Has Scientific Talent Shifted from Depth to Breadth?Evidence across Papers, Knowledge Inputs, Careers, and Teams
Xiaoshn Nee, Haobo Zhong, Xiaomin Ni
cs.DL · cs.AI · cs.HC
Quick verdict
Finds science stays deep as teams grow and outputs focus; rigorous multilevel OpenAlex study, but equivalence bounds were set post hoc and code not yet public.
Summary
This empirical science-of-science study tests whether generative AI is shifting scientific expertise from depth to breadth, using an OpenAlex snapshot retrieved 12 September 2026. Four linked levels are analysed: 47,959 articles balanced across six fields and 2010–2025; an independent reference sample covering 51,736 resolved cited works; reconstructed publication histories for 1,754 sampled contributors; and citation returns among 959 complete contributor events with mature four-year FWCI windows ending 2022. From 2010 to 2022, modelled team size rose 37.3% [34.4%, 40.3%] while paper topic breadth declined 0.0144 on a 0–1 hierarchical distance scale; cited knowledge breadth was stable to modestly broader (annual +0.00228, HC3 p=0.048, but p=0.112 with clustering). Individual five-year breadth rose 0.0190 [-0.0078, 0.0459] by 2019–2022, judged within a ±0.05 equivalence bound; one SD of focal depth was associated with 8.2% higher 1+FWCI [1.9%, 14.9%]. Post-2022 departures from pre-trends were unsystematic and showed no clear gradient across 83 subfields' baseline AI intensity. Six primary tests are Holm-corrected, with TOST equivalence and placebo analyses. Main limitations: equivalence bounds chosen post hoc, OpenAlex author-disambiguation risk, contributor sampling capped at teams of 1–20, only three recent years, and no code deposited yet.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Team size grew an estimated 37.3% [34.4%, 40.3%] from 2010 to 2022 while paper topic breadth declined 0.0144 on a 0–1 hierarchical distance scale, so outputs concentrate as collaborations expand.
- Cited knowledge breadth stayed stable to modestly broader, revealing a divergence between focused output topics and the reach of knowledge inputs.
- Established contributors' five-year prior breadth rose only 0.0190 [−0.0078, 0.0459] by 2019–2022, within the study's ±0.05 equivalence bound, rejecting a strong depth-to-breadth shift in individual repertoires.
- In mature citation windows, one SD of focal depth was associated with 8.2% higher 1+FWCI [1.9%, 14.9%], while average breadth and breadth×depth associations were practically small under the declared bounds.
- Post-2022 deviations from earlier trends were not systematic and did not vary clearly with baseline AI intensity across 83 subfields.
Strengths
Multilevel linked design makes divergent measures interpretable
Figure 1 and §2.1 link the article sample, an independent reference sample, reconstructed contributor histories and citation returns within one topic hierarchy (eqs. 1–4). The central finding — output topics contracting (−0.0144) while cited inputs stay stable-to-wider (+0.0273 cumulative, Table 3) — is only decodable because both breadth measures aggregate the same hierarchical distance over different objects.
Inference discipline unusual for this literature
Six primary nulls were specified after data construction but before estimating capability/outcome models, with Holm adjustment (Table 2, §5.8); only the two organizational/output trends survive (both Holm p<0.001), and the null results are reported as prominently as the rejections.
Equivalence testing with declared SESOIs
Extensive triangulation and robustness
Honest scope statement
Weaknesses
Equivalence bounds chosen post hoc yet carry the headline
Cited-knowledge trend is the fragile leg of the divergence story
Career reconstruction rides on OpenAlex author disambiguation
Pooled contraction generalizes weakly across fields
Individual estimand excludes hyperauthorship exactly where teams grow fastest
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper asks whether generative AI is shifting scientific talent from deep specialization toward broad individual knowledge, answered with a multilevel descriptive analysis of OpenAlex records (snapshot 12 September 2026). Four linked levels: (1) a balanced article sample — 500 records per six fields per 16 years, 47,959 after exclusions; (2) an independent cited-knowledge sample — 35 papers per field-year, up to 20 references each, 51,736 of 55,223 identifiers resolved (93.7%), 3,215 eligible papers; (3) reconstructed prior histories for 1,754 unique contributors (1,316 eligible events) drawn uniformly from teams of 1–20 authors; (4) citation returns on 959 complete events with mature four-year FWCI windows through 2022. Six primary null hypotheses were fixed after data construction and before estimating the capability/outcome models (Table 2): annual paper-breadth and log(1+team size) trends; individual breadth and depth changes, 2019–2022 vs 2010–2014; the breadth-by-depth citation interaction; and the mean post-2022 forecast residual. Only the two organizational/output trends reject (both Holm p<0.001).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
Did this paper find that AI is turning scientists into generalists?
No. Using OpenAlex data for six fields (2010–2025), the study finds established contributors' topic breadth changed only modestly (+0.0190 by 2019–2022, within the study's ±0.05 equivalence bound), and post-2022 changes did not vary clearly with how AI-intensive a subfield was at baseline. The clearest changes were organizational: team size grew about 37.3% while paper topics became more focused.
What data does the study use?
Public OpenAlex records retrieved 12 September 2026: 47,959 randomly sampled articles across six fields and 16 years (2010–2025), an independent sample of 3,215 papers whose 51,736 cited works were resolved, and reconstructed full publication histories for 1,754 randomly selected authors, 959 of whom enter the citation-return model with complete four-year citation windows.
Is depth or breadth better for citation impact according to the paper?
In mature four-year citation windows, one standard deviation of accumulated depth in the paper's own subfield was associated with 8.2% higher 1+FWCI [1.9%, 14.9%]. Average breadth (+1.8%) and the breadth-by-depth interaction (+0.7%) were small and fell within the study's practical-equivalence bounds. These are conditional associations, not causal effects.
Is the paper's code and data available?
The paper states that analysis code and data are organized in a replication package for deposit in a public repository before journal submission. As of this preprint version, the package had not yet been deposited, so the exact numbers cannot currently be checked against released code.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 2–7 and A1–A2 are available only as extracted captions; visual point-level patterns (e.g., field trajectories in Figure A1) could not be independently verified.
- The abstract appears truncated mid-sentence in the extracted text; conclusions were cross-checked against §3–§4 and Tables 2–3.
- The replication package is referenced but not yet deposited, so reported numbers could not be validated against code or the exact OpenAlex snapshot.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.