Study of 120,000 papers finds LLMs make scientific teams more specialized

Bianchini and Coroama of the University of Strasbourg have released a preprint, "LLMs and the Organization of Scientific Work," drawing on author contribution (CRediT) data from roughly 120,000 PLOS ONE papers published between 2018 and 2025 to examine how the division of labor in research teams changed after ChatGPT's release. The core finding: after adopting AI, teams became more specialized rather than more generalist — each author handled fewer tasks, and task overlap between co-authors declined. This seems counterintuitive, since LLMs should broaden individual capabilities; but the authors argue it makes sense if excessive generalization hampers team coordination.

Confirmed

• The paper analyzed data from roughly 120,000 PLOS ONE papers published between 2018 and 2025
• Participation declined across most CRediT tasks, most notably in ideation and writing; the authors explain that teams no longer need multiple people performing the same task redundantly
• Team specialization increased: fewer tasks per person and less task overlap, with newly formed teams showing the largest shifts
• The researchers use a conceptual framework to argue that LLMs expand individual capability, reduce reliance on complementary human labor, and enable more specialized divisions of labor

Not yet confirmed

• The paper does not examine LLMs' impact on research productivity, quality, or novelty — the authors name this as a clear direction for follow-up work
• The claim that specialization boosts efficiency but reduces learning opportunities remains the authors' speculation

Why it matters

The study provides large-scale empirical evidence that LLMs are reshaping how science is organized. The authors caution that increasingly specialized roles may improve efficiency while cutting learning opportunities, hitting early-career researchers hardest — by automating intellectual activity, LLMs may make scientists more productive while stripping away the very experience they need to develop scientific judgment.

2026-11-05 ~ 2026-11-05 · 6 related posts

• [source] New preprint analyzes 120k papers to show how LLMs reshaped scientific division of labor — JMateosGarcia · 2026-11-05
• Participation in most CRediT research tasks dropped after ChatGPT, especially ideation and writing — JMateosGarcia · 2026-11-05
• Counterintuitive: AI makes research teams more specialized, not individually broader — JMateosGarcia · 2026-11-05
• [source] AI makes research teams more specialized, not less: counter-intuitive finding — JMateosGarcia · 2026-11-05
• Authors speculate LLM-driven specialization may hurt early-career researcher learning — JMateosGarcia · 2026-11-05
• [source] 120k-paper study: LLMs made research teams more specialized, with new teams changing most — JMateosGarcia · 2026-11-05