Scientific Idea Diffusion Decline

Cross-field transmission of new scientific ideas has been contracting for four decades. New concepts increasingly stay near their field of origin, and a growing body of evidence ties this to specialization — both in who produces ideas and in the language used to introduce them. Diffusion, not just discovery, looks like a binding constraint on the social value of science.

The core finding (Berkes & Gaetani, 2026)

Using the full PubMed corpus, Berkes & Gaetani identify newly introduced scientific concepts and measure diffusion breadth as the semantic distance between the papers that introduce a concept and the papers that adopt it within ten years. Concepts introduced in 2009 diffused roughly half a standard deviation less broadly than those from 1980 — and this is not an artifact of science becoming more homogeneous (placebo simulations show articles have, if anything, grown modestly more dispersed). ¹

The mechanism is a communication tradeoff:

• Comprehension of an idea declines with scientific distance from the originating field, and declines more steeply the deeper the idea sits from basic principles.
• Accessible language has a bridging effect (distant audiences can understand the idea) and a dilution effect (precision drops for all audiences, including nearby peers).
• As knowledge accumulates and new ideas build on deeper prior work, dilution costs rise faster than bridging benefits — so scientists rationally choose more specialized language, and diffusion narrows endogenously. ²

Empirics line up with the model: knowledge depth of new concepts has risen steadily, introduction language has grown less accessible, and deeper/less-accessible concepts diffuse less broadly. An own-field bias in institutions (tenure, peer review, grants all judge within-field impact) pushes accessibility below the social optimum — a wedge between private and social returns to diffusion. Their policy menu: weight cross-field influence in funding, fold reach into promotion, reward accessibility in journals, and invest in translation infrastructure — including LLMs as idea-translators.

The corroborating literature

• Park, Leahey & Funk (Nature, 2023): across 45M papers and 3.9M patents over six decades, work is increasingly less disruptive (CD index: does citing work also cite the focal work’s predecessors?). Declines are universal across fields and linked to a narrowing in the use of prior knowledge — the same constriction Berkes & Gaetani measure from the language side. ³
• Clancy (2026), an innovation-economics specialist, contextualizes: new medical phrases increasingly come from scientists who have worked in fewer fields, phrases from more specialized scientists diffuse less far, and — the compounding dynamic — new ideas are introduced in steadily less accessible language. This sits alongside “ideas are getting harder to find” (Bloom et al.): rising specialization impedes exactly the cross-field connection-making that drives innovation. ⁴

Why it matters here

Diffusion is a comprehension barrier, not an access barrier — echoing Cohen & Levinthal’s absorptive capacity: you can’t use what you can’t process. That reframes a lot:

• Falling returns to R&D may be partly a transmission failure, not a discovery failure — a constant flow of good ideas yields shrinking social returns when each reaches fewer fields.
• Accessible synthesis (living literature reviews, good textbooks, translator-LLMs, this wiki) is not mere popularization — on this evidence it is a measurable lever on the value of the whole research enterprise.

Cross-domain connections

• filter-bubble — two flavors of the same disease: Pariser’s bubble is personalization narrowing serendipitous exposure; this is specialization narrowing cross-field exposure. Both are comprehension/exposure scheduling problems, the domain-scale dual of spaced-repetition’s personal-scale one.
• institutionally-constrained-technology-adoption — same deep structure: actors (rulers / scientists) rationally choose against the socially better option (superior tech / accessible language) because internal incentives (coup risk / own-field peer reward) dominate. Institutions wedge private and social returns.
• john-von-neumann — the existence proof for cross-field transmission: game theory, computing, quantum mechanics, automata theory, and anticipating DNA’s logic from one mind. The diffusion measure in this literature is, in effect, quantifying how rare von Neumann-scale bridging has become.
• demand-governance-wedge — another case where individually rational demand/adoption diverges from collective preference, measured with survey-experiment rigor.
• chinese-innovation-patent-ecosystem — patent text is the other great corpus where idea-flow gets measured; China’s falling citation-dependence on foreign knowledge is diffusion decline at the international scale.
• technological-singularity — Vinge predicted idea spread would accelerate toward the Singularity (“lead time seems more like eighteen months”); this literature is quantitative evidence in the opposite direction, and against a collapsing Δ.
• tabbys-star — the rise-and-resolution of the WTF anomaly tracks how a scientific claim diffuses, peaks, and is displaced by better-supported explanations
• aboriginal-oral-tradition-deep-time-memory — the dismissal of oral tradition as unreliable was itself a dominant paradigm overturned by empirical testing

Open questions

• Biomedicine-only evidence so far (PubMed corpus) — does the contraction replicate in physics, CS, economics itself?
• Will LLM translation actually widen diffusion, or just flood distant fields with low-quality comprehension?
• Is declining disruptiveness (Park et al.) caused by narrowing diffusion, or are both symptoms of a deeper specialization equilibrium?

Sources

• Berkes, E. & Gaetani, R. (2026). “The Decline in the Transmission of Scientific Ideas.” NBER Summer Institute working paper. 2026 — common words
• Park, M., Leahey, E., & Funk, R. J. (2023). “Papers and patents are becoming less disruptive over time.” Nature, 613, 138–144. 2023
• Clancy, M. (2026). “The slowdown in idea diffusion.” Abundance and Growth (expert explainer, June 2026). 2026 — The slowdown in idea diffusion - by Matt Clancy

Footnotes

1. 2026 — common words ↩

2. 2026 — common words ↩

3. 2023 ↩

4. 2026 — The slowdown in idea diffusion - by Matt Clancy ↩