LLMs and the Organization of Scientific Work
A · Strong evidence University of Strasbourg economists use CRediT contribution statements from about 120,000 PLOS ONE papers in event-study designs with trend-break tests, separating new from established teams; observational but method and data are clear.
Summary based on the full text·Original: about 80 minutes to read
In short
An arXiv paper (2 November 2026) by Stefano Bianchini and Georgiana-Ioana Coroama. If LLMs widen what one researcher can do, teams should need fewer co-authors per task and divide work more sharply. Using CRediT statements of who did what in PLOS ONE papers, they find participation in tasks falling below its prior trend after ChatGPT, more specialisation, and less overlap between co-authors. The change is concentrated in teams whose members had not worked together before.
Key numbers
- approximately 120,000PLOS ONE articles with CRediT statements, 2018-2025
- 2.6 percentage pointsdrop in team participation in ideation by 2025 relative to 2022
- 3.0 percentage pointsdrop in writing participation by 2025 among newly formed teams
- 0.087fewer task domains per author by 2025 in new teams, about 0.60 fewer author-task assignments per paper
As reported by the source.
English·Web page·opens arxiv.org
What it reports
- Before LLMs, writing involved 4.85 contributors on average (77.8% of team members), execution 4.37, ideation 4.07, data and coding 2.54.
- By 2025, participation in ideation, execution and coordination fell about 2.6, 0.9 and 1.4 points from 2022, all below extrapolated pre-LLM trends (break tests p < 0.001, 0.002, 0.044).
- Newly formed teams (n = 102,228) drive the pattern: writing -3.0 points, ideation -2.7, execution -1, coordination -1.6 by 2025.
- Authors in new teams cover 0.087 fewer task domains, about 0.60 fewer author-task assignments per paper at a mean team size of 6.90; task overlap falls 3.7% of its baseline.
- Established teams (n = 14,983), where all members had collaborated before, show much greater inertia.
How they got it
Event studies on paper-level shares of team members credited with each CRediT task, with field fixed effects and team controls, 2022 as reference year, and break tests against pre-LLM trends. The sample covers about 96% of the eligible corpus.
Limits
- Observational: LLM use is not measured per paper; timing around ChatGPT's release identifies the effect.
- One journal; CRediT statements are self-declared and may follow norms that themselves changed.
- Effects are a few percentage points.
Why it matters for companies
The same shift is likely in R&D, analytics and consulting teams: fewer people touch each task and roles narrow. New teams reorganise fastest; established teams keep old habits, which is where deliberate redesign of who does what pays. Contribution logs, like CRediT here, are a cheap way to see whether AI is changing team structure before productivity numbers show anything.
Summary based on the full text. Read the original for the full argument and data.Read the original on arxiv.org
- Tags
- Future of work, Science