Pith. sign in

REVIEW 5 major objections 6 minor 39 references

Return of the solo author: The changing division of labor in science in the age of generative AI

T0 review · 5 major / 6 minor · reviewed 2026-07-14 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read The long decline in solo-authored science papers halted and partially reversed after ChatGPT's late-2022 release.

desk verdict Solid large-scale descriptive break in the solo-authorship tail around late 2022, carefully multi-checked; the LLM-substitution reading is correlational and only partly bounded against database and multi-shock confounds. read the letter →

arxiv 2607.10780 v1 pith:PTPIBFBH submitted 2026-07-12 cs.CY cs.DLcs.SI

Akira Matsui This is my paper · ORCID
classification cs.CYcs.DLcs.SI
keywords ScienceofScientificcollaborationAuthorshipLargelanguagemodelsArtificialintelligenceDivisionlaborTeamSolo
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Science has long shifted from solo work toward larger teams that divide cognitive labor. This paper asks whether generative AI extends that trend or instead lets researchers complete some tasks alone that once required coauthors. Using hundreds of millions of works across 26 fields, it shows the decades-long fall in solo authorship stopped and partially rebounded right after ChatGPT became public. The rebound is strongest where coauthor tasks are easier to automate, appears among authors who previously published only with others, and produces solo papers that stay near those authors' prior collaborative topics while narrowing and tilting computational. Solo papers without credited human coauthors thus become a visible probe of which research labor AI can replace, pointing to a reconfiguration inside papers rather than simply bigger or smaller teams.

What carries the argument

The solo-authored left tail of the author-count distribution, treated as an observable probe of labor substitution: a solo paper is work completed without credited human coauthors, so changes in its share, who produces it, and what it contains mark the boundary of tasks generative AI can take over.

What would settle it

If independent measures of LLM adoption by field and author, or a later window after journal policies and indexing stabilize, show no corresponding solo-share break once venue composition, author disambiguation, and preprint volume are held fixed, the substitution reading would fail.

Watch

Extended reading notes

Core claim

The decades-long decline in the share of solo-authored papers halted and partially reversed around ChatGPT's public release in late 2022. The break is broad across most fields and publication filters, is not explained by new entrants or field-mix changes, and concentrates among authors who had recently or never published alone. Recovered solo papers stay close to the authors' own earlier coauthored content, contract in breadth, and shift toward computational topics, consistent with generative AI substituting for execution labor that coauthors once supplied.

Load-bearing premise

That the shared timing of the break at ChatGPT's public release, together with the field ordering by task substitutability, can be read as evidence of LLM substitution rather than concurrent confounds such as post-pandemic publishing shifts or database coverage changes.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. The paper uses the full OpenAlex corpus (1990–2025; ~300M works, 26 fields) to study the left tail of the author-count distribution rather than mean team size. It reports that the long decline in the share of solo-authored papers flattens or partially reverses around ChatGPT’s public release (Nov 2022), with a companion flattening of mean author counts in many fields. The break is heterogeneous: stronger in fields where coauthor tasks are more substitutable by writing/coding/analysis tools, and weak or absent in lab/instrument-heavy fields. Author-level, composition-standardized probabilities show the rebound among previously coauthor-only and never-solo authors, including seniors. Content analyses (SPECTER2 DiD netting field-wide drift; within-author breadth/exploration) find post-2022 solo papers tilt toward computational work, stay near authors’ prior collaborative content, and narrow in scope. The authors interpret solo authorship as an observable probe of AI substitution for parts of coauthor labor, while stating the design is correlational and anchored on a dating convention.

Significance. If the descriptive break and its author/content signatures hold, this is a substantial contribution to the science of science and to empirical work on generative AI’s effect on knowledge production. Focusing on the solo tail rather than the mean is a clear conceptual advance, and the multi-filter design (peer-reviewed, core, preprint, clean all), composition standardization, history-conditioned hazards P_t(k), balanced venue panels, event-study/donut timing checks, and embedding DiD are serious empirical strengths. The paper also carefully separates acceleration vs substitution predictions and assigns them to different parts of the author-count distribution. The result is falsifiable in principle (field ordering, never-solo hazards, content geometry) and would matter for authorship norms, credit, and training pathways even if the causal attribution remains incomplete.

major comments (5)
  1. [Abstract; Results (mechanism); Limitations; SI D, L, M] The central interpretive claim—that the 2022 left-tail break is an empirical probe of LLM substitution for coauthor labor—rests on a common temporal anchor plus field ordering and content signatures (Abstract; Results “The mechanism is consistent with LLM substitution”; Discussion). The manuscript correctly labels the design correlational and a dating convention (Results; Limitations). Even so, Abstract/title framing and the mechanism section still invite a causal reading that SI D, L, and M only partially bound. Concurrent shocks (post-pandemic publishing, other 2022–23 AI tools, journal policy shifts) remain inside the post window. Please either (i) demote the substitution language consistently to “timing-consistent with” and lead with the descriptive break, or (ii) add a sharper multi-shock falsification (e.g., placebos at other 2020–2024 AI/tool releases; field-by-field comparison of
  2. [Limitations; SI Section L; Materials and Methods (author IDs)] OpenAlex infrastructure changes are a load-bearing confound risk for authorship counts. SI L shows the peer-reviewed break attenuates from +1.72 to +0.75 pp/yr on balanced 2018–2024 venues (Spearman 0.84 across fields), which bounds venue entry/exit but explicitly cannot rule out within-venue metadata or author-disambiguation changes. MAG discontinuation (end-2021) and the July 2023 author-disambiguation revision fall inside or adjacent to the post window (SI L; Limitations). Because solo status is defined from author IDs, disambiguation revisions can mechanically create or destroy solo papers. A main-text robustness that freezes author IDs / re-runs on a pre-revision snapshot, or at least reports sensitivity of Δβ to excluding mid-2023, is needed before the within-author switching claim can be treated as database-robust.
  3. [Abstract; Results; SI Sections H and L] Field-level and author-level breaks do not coincide in Engineering, the largest share-level rebound (+2.5 pp/yr peer-reviewed), where author-level solo probability barely moves and the balanced-venue break falls to +0.65 (SI H, L; Results). The paper notes this, but main-text claims that “the break appears among authors who had written only with others” and that composition “does not explain the result” (Abstract; Results) over-generalize. Engineering should be treated as an explicit exception in the Abstract/Results, and the substitution narrative should be restricted to fields where within-author Δβ is positive (e.g., CS, Psychology, Economics in SI Fig. S8).
  4. [Results (mechanism); Discussion; SI Section K] SI Section K (external Liang et al. corpus) finds multi-author papers carry at least as much estimated LLM-modified writing as solo papers in every arXiv field. That is a useful check against a pure surface-editing account of the solo rebound, but it is currently buried and not integrated with the main mechanism claim. If writing assistance is pervasive on teams, the substitution story must emphasize non-writing execution (coding, analysis, drafting structure) or the decision margin to publish without coauthors—not text polish. Please move a concise version of this result into the main Results/Discussion and state what it does and does not identify.
  5. [Limitations; Results (preprint vs peer-reviewed); Abstract] Quality and paper-mill inflation remain open for the left-tail recovery (Limitations). The rebound is largest for preprints, where lag is shortest but low-cost solo output is also easiest. Peer-reviewed and core filters still show positive breaks with CIs excluding zero, which is important, yet venue-based filters cannot verify refereeing quality. Without any quality/impact/retraction/duplicate screen on recovered solo papers, the claim of a “reconfiguration of cognitive labor” risks conflating genuine substitution with an influx of low-value solo output. At minimum, report citation or journal-tier distributions for pre- vs post-2022 solo papers under the peer-reviewed filter, or flag quality as outside scope more prominently in the Abstract.
minor comments (6)
  1. [Figure 1; SI Section C] Figure 1 reports Δβ as change in annual slope of the solo share; clarify in the caption whether monthly fits are rescaled to annual units and how seasonal adjustment (used for mean authors in SI C) is handled for the solo share.
  2. [Materials and Methods; SI Section F] Eq. (1)–(2) for P_t(k) and d_it are clear, but the main text sometimes uses d_active without restating that it is predetermined and resets after every solo year; a one-sentence reminder would help non-specialist readers.
  3. [Figure 3; SI Section I] Figure 3a UMAP atlas is descriptive only (as stated), but the two language-defined clusters (Turkish, Indonesian) are easy to misread as topical; consider a footnote or inset noting they are excluded from the English-only axis analysis.
  4. [Acknowledgments; Materials and Methods] The Acknowledgments disclose extensive Claude use for code and manuscript revision. Given the paper’s topic, a brief Methods note on which analyses were AI-assisted vs human-verified would strengthen reproducibility norms the paper itself discusses.
  5. [Results; SI Figure S3] SI Fig. S3 heatmap significance uses bootstrap conditional on the observed source universe; state that limitation once in the main text when citing field-level significance counts (23/26 fields).
  6. [Title] Minor wording: title truncates “generative A” in the provided header; ensure the published title is complete (“generative AI”).

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: observational trend breaks, composition standardization, and embedding DiD are estimated from external OpenAlex data against an external timestamp, not forced by construction or self-citation.

full rationale

This is correlational bibliometric social science, not a first-principles derivation. The load-bearing quantities (field-level solo share, author-level P(solo|history), mean author count, SPECTER2 DiD displacement, keyword-axis projections) are computed from the OpenAlex snapshot; the November 2022 ChatGPT release is an external dating convention, not a fitted target. Pre/post slopes, composition cells (field × age × citations × productivity), history thresholds d_it ≥ k, and solo×post interactions are standard estimators whose outputs are not definitionally identical to their inputs. Semantic axes are keyword-anchored then validated on non-anchor papers and field rankings; the DiD nets field-wide drift rather than tautologically recovering the keywords. There are no self-citations that carry the central claim, no uniqueness theorems, no ansatz smuggled from prior author work, and no renaming of a known pattern as a new derivation. Interpretive reading of the break as LLM substitution is acknowledged by the paper as correlational (Limitations) and is a causal-inference concern, not circularity by construction. Score 0 is the honest finding.

Assumptions & free parameters 6 free parameters · 6 assumptions · 2 invented entities

The central claim rests on observational identification choices and measurement proxies rather than free physical constants or invented particles. Load-bearing inputs are the ChatGPT dating convention, OpenAlex field/author/venue constructions, peer-review and preprint operationalizations, composition cells for author probabilities, and SPECTER2/keyword axes for content. No new physical entity is postulated; the ‘solo paper as substitution probe’ is an interpretive instrument built from existing authorship metadata.

free parameters (6)
  • ChatGPT release cutoff (Nov 2022 / year 2022)
    Common temporal anchor for all pre/post slope comparisons; assignment of 2022 to windows is a design choice that defines the measured break.
  • Symmetric four-year and asymmetric rolling windows for Δβ
    Window lengths and whether the cutoff year opens the post period change estimated slope breaks; robustness is shown but the primary windows remain analyst choices.
  • History threshold k / d_active for never-recent-solo conditioning
    Thresholds k∈{1,2,3,5,10,15} define which authors enter history-conditioned probabilities; results strengthen with k but k is chosen, not estimated from theory.
  • Composition cells (field × academic age × citations × productivity)
    Binning scheme for direct standardization of author-level solo probability; alternative bins could reweight the composition-adjusted series.
  • Author output exclusion thresholds (>2000 lifetime works or >50/year)
    Hand-set filters to drop likely institutional placeholder profiles; affect who enters author-level analyses.
  • Content cohort sampling caps (≤5000 papers/field and top subfields) and UMAP/grid smoothing choices
    Stratified sample sizes, 300×300 grid, Gaussian σ=3, and mask threshold shape the density map localization (quantitative claims use full 768-d space).
assumptions (6)
  • domain assumption OpenAlex author IDs, primary-field assignments, and venue metadata are sufficiently accurate for field-level and author-year solo rates over 1990–2025 despite known disambiguation and coverage changes.
    Materials and Methods and SI Section L rely on OpenAlex profiles and continuous-venue panels; residual within-venue metadata change cannot be fully excluded.
  • domain assumption Venue-based peer-reviewed/core/preprint filters are valid proxies for publication quality strata even though OpenAlex lacks a true peer-review flag.
    SI Section A defines peer reviewed as article in journal/conference plus denylist; the paper notes this cannot verify refereeing.
  • domain assumption A solo-authored paper is work completed without credited human coauthors and thus a usable behavioral boundary for substitution of scientific labor.
    Abstract/Discussion treat solo authorship as the empirical probe; uncredited humans and AI remain invisible by construction.
  • domain assumption SPECTER2 embeddings plus keyword-anchored axes capture scientifically meaningful content shifts (especially computational vs equipment work) net of field-wide drift.
    Fig. 3 and SI Section I; validation is face/field-level, not ground-truth labels of LLM use.
  • ad hoc to paper Field differences in coauthor-task substitutability (writing/coding/stats vs lab/instrument work) explain heterogeneity in the solo rebound.
    Results interpret ordering as substitutability; occupation-based exposure scores are not mapped cleanly onto OpenAlex fields (Limitations).
  • standard math Piecewise linear trends with paper-count weights adequately represent the change in solo-share dynamics around the cutoff.
    Standard WLS/piecewise and event-study estimators in Results and SI Sections B–C, M.
invented entities (2)
  • Solo-specific content displacement (DiD of solo vs team cohorts in embedding space)
    purpose: Isolate post-2022 solo content movement net of field-wide drift shared with team papers.
    Constructed estimand (Fig. 3), not a physical entity; independent meaning depends on embedding validity and cohort sampling.
  • History-conditioned solo-authorship probability P_t(k) independent evidence
    purpose: Separate within-author switching into solo work from aggregate composition change.
    Defined in Materials and Methods Eq. (1)–(2); a measurement construct rather than a new scientific object with external existence.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Return of the solo author: The changing division of labor in science in the age of generative AI." pith.science (2026). https://pith.science/paper/PTPIBFBH

@misc{pith2026260710780,
  author       = {Pith},
  title        = {Pith review of: Return of the solo author: The changing division of labor in science in the age of generative AI},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PTPIBFBH}},
  note         = {Machine review of arXiv:2607.10780}
}
read the original abstract

Modern science has experienced a long shift from individual work to team production. Generative artificial intelligence (AI) might appear to extend this trajectory by lowering research costs and enabling larger-scale collaboration. Yet if tasks once performed by coauthors can be delegated to AI, the same technology may also weaken the need for collaboration in parts of the research process. Here, we examine this tension by moving beyond average team size and focusing on the solo-authored tail of the author-count distribution. Analyzing over 300 million works across 26 fields, we find that the decades-long decline in solo authorship halted and partially reversed with ChatGPT's public release in late 2022. We also reveal that this is an uneven phenomenon: it is strongest in fields where coauthors' work is more readily replaceable, and weak or absent in fields that depend on physical collaboration. At the individual level, the recovery is not explained by the entry of new researchers or by changes in field composition. Instead, the break appears among authors who had written only with others, including those with no prior solo publications, and among long-established authors as well as newcomers. Their solo papers stay close to their own coauthored work while narrowing in scope and shifting toward computational topics. Because a solo paper is work without credited human coauthors, this study offers an empirical probe of how generative AI can substitute for scientific labor, and evidence of a reconfiguration of cognitive labor within papers rather than of team size.

Figures

Figures reproduced from arXiv: 2607.10780 by the authors.

Figure 1
Figure 1. Piecewise-linear break in the monthly solo-author share at the public release of ChatGPT. Panels show the monthly share of solo-authored papers from January 2018 through December 2024 for all fields pooled and for five focal fields: Engineering, Com￾puter Science, Mathematics, Psychology, and Economics. Colors distinguish four coverage variants, with peer reviewed used as the main specification and the other variant… view at source ↗
Figure 2
Figure 2. Author-level (marginal) probability that an active author publishes at least one solo-author paper in year t, and its 2022 trend break. (A) Composition-adjusted P(≥ 1 solo in year t) by calendar year (2011– 2025), among authors with at least one publication in year t under the given publication-filter variant (clean all, peer reviewed, core, preprint); the dashed vertical line marks the public release of ChatGPT (No… view at source ↗
Figure 3
Figure 3. Going solo shifts content toward computational work: a displacement distinct from field-wide drift. Cohorts: A, solo-authored post-2022; B, solo-authored 2018–2021; C, team-authored post-2022; D, team-authored 2018–2021 (cohort construction, embedding, and estimation in Materials and Methods). Displacement statistics in panels b and c are computed in the full 768-dimensional SPECTER2 embedding space; the two-dimensi… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 3 linked inside Pith

  1. [1]

    death of the renaissance man

    BF Jones, The burden of knowledge and the “death of the renaissance man”: Is inno- vation getting harder?Rev. Econ. Stud.76, 283–317 (2009)

    2009

  2. [2]

    S Wuchty, BF Jones, B Uzzi, The increasing dominance of teams in production of knowledge.Science316, 1036–1039 (2007)

    2007

  3. [3]

    C Haeussler, H Sauermann, Division of labor in collaborative knowledge production: The role of team size and interdisciplinarity.Research Policy49, 103987 (2020)

    2020

  4. [4]

    S Fortunato, et al., Science of science.Science359, eaao0185 (2018)

    2018

  5. [5]

    (Cambridge University Press), (2021)

    D Wang, AL Barab ´asi,The Science of Science. (Cambridge University Press), (2021)

    2021

  6. [6]

    E Leahey, From sole investigator to team scientist: Trends in the practice and study of research collaboration.Annu. Rev. Sociol.42, 81–100 (2016)

    2016

  7. [7]

    M Thelwall, N Maflahi, Research coauthorship 1900–2020: Continuous, universal, and ongoing expansion.Quantitative Science Studies3, 331–344 (2022)

    1900

  8. [8]

    R Guimer `a, B Uzzi, J Spiro, LAN Amaral, Team assembly mechanisms determine collaboration network structure and team performance.Science308, 697–702 (2005)

    2005

Show all 39 references
  1. [9]

    S Milojevi ´c, Principles of scientific research team formation and evolution.Proc. Natl. Acad. Sci.111, 3984–3989 (2014)

    2014

  2. [10]

    L Wu, D Wang, JA Evans, Large teams develop and small teams disrupt science and technology.Nature566, 378–382 (2019)

    2019

  3. [11]

    P Cunningham, P Smyth, B Smyth, Shifting norms in scholarly publications: Trends in readability, objectivity, authorship, and AI use (arXiv:2510.21725) (2025)

    2025

  4. [12]

    F Xu, L Wu, J Evans, Flat teams drive scientific innovation.Proceedings of the National Academy of Sciences119, e2200927119 (2022)

    2022

  5. [13]

    S Noy, W Zhang, Experimental evidence on the productivity effects of generative artificial intelligence.Science381, 187–192 (2023)

    2023

  6. [14]

    KZ Cui, et al., The effects of generative AI on high-skilled work: Evidence from three field experiments with software developers.Management Science(2026)

    2026

  7. [15]

    A Korinek, Generative AI for economic research: Use cases and implications for economists.Journal of Economic Literature61, 1281–1317 (2023)

    2023

  8. [16]

    P Praus, A note on the topic of single-author articles in science.Scientometrics130, 3071–3088 (2025)

    2025

  9. [17]

    A Goolsbee, C Syverson, Fear, lockdown, and diversion: Comparing drivers of pan- demic economic decline 2020.Journal of Public Economics193, 104311 (2021). 35

    2020

  10. [18]

    R Chetty, JN Friedman, M Stepner, The economic impacts of COVID-19: Evidence from a new public database built using private sector data.Quarterly Journal of Eco- nomics139, 829–889 (2024)

    2024

  11. [19]

    T Eloundou, S Manning, P Mishkin, D Rock, GPTs are GPTs: Labor market impact potential of LLMs.Science384, 1306–1308 (2024)

    2024

  12. [20]

    EW Felten, M Raj, R Seamans, Occupational Heterogeneity in Exposure to Generative AI, (SSRN Working Paper), Technical report (2023)

    2023

  13. [21]

    G Tripodi, et al., Tenure and research trajectories.Proceedings of the National Academy of Sciences122, e2500322122 (2025)

    2025

  14. [22]

    D Kobak, R Gonz ´alez-M´arquez, E ´A Horv ´at, J Lause, Delving into LLM-assisted writing in biomedical publications through excess vocabulary.Science Advances11, eadt3813 (2025)

    2025

  15. [23]

    W Liang, et al., Quantifying large language model usage in scientific papers.Nature Human Behaviour9, 2599–2609 (2025)

    2025

  16. [24]

    M Kwiek, W Roszka, Are female scientists less inclined to publish alone? the gender solo research gap.Scientometrics127, 1697–1735 (2022)

    2022

  17. [25]

    T Lazebnik, A Rosenfeld, How lonely or influential is the lone wolf? an analysis of individual scholars’ solo-authorship dynamics.Scientometrics130, 3053–3069 (2025)

    2025

  18. [26]

    A Zeng, et al., Increasing trend of scientists to switch between topics.Nature Commu- nications10, 3439 (2019)

    2019

  19. [27]

    L McInnes, J Healy, J Melville, UMAP: Uniform manifold approximation and projec- tion for dimension reduction (arXiv:1802.03426) (2018)

    2018 arXiv

  20. [28]

    Social Studies of Science46, 417–435 (2016)

    V Larivi `ere, et al., Contributorship and division of labor in knowledge production. Social Studies of Science46, 417–435 (2016)

    2016

  21. [29]

    HW Shen, AL Barab ´asi, Collective credit allocation in science.Proceedings of the Na- tional Academy of Sciences111, 12325–12330 (2014)

    2014

  22. [30]

    S Jabbehdari, JP Walsh, Authorship Norms and Project Structures in Science.Science, Technology, & Human Values42, 872–900 (2017)

    2017

  23. [31]

    Nature Editorial, Tools such as ChatGPT threaten transparent science; here are our ground rules for their use.Nature613, 612 (2023)

    2023

  24. [32]

    M Andal ´on, C de Fontenay, DK Ginther, K Lim, The rise of teamwork and career prospects in academic science.Nature Biotechnology42, 1314–1319 (2024)

    2024

  25. [33]

    JP Alperin, J Portenoy, K Demes, V Larivi`ere, S Haustein, An analysis of the suitability of OpenAlex for bibliometric analyses (arXiv:2404.17663) (2024). 36

    2024 arXiv

  26. [34]

    JH Culbert, et al., Reference coverage analysis of OpenAlex compared to Web of Sci- ence and Scopus.Scientometrics130, 2475–2492 (2025)

    2025

  27. [35]

    J Priem, H Piwowar, R Orr, OpenAlex: A fully-open index of scholarly works, au- thors, venues, institutions, and concepts (arXiv:2205.01833) (2022)

    2022 arXiv

  28. [36]

    H Bouamor, J Pino, K Bali

    A Singh, M D’Arcy, A Cohan, D Downey, S Feldman, SciRepEval: A multi-format benchmark for scientific document representations inProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, eds. H Bouamor, J Pino, K Bali. (Association for Computationa...

    2023

  29. [37]

    (Association for Com- putational Linguistics, Stroudsburg, PA, USA), (2020)

    A Cohan, S Feldman, I Beltagy, D Downey, D Weld, SPECTER: Document-level rep- resentation learning using citation-informed transformers inProceedings of the 58th Annual Meeting of the Association for Computational Linguistics. (Association for Com- putational Linguistics, Stro...

    2020

  30. [38]

    OurResearch, We’re building a replacement for Mi- crosoft Academic Graph (https://blog.openalex.org/ were-building-a-replacement-for-microsoft-academic-graph/) (2021) Accessed July 11, 2026

    2021

  31. [39]

    OpenAlex, Author disambiguation, OpenAlex Support (https://help.openalex.org/hc/en-us/articles/ 24347048891543-Author-disambiguation) (n.d.) Accessed July 11, 2026. 37

    2026

Pith tools

Reviewed July 14, 2026 · model on record in the stance chip above.