least squares regression, quantile regression, Poisson regression, logistic regression, and propensity score matching on full-text publications from PLoS family (105,994 publications) and Nature portfolio (41,080 publications) from 2020-2025.

Sample

N = 147074, 9 groups

Primary method

Multiple statistical approaches were employed: (1) Mann–Whitney U test for comparing means across non-normally distributed groups (team age and team size tested with Kolmogorov-Smirnov K-S test for normality, D=0.141 for team age, D=0.031 for team size, both p<0.01); (2) Ordinary Least Squares (OLS) regression with robust standard errors for team age; (3) Quantile regression at 25th, 50th, and 75th percentiles for team age heterogeneity analysis; (4) Poisson regression for team size (count variable); (5) Logistic regression for binary outcome (Top 5% FWCI); (6) Propensity Score Matching (PSM) using 1:1 nearest neighbor matching with 0.2 standard deviation caliper; (7) Paired t-tests for Average Treatment Effect on the Treated (ATT) significance. Multiple fixed effects included: time (year-month combination), discipline, and journal.

Main result

The study found that "research teams using AI-assisted writing tend to be younger and smaller" and importantly, "this shift toward more compact, junior-leaning teams does not come at the expense of scientific impact. On the contrary, we observed a higher probability of research teams that employed AI-assisted writing producing highly impactful publications." More specifically, "in both PLoS and Nature, AI-assisted publications were more likely than human-written publications to rank in the top 5% of FWCI."

Reports effect sizes.

Research paradigm

Positivist empiricism with observational data analysis

Author conclusions

The authors conclude: "Our findings suggest that AI has the potential to ease this burden, offering an alternative path" to the decades-long trend toward larger teams driven by knowledge burden. They further state: "The observed shift toward smaller and younger teams among AI‑assisted research groups calls for a careful evaluation of how AI may be reshaping scientific teamwork, both its strengths and potential unintended consequences." Additionally, "Funding agencies and institutions may need to reconsider policies that implicitly favor large, established individuals or teams, ensuring that smaller, AI-augmented groups are not disadvantaged."

Risk of bias

Selection bias: Study limited to two open-access publishers (PLoS and Nature), which may not represent broader publication landscape; Publication bias: Early-adopted AI publications may be systematically different from non-adopters; Confounding: Unobserved variables may drive both AI adoption and team composition changes; Measurement bias: AI detection algorithm has 5% false-positive rate by design (95th percentile threshold); Temporal bias: Post-2022 publications have short citation accumulation windows, potentially underestimating impact; Selection bias: Study limited to two open-access publishers (PLoS and Nature), not representative of all scientific publishing; Observational design: Cannot establish causality; only associations observed; Confounding variables: Despite propensity score matching and extensive controls, unobserved confounding possible; Measurement error: AI detection algorithm (Liang et al., 2025) uses 95th percentile threshold; may misclassify human-written papers as AI-assisted with 5% false positive rate; Citation window bias: Recent publications (post-2022) have short citation accumulation periods, potentially underestimating true impact; Career age measurement bias: Calculated only from publications in PLoS and Nature portfolio, not from comprehensive career records; Selection bias: Study limited to two publishers (PLoS family and Nature portfolio), not representative of all journals; Detection bias: AI detection algorithm may have false positive/negative rates; thresholding at 95th percentile caps false-positive rate at 5%; Confounding: Observational design prevents strong causal inference; unobserved confounding possible despite PSM controls; Temporal bias: Papers published after 2022 have short citation accumulation periods, FWCI may not capture true long-term impact; Measurement bias: Career age calculated only from publications in PLoS and Nature Portfolio, may miss earlier work in other venues

Limitations

The authors state: "First, it only analyzes publications from two sources: the PLoS family and the Nature portfolio
The findings therefore apply directly only to these two open‑access publishers
Whether they generalize to other journals or fields remains an open question for future research
Second, this study uses observational data, which means we cannot make strong causal claims
Although propensity score matching and extensive controls help strengthen our inferences, the possibility of unobserved confounding persists
Third, we focus on only two dimensions of team structure, team size and team age

Open questions raised

Generalizability beyond PLoS and Nature to other journals and fields
Causal inference: limited ability to make causal claims from observational data
Team composition: Need to examine intellectual aspects such as expertise diversity, cognitive roles, and division of labor
Impact measurement: Long-term scientific impact beyond citation windows for recent publications
Mechanism understanding: How AI specifically reduces the need for senior researchers and affects writing tasks
Generalizability beyond PLoS and Nature: Whether findings apply to other journals or fields remains an open question

Data: PLoS family full-text corpus: https://github.com/PLoS/allofPLoS; Nature portfolio dataset: provided by Liang et al. (2025) with derivation described; OpenAlex bibliographic database: https://openalex.org/; PLoS family full-text corpus: https://github.com/PLoS/allofPLoS (105,994 publications, January 2020 - May 2025); Nature portfolio dataset: provided by Liang et al. (2025) (41,080 publications, January 2020 - September 2024); OpenAlex bibliographic database: https://openalex.org/ (downloaded November 2025); PLoS family full-text corpus (105,994 publications, January 2020 - May 2025) - https://github.com/PLoS/allofPLoS; Nature portfolio dataset (41,080 publications, January 2020 - September 2024) - provided by Liang et al. (2025); OpenAlex bibliographic database (200+ million publications) - https://openalex.org/Code: None explicitly mentioned in the paper.Extracted from: pdfAgreement 50%

Explore related topics

Productivity Effects Research Productivity Empirical Study