Brief 004 · Science & discovery

Is science getting faster—or merely producing more?

Inputs, papers, reliable findings and useful translation are different things. This briefing keeps their denominators apart, then asks where AI has demonstrated acceleration.

Official series updated 2026-07-13World → Netherlands → selected countriesInspect sources
  1. 01Where are we now?
  2. 02What direction?
  3. 03Who benefits?
  4. 04What else explains it?
  5. 05What changes our mind?

01 · Where are we now?

Five clocks,
not one speedometer.

No composite “science speed” score. Each layer answers a different question and can move without the others.

Inputs

Capacity

R&D money and researchers make work possible; they do not guarantee discovery.

Volume

Recorded output

Indexed articles count publication, not importance or reliability.

Reliability

Results that travel

Replication, inspectable data and software test whether findings hold.

Translation

Use beyond laboratory

Clinical and regulatory gates remain after discovery.

Acceleration

Demonstrated gain

Benchmarks and experiments—not forecasts—show bounded task improvement.

World

Structural capacity

Global intensity and output, with unequal country contributions.

Europe / Netherlands

Institutional context

Dutch investment plus European open-science standards.

Selected countries

Mechanism comparison

United States scale and Korean intensity; not a league table.

02 · What direction are we moving?

Inputs and papers rose.
Meaning needs another measure.

World Bank series show world R&D intensity at 2.60% of GDP and 3.27 million indexed science and engineering articles in 2023. Coverage and definitions travel with those values.

Official annual observations

R&D expenditure · latest available

View chart as data
R&D expenditure, latest available 2023 observation
GeographyYearValueUnit
World20232.5974% of GDP
Netherlands20232.2672% of GDP
United States20233.4472% of GDP
Korea20234.9435% of GDP

Source: World Bank / UNESCO UIS, R&D expenditure, 2000–2023 ↗ · Scope & caveats ↓

Paired publication figure

More knowledge—or more throughput?

Annual observations
Panel A · Global scientific and technical journal articlesWorld Bank WLD · 2000–2023
Global scientific and technical journal articles, 2000 to 2023Annual article count rises from about 1.07 million in 2000 to 3.27 million in 2023.200020231.07m3.27m
Complete annual table
World scientific and technical journal articles, annual count
YearArticles
20001,070,967
20011,109,574
20021,154,717
20031,213,775
20041,324,894
20051,484,401
20061,570,081
20071,636,696
20081,725,502
20091,827,111
20101,909,442
20112,003,515
20122,062,566
20132,125,848
20142,195,344
20152,252,543
20162,336,757
20172,425,390
20182,522,113
20192,708,531
20202,868,013
20213,111,013
20223,220,952
20233,271,296

Source: World Bank / NSF NCSES, journal articles, 2000–2023 ↗ · Scope & caveats ↓

Panel B · AI-related computer-science publicationsStanford AI Index · 2013–2024
AI-related computer-science publication count and share, 2013 to 2024Two vertically separated, independently scaled plots share a time axis. Count rises from 101,885 to 257,891; share rises from 21.63 to 40.94 percent. A 2022 marker provides context only.Count101,885257,891Share of CS21.63%40.94%2022 · context only20132024

Growth predates ChatGPT. The 2022 marker is context, not attribution. 2024 count growth was 6.3%. These are papers about AI, not AI-generated or AI-assisted papers. No trustworthy global annual AI-generated-paper series exists: disclosure is inconsistent and detectors are unreliable.

Complete annual table
AI-related computer-science publications and share of CS publications
YearPublicationsAI share of CS
2013101,88521.63%
2014104,40921.75%
2015105,74121.67%
2016107,26921.97%
2017116,93723.42%
2018139,71125.84%
2019164,19928.71%
2020181,10730.31%
2021204,05436.01%
2022202,73439.03%
2023242,66441.76%
2024257,89140.94%

Source: Stanford HAI, 2026 AI Index Figure 1.6.1, 2013–2024 ↗ · Scope & caveats ↓

Method boundary: OpenAlex + CSO Classifier v3.3; English-language publications with a CS label. Cross-disciplinary AI may be undercounted; topic labels overlap; metadata and venue assignment lag. Category and venue switchers omitted because labels do not stack cleanly.

Interpretation: Volume measures activity and throughput—not quality, reliability, translation, AI authorship or social benefit. Read it with replication, open-science, FDA translation and AlphaFold evidence below.

Boundary: article counts can rise because researcher numbers, indexing, collaboration or incentives change. They are not a discovery-speed measure.

03 · Who benefits or carries cost?

Capacity clusters.
Open infrastructure changes who can participate.

R&D intensity differs sharply even among high-income countries. Absolute article counts additionally reflect country size. Open data and research software can lower access friction, but UNESCO's recommendation measures a standard—not adoption or impact.

Distribution caveat

Intensity and scale answer different questions.

Korea's R&D share exceeds the U.S. share; U.S. article volume is much larger. Neither proves higher discovery quality.

04 · Competing explanations

More output can mean
more science—or more counting.

Translation · regulatory process

Discovery is first gate, not finish line.

FDA separates discovery, preclinical work, clinical research, review and post-market monitoring. Faster modelling does not erase safety and efficacy work.

Rivals: more researchers, broader indexing, collaboration and publication incentives can all lift output. Higher complexity, cost and safeguards can slow translation even when discovery tools improve.

05 · What would change our conclusion?

Show reliable gains
across whole pathways.

Demonstrated AI acceleration · benchmark

AlphaFold changed protein-structure prediction.

Across 87 CASP14 domains, median backbone error was 0.96 Å versus 2.8 Å for the next-best method. Strong task evidence; not experimental confirmation, drug approval or clinical benefit.

  1. Faster hypothesis-to-independent-replication times across fields without weaker standards.
  2. Shorter clinical stages with safety and efficacy maintained.
  3. Repeated AI studies that transfer from task metrics to validated discoveries.
  4. Broader access to capacity, data and software—not only concentration.
Hypothesis

AI may compress search and modelling while moving bottlenecks to validation.

Possibility

Open data and software may compound reuse and verification.

Data gap

No comparable global series measures idea-to-reliable-adoption time across fields.

Method & source ledger

Every claim keeps
its scope and role.

Official API observations refresh deterministically. Papers, benchmarks and normative sources stay manually curated. Script fails closed when indicator identities, countries, pages or observation counts change.

World Development Indicators: R&D expenditure

World Bank / UNESCO UIS · official statistical series · 2000–latest

Open source ↗

World Development Indicators: scientific journal articles

World Bank / NSF NCSES · official statistical series · 2000–latest

Open source ↗

2026 AI Index, Figure 1.6.1

Stanford HAI · OpenAlex + CSO Classifier v3.3 · 2013–2024

Open report ↗

Estimating the reproducibility of psychological science

Open Science Collaboration · peer-reviewed primary study · 2015

Open source ↗

Highly accurate protein structure prediction with AlphaFold

DeepMind / CASP collaborators · peer-reviewed benchmark · 2021

Open source ↗

Recommendation on Open Science

UNESCO · official normative framework · 2021

Open source ↗

The Drug Development Process

U.S. FDA · official regulatory guidance

Open source ↗