The last fifty years produced one of the most robust findings in the economics of science: team-authored papers are rising and solo authorship is in secular decline. Jones (2009) explains this through the “burden of knowledge” — as the cumulative stock of scientific knowledge expands, researchers must specialise ever more narrowly to reach the frontier, requiring collaboration to aggregate expertise. Kuld and O’Hagan (2018) trace this collapse in economics specifically, from roughly 50% solo-authored papers in 1996 to barely 25% by 2014.
Generative AI disrupts this equilibrium. If a large language model can perform, at near-zero marginal cost, the routine cognitive tasks previously supplied by junior co-authors — drafting, literature synthesis, reformatting, translating statistical output into readable prose, and now writing and debugging code — the marginal value of adding a human collaborator falls sharply. The knowledge production function shifts: what previously required a team can be accomplished by a suitably equipped individual.
| Finding | Estimate | Context |
|---|---|---|
| Solo-authorship rise (DiD) | +2.8–3.7 pp | Writing- vs lab-intensive fields, post-2022Q4 (arXiv +2.8pp, p = 0.011; SSRN +3.7pp, p = 0.014) |
| Co-authors displaced (synth.) | −0.47 | Economics on arXiv vs synthetic counterfactual (p < 0.001) |
| Writing channel: GenAI wave (2022Q4–23Q2) | +2.0 pp** | ChatGPT: economics and social sciences activate first |
| Coding channel: Codex break (May 2025) | +6.7 pp*** | CS/stats solo 15.1% → 24.7%; DiD vs non-coding fields |
| Agentic wave (2024Q2–25Q3) | +2.7 pp*** | Statistics and ML activate only in this wave |
| Writing-adoption ≠ solo work | r = −0.86 | Fields adopting AI-writing most (physics, bio) go solo least |
| Author-level mechanism | −0.12 / SD | AI-style abstract authors form smaller teams (p = 0.031) |
| Productivity paradox | −3.3% p.a. | AI-exposed authors publish less, not more, post-ChatGPT |
** p < 0.05; *** p < 0.01. Standard errors clustered/robust at the field level. Parallel trends hold pre-2022 across all primary specifications.
Prior work is either purely descriptive (Ben-Zion et al. 2026; Cunningham et al. 2025) or focused on output quantity rather than collaboration structure (Kusumegi et al. 2025, Science). This paper adds causal identification through DiD, a synthetic control, and within-platform interrupted time-series — all pointing in the same direction.
The SSRN economics and finance series declined from 38% solo-authored in 2018 to 30% in 2022, tracking the Kuld and O’Hagan (2018) secular trend exactly. After 2022Q4 it reversed, reaching 40% by 2025 — the first documented uptick in decades. The reversal is absent in lab-intensive fields and survives platform-composition controls.
Writing substitution (ChatGPT, Nov 2022) drives the effect in economics and social sciences; coding substitution (Codex, May 2025) then activates CS and statistics, whose solo share jumped 15.1% → 24.7% within eight months (+6.7 pp DiD) and rose +14.7 pp cumulatively by 2025 — despite no response to the writing wave, a timing pattern no COVID story predicts. The cross-field correlation between AI-writing adoption and solo-share change is −0.86, ruling out a naive “more AI text → more solo.”
Authors with more AI-characteristic abstracts form teams ~0.12 co-authors smaller per SD (p = 0.031); economics adopts AI-writing least yet goes solo most — the effect runs through substituting writing-task collaborators, not adoption per se.
The Codex agent does for coding-intensive fields in 2025 what ChatGPT did for writing-intensive fields in 2022: the same mechanism on a different task.
AI-exposed authors publish 3.3% fewer papers per year post-ChatGPT, consistent with reallocation toward fewer but more ambitious solo projects.
Jones (2009) Rev. Econ. Stud. 76(1). • Kuld & O’Hagan (2018) Scientometrics 114(3). • Kusumegi et al. (2025) Science 390. • Kobak et al. (2025) arXiv:2406.07016. • Liang et al. (2024) arXiv:2403.07183. • Liang et al. (2025) Nature Human Behaviour 9. • Noy & Zhang (2023) Science 381. • Wuchty, Jones & Uzzi (2007) Science 316.