The Solo-Authorship Rebound

Sources#

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

Summary#

Akira Matsui's single-author study of the full OpenAlex corpus (300M+ works, 126M peer-reviewed, 26 fields, 1990–2025) asks what generative AI does to the division of cognitive labor inside a paper, and answers it by looking somewhere nobody had: the left tail of the author-count distribution.

The framing is the paper's best idea. Two views make opposite predictions — the acceleration view (LLMs lower the cost of producing papers, so teams grow and gift authorship spreads) and the substitution view (LLMs take over writing, coding and statistical work that coauthors used to supply, so researchers finish alone). Mean team size cannot distinguish them, because the mean rises under both and is insensitive to the left tail. So measure the left tail. The finding: the decades-long decline in solo authorship halts or reverses at ChatGPT's public release, broadly and simultaneously across fields.

The paper's conclusion is deliberately narrow, and worth quoting because the headline invites over-reading: the results "do not settle this conflict in favor of one view. Rather, they assign the two views to different parts of the distribution: the substitution view to the left tail, where our evidence is direct, and the acceleration view to the upper tail, which our tail-based design does not measure." Solo papers remain a small minority; the level "continues to sit far below its pre-decline value." The claim is about the direction of change in the left tail, not its level.

Evidence note. empirical — the largest-N result in the vault's economics cluster, and the weakest identification among its empirical sources. There is no control group for the main estimate: every field and every author is treated at the same instant. See Identification, honestly below before citing any number as an effect of LLMs.

Why a solo paper is an instrument#

The methodological contribution, independent of whether the causal story holds:

"Whether and how researchers use LLMs is rarely disclosed and is hard to observe directly. A solo-authored paper, by contrast, is an observable behavioral trace: it is, by construction, work completed without credited human coauthors, so the set of papers that can be written alone marks the boundary of what can be completed without credited collaborators. When AI moves that boundary, the movement appears first in the left tail."

This sidesteps the disclosure problem that defeats every survey-based measure of AI use in science, and it does not depend on detecting LLM text. It is the same move the vault's realized-consumption and usage-telemetry sources make against survey instruments — find a trace the behavior leaves whether or not anyone admits to it.

The field ordering — the paper's real argument#

Under the main peer-reviewed specification, the trend break at the 2022 cutoff (Δβ, percentage points per year, SI Fig. S3):

Reverses strongly	~+1.7 cluster	No reversal
Engineering +2.5 · Business, Management and Accounting +1.9	Mathematics · Decision Sciences · Psychology · Computer Science	Chemistry −0.1 · Physics and Astronomy −0.0 · Arts and Humanities −0.5

23 of 26 fields turn positive; 22 are significant at p < 0.05. The three exceptions are exactly the table's third column, and only Arts and Humanities is still significantly declining.

The proposed reading is substitutability of the coauthor's execution work: fields where the collaborator supplied writing, coding or statistical support reverse; fields organized around laboratory work and instrument operation do not. The disconfirming test the paper runs on itself is the useful part — team-based organization as such is not what matters. Computer Science and Psychology run on the same large, hierarchical, grant-funded lab model as Chemistry and still reverse strongly; Physics and Astronomy, organized around large instrument-based collaborations, shows no rebound at all. Arts and Humanities is a third kind of null: solo authorship was already the norm, so there is little room for a substitution-driven rebound.

This ordering — not any single coefficient — is what should be cited. It parallels the independently documented field gradient in LLM uptake (Kobak et al. 2025 on excess vocabulary; Liang et al. 2025), and it agrees with the content evidence below, which uses no field labels at all.

Two estimators, two sets of numbers for the same field — do not mix them. The figures above are annual Δβ_c from symmetric four-year windows (pre [c−4, c−1], post [c, c+3]), SI Fig. S3. Fig. 1's monthly piecewise fit at the November 2022 release is a different estimator and reports Engineering +3.81, Psychology +2.11, Computer Science +2.09, Mathematics +1.77, Economics +1.25, all fields pooled +1.72 under the same peer-reviewed filter. Both are in the paper; quoting "Engineering +2.5" and "Engineering +3.8" as if one were a correction of the other is the error to avoid.

It is not composition, and not the usual solo authors#

The obvious deflation — the pool of authors changed — is the hypothesis the paper works hardest to kill, and this is the strongest section.

Composition-standardized author-level probability. An author-year panel, with P(≥1 solo paper in year t) direct-standardized over cells of field × academic age × lifetime citations × productivity, holding the observable author mix fixed. A pre-2022 decline of roughly −0.9 pp/yr halts or reverses at 2022 under every filter variant.
Conditioning on individual history. Restrict to authors with no solo paper in their k−1 most recent active years. The break is present at every threshold and strengthens with k: peer-reviewed Δβ rises from +0.50 (k=1) to +0.79 (k=15), where the estimand approximates the hazard of a first-ever solo-authored paper. The effect is at least as strong among the authors least accustomed to publishing alone — so this is authors switching into solo work, not habitual solo authors publishing more.
Seniority is not a barrier. Among authors with no recent solo publication, the rebound appears in every academic-age class, and under the main filters the largest break is in the most senior class (26+ years: +0.36 clean-all, +0.34 peer-reviewed). The gradient is weak, non-monotone, and reverses under the preprint filter — so "senior researchers are the ones switching" is directional at best.
Productivity is the sharpest cut. The rebound is largest among the least prolific authors (1–5 lifetime works, Δβ = +1.16 clean-all; +2.83 for preprints), smallest for the 6–20 group (+0.21). Read as LLMs lowering the fixed cost of producing a paper without a human coauthor.

Field-level and author-level breaks come apart, and the paper says so. Engineering has the largest field-level break and an author-level probability that does not move (−0.07/−0.14, CIs spanning zero); its field break attenuates from +2.47 to +0.65 inside continuously observed venues. So the biggest headline number in the paper is the one field where the mechanism is compositional, not within-author. The cleanest within-author breaks are Economics (+0.85/+1.27), Psychology (+0.66 to +1.34) and Computer Science (+0.75/+0.80).

What the solo papers are#

The content analysis is the best-identified piece of work in the paper, because it is the only part with a control group: four cohorts (solo/team × pre-2022/post-2022, 636,454 papers) embedded with SPECTER2, so field-wide drift since 2022 is differenced out and only the solo-specific displacement remains.

A tilt toward computation, and nothing else. Of four pre-specified keyword-anchored semantic axes, only computational-vs-experimental moves: +0.040 s.d., P = 4×10⁻⁸. Data-rich (+0.016, P=0.15), review-vs-original (+0.014, P=0.07) and theoretical-vs-applied (+0.010, P=0.19) are flat. Recovered solo papers are ordinary original research that has tilted toward computation — not reviews, not opinion, not a retreat into theory.
Narrower, not elsewhere. Comparing authors' own post-2022 output: solo-writers' content breadth (mean pairwise cosine distance among their own papers) is 0.089 vs 0.115 for persistent coauthors — 23% narrower (adjusted −0.028, P < 10⁻¹⁶, n = 35,893). But exploration — distance from the author's own pre-2022 collaborative centroid — is statistically indistinguishable (+0.001, P = 0.07, n = 45,642), an order of magnitude smaller than the breadth effect.

That pair is the discriminating test. A change of interest predicts movement away from past collaborative content; carrying a slice of the old work alone predicts staying put while covering less. The data match the second on every dimension measured. The author-level comparison is a covariate-adjusted descriptive contrast, not a matched design — but the exploration null itself argues against the most natural selection story (that authors go solo because their interests already moved).

Identification, honestly#

The design. An interrupted time series on the universe of papers, with the interruption placed at a date chosen because it is "the only sharp timestamp common to all fields and authors." The paper is explicit that this is a dating convention and that "our evidence is correlational," resting on three patterns pointing the same way: field differences in substitutability, conditioning on author histories, and the contrast with non-reversing fields.

What is genuinely strong.

The break survives four independent coverage filters, two window definitions with no shared construction (symmetric 4-year vs asymmetric [τ−1,τ] against a 7-year baseline), a cutoff scan over 2018–2022, composition standardization, and history conditioning.
The cutoff scan is the best single defense against "you picked the date." Positive Δβ_c appears at earlier cutoffs too — but those are decelerations of a decline that remained ongoing. Only at c = 2022 does the post-window slope itself turn positive in a broad set of fields (11 of 26, versus at most 3 at any earlier cutoff).
The event study shows a decline steepening right up to the break (−7.2 pp cumulative by 2022, with the year-over-year change reaching −2.9 pp in 2022) and then stopping dead: +0.1 (2023), −0.6 (2024), −0.3 (2025), with no earlier year showing a flattening of comparable size. The pre-trend is deliberately left visible.
The pandemic-rebound alternative is tested and fails on its own predictions: the fields that dipped deepest at c=2020 (Arts and Humanities −1.3, Chemistry −1.2) are precisely the ones that do not rebound; Mathematics and Decision Sciences show no dip yet sit in the top 2022 cluster; and the recovery is not biomedical-led, which reversion-to-trend requires.
A companion break in the mean number of authors per paper (its decades-long rise flattens or turns down after late 2022 in most fields) means the finding is not an artifact of the tail-based metric.

What is genuinely weak, in descending order of how much it should discount the headline.

Roughly half the pooled break is venue composition. Inside a balanced panel of sources publishing in every year 2018–2024 (45,621 sources carrying 72–77% of peer-reviewed papers), the pooled break falls from +1.72 [+1.12, +2.32] to +0.75 [+0.41, +1.10]; within those venues the pre-2022 decline of −1.18 pp/yr flattens to −0.42 rather than turning positive. The break stays positive and significant, and the cross-field ordering is preserved (Spearman 0.84 peer-reviewed, 0.91 core) — but anyone quoting +1.72 is quoting a number half of which is which journals OpenAlex indexed.
No untreated unit. The field ordering is the only cross-sectional variation, and the paper concedes it is "an ordering rather than a quantitative test": occupation-based exposure scores do not map onto OpenAlex fields, and 26 fields have little power. The cross-field pattern persuades by shape, not by significance.
The database changed under the panel. OpenAlex retired Microsoft Academic Graph at the end of 2021 and overhauled author disambiguation in July 2023 — both inside the estimation window and both plausibly affecting measured author counts. The balanced-venue panel bounds the venue-turnover margin; the paper states plainly that metadata changes within continuously indexed venues cannot be excluded. This is the confound with no answer in the paper.
Timing checks are reassuring on point estimates and not on precision. The monthly donut (dropping Nov 2022 – Jun 2023) leaves the estimate close to published (+1.38 vs +1.72) but with a CI of [−0.03, +2.79] — i.e. it no longer excludes zero. The annual donut excluding 2022 is tight (+0.82 [+0.79, +0.85]). One check cuts the other way: under unrestricted coverage with the post window limited to 2023–24, the estimate turns negative (−1.50), which is why the headline variants are venue-restricted.
Quality is not measured at all, and the rebound is largest in the preprint variant — consistent with substitution and equally consistent with an influx of low-cost, possibly LLM-generated solo output. The paper's defense is that the break survives under peer-reviewed and core filters, so a preprint-confined influx cannot account for it; its own concession is that venue-based filters cannot verify refereeing and cannot rule out mill-style output inside indexed venues.

Assessment. "Something changed in the left tail of the author-count distribution at the end of 2022" is well established and heavily stress-tested. "LLMs caused it" is not identified and is not claimed to be. The load-bearing evidence for the mechanism is the disciplinary ordering plus the content signature, which agree with each other and were produced by independent instruments (one uses field labels, one uses none) — that agreement is worth more than any coefficient in the paper. Treat individual Δβ values as descriptive magnitudes with a large composition component, and the ordering as the finding.

The external check that cuts both ways#

The most interesting robustness exercise, and a genuinely surprising result. Re-estimating Liang et al.'s population-level LLM-modification rate (α) by author count on the same public corpus, multi-author papers carry at least as much LLM-modified writing as solo papers in every arXiv field (+1.7 to +3.7 pp, largely non-overlapping intervals); bioRxiv is the only reversal and rests on ~300 solo papers. The regional gradient (steepest for non-native-English corresponding authors, highest for China) points at English-language polishing by people who are more likely to sit on larger teams.

The author uses this correctly and narrowly: writing assistance is pervasive rather than concentrated in the solo tail, so surface-level editing cannot by itself explain a shift specific to solo authorship. But read from the other side, it removes the most obvious direct corroboration one would want — there is no measurement anywhere showing that the authors who went solo are the ones using LLMs. α measures text-level modification, not substitution for a collaborator's contribution, and the paper says so.

What this does to the vault's other claims#

It is the closest thing yet to the commons argument observed as behavior rather than forecast — in one profession, on the mechanism side only. Lovett's structural prediction is that AI removes entry-level work and with it the pathway by which expertise regenerates. Matsui's Implication 2 states the same mechanism for science and this paper measures its precondition: "The more senior researchers can hand execution work to LLMs instead of to junior coauthors, the weaker this training pathway through collaboration may become." A solo paper is one apprenticeship slot that did not exist. Three honest limits on that reading: the seniority gradient carrying it is weak and filter-dependent; solo papers remain a small minority of output; and nothing here measures whether any junior actually learned less. This is the removal of the slot, not the depletion of the commons — exactly the distinction Lovett insists on.

It sharpens AI as Primary Author by inverting its visibility. In software, the authorship shift is measurable — acceptance rates, PR provenance, telemetry. In science it is invisible by construction, which is Matsui's Implication 1: contribution statements and collaboration networks were the two instruments for inferring who did what, and "for solo-authored papers in which an LLM supplies the execution work, neither approach applies: the contributing agent appears in neither the author list nor the collaboration network." The credit gap moves from unlisted humans (long documented) to machines that cannot be listed at all. Both pages land on the same asymmetry: authorship moved, accountability did not — the solo author signs for everything.

It is role averaging at the level of a paper, with a twist the page does not predict. The solo author absorbs the coauthors' execution roles and keeps the judgment layer, which is the thesis exactly. But the measured consequence is contraction: 23% narrower content breadth, with no movement into new territory. The vault's other averaging cases show the residual judgment half growing (recruiters' evaluation stage 2.62 → 7.24 days in Controlled Variance: AI's Edge as Reduced Dispersion). Here the averaged role covers less ground, not more. Whether that is scope discipline or capacity limit is unmeasured, and it is the sharper of the two open questions this source leaves.

It complicates Returns to Expertise in Agentic Coding on a margin that page cannot see. Expertise amplifies an agent 2× in actions and 5× in output; here the switch into solo work is largest among the least prolific authors (+1.16 vs +0.21 for the 6–20 group) and only weakly largest among the most senior. These are different outcomes — the probability of publishing without a coauthor is not productivity or quality — so this is not a contradiction. It is a levelling signature on the fixed cost of producing output at all, sitting beside an amplification signature on output per prompt, and nothing in either measures whether the low-output authors' solo papers are any good.

It is a field-level exposure ordering built from behavior, which is what Exposure Taxonomy: Observed, Theoretical, Reported, Anticipated finds the existing instruments lack. Seven occupational instruments disagree because they differ by data source rather than construct; this one is derived from an observed outcome instead of a task rating, and it ranks 26 scientific fields rather than 872 occupations. The paper concedes the two cannot be joined (Eloundou and Felten scores "do not map cleanly onto the OpenAlex fields"), which is itself the taxonomy problem restated in a domain where a behavioral measure exists.

It fits the division-of-labor boundary from the other end. That page found, inside one research pipeline, that LLMs handle the mechanical, quote-anchored work and fail at interpretive synthesis. This paper detects the population-scale shadow of the same split: the execution layer — coding, data handling, statistical analysis — is what gets handed over, and what the solo author keeps is the part that was never delegable.

Connections#

The Tragedy of the Cognitive Commons — the mechanism's precondition observed at population scale in one profession: a solo paper is a coauthorship slot that did not exist, which is the training pathway Lovett argues AI removes. Removal of the slot, not depletion of the commons — nothing here measures whether anyone learned less
AI as Primary Author — the same authorship shift with the visibility inverted: measurable in code telemetry, invisible by construction in science, where the substituting agent appears in neither the author list nor the collaboration network
Role Averaging, Not Role Elimination — averaging at the paper level, with a result the page does not predict: the averaged solo role is 23% narrower in content, not broader
Organizational Complements to AI — where the HAT prediction ledger lives; this supplies P1's first labor-composition outcome that breaks sharply at a capability event rather than ramping
Returns to Expertise in Agentic Coding — the levelling counterpart on a different outcome: the switch into solo work is largest among the least prolific authors, while expertise is what amplifies the agent per prompt
Exposure Taxonomy: Observed, Theoretical, Reported, Anticipated — a substitutability ordering derived from observed behavior rather than task ratings, over 26 scientific fields; the paper concedes it cannot be mapped onto the occupational instruments
LLM-Assisted Grey-Literature Theory Building — the same execution/interpretation boundary seen inside one research pipeline; this is its population-scale shadow
Controlled Variance: AI's Edge as Reduced Dispersion — the vault's randomized substitution case, and the contrast on what happens to the residual: there the retained judgment stage expanded 2.8×; here the retained scope contracted 23%
Verification as the New Bottleneck — the solo author is the sole verifier of work an LLM executed, with no coauthor left to catch it

Open Questions#

Roughly half the pooled break is venue composition (+1.72 → +0.75 pp/yr inside continuously observed venues), and OpenAlex changed its author-disambiguation pipeline in July 2023 — inside the window. Does the break survive in a corpus with stable curation and stable disambiguation (Scopus, Web of Science, or a single large publisher's internal records) over the same period?
Solo papers narrow 23% in content breadth while showing no movement toward new territory. Is that scope discipline (the author does what they can verify alone) or capacity limit (the LLM covers the execution but not the range a second mind supplied) — and does the quality of solo output diverge from coauthored output on citations, replication or retraction? The paper measures quantity and content and explicitly not quality.
The break's attribution rests on a cross-field ordering the author concedes is an ordering, not a test. If the halt is LLM-driven it should track LLM capability, so the ordering should shift as models improve — fields whose execution work is newly automatable (lab protocol design, instrument control) should join late. If it is a one-time re-sorting, the ordering freezes and the halt decays.

Sources#

Return of the solo author: The changing division of labor in science in the age of generative AI — Akira Matsui (Center for Computational Social Science, Kobe University), Return of the solo author: The changing division of labor in science in the age of generative AI, arXiv 2607.10780 (2026-07-12, 37pp, sole author), empirical. §Results (the field ordering and the substitutability reading; the 23-of-26 / 22-significant counts; the 11-of-26 post-window-positive count at c=2022); §"Author composition alone does not explain the trend break" (composition standardization, the k-threshold conditioning); §"The mechanism is consistent with LLM substitution" (productivity and seniority profiles); §"Going solo shifts content toward computational work" (the four-cohort DiD, the semantic axes, the breadth/exploration contrast); §Discussion (the left-tail/upper-tail assignment of the two views); §Implications (authorship and credit; the training-pathway concern); §Limitations (the correlational concession, the quality/paper-mill caveat, the OpenAlex-pipeline caveat); SI §B (the symmetric-window Δβ_c construction), §D (the pandemic-rebound test), §E (the asymmetric-window robustness), §F–H (thresholds, productivity/age, author-level field breaks), §I–J (embedding cohorts and the author-level scope comparison), §K (the Liang et al. re-estimation by author count), §L (the balanced-venue panel and the +1.72 → +0.75 attenuation), §M (event study, donut specifications, the −1.50 unrestricted-coverage caveat). Figure-heavy source, quoted from prose and captions. The document has exactly one table (SI Table S1, the semantic-axis keyword sets) and it carries no results — every trend estimate lives in a figure or in prose, so the vault's table-parse hazard does not apply here. Fig. 1 (monthly piecewise fits, per-panel Δβ per coverage variant) and SI Fig. S3 (the Δβ_c heatmap, fields × cutoff years × filter variants) were opened and reconciled against the prose before writing; the two estimators disagree by field and both sets are recorded above deliberately. Parse note: ten mojibake repairs were applied to the bibliography and one to body prose at ingest (Barabási ×2, Guimerà, Milojević, González-Márquez, Horvát, Larivière ×2, Andalón) — cosmetic, affecting no quantitative claim