Artificial intelligence adoption in the physical sciences, natural sciences, life sciences, social sciences and the arts and humanities: A bibliometric analysis of research publications from 1960-2021

• AI has seen historical ups and downs, but scholarly publishing and activity around AI have surged in recent years.Recent increases also include investment, policy initiatives, and university AI courses.Source: §I · p. 1
• Research is among the earliest and most enthusiastic adopters of AI, which may improve research cost-effectiveness, speed, safety and quality across fields.Some researchers argue AI could reshape discovery and the organisation of science, amid concerns about declining research productivity.Source: §I · p. 1
• The study examines AI adoption across physical, natural, life and social sciences, and arts and humanities, over 1960–2021 using a broad definition spanning 214 phrases.Source: §I · p. 1
• The paper examines AI's development, application and diffusion across research fields, then discusses knowledge discovery, boom-bust cycles and productivity.It concludes that AI's future impact may range from useful to paradigm-changing for researchers in most disciplines.Source: §I · pp. 1–2

• Earlier studies reported growing AI publication activity in neural networks and specialized disciplines, while other work examined AI references and the multidisciplinary development of AI.Examples include neural-network diffusion across fields, AI research references, and publishing patterns across disciplines.Source: §II · p. 2
• The study extends prior work by examining AI development and application across practically all research fields, focusing on differences and complementarities between fields.Source: §II · p. 2
• The 1960–2021 analysis covers a longer period than earlier studies, including the first two AI springs and winters and early diffusion beyond computer science in the 1980s.Source: §II · p. 2
• The study uses 214 OECD-derived AI search phrases, compared with 30 phrases focused on neural networks in Bianchini et al. and a narrower strategy in Liu et al..Source: §II · p. 2
• Using the ASJC classification, the analysis covers all research fields at a more granular level and applies the taxonomy consistently across the period through The Lens.Source: §II · p. 2

• Bibliometric analysis is used to examine patterns associated with AI adoption over the past 62 years.Source: §III · p. 2
• The approach uses publication and citation data to provide quantitative insights into structural relationships in fields of interest.Source: §III · p. 2
• The methods describe data sources, a research-field taxonomy, an AI search strategy and reporting metrics, following the procedure outlined by Donthu et al..Source: §III · p. 2

• The Lens contains over 224 million scholarly publications and over 137 million intellectual-property patents, with records dating back to the 1950s.Source: A. The Lens Database of Scholarly… · p. 2
• The study extracted records published through 31 December 2021 and accessed The Lens through its GUI and API, using Python scripts for customised searches.Source: A. The Lens Database of Scholarly… · p. 3
• The Lens draws publication data from Microsoft Academic Graph, Crossref, PubMed, Impactstory and CORE.Source: A. The Lens Database of Scholarly… · p. 3
• The authors chose The Lens for its open-access, non-commercial model, comprehensive multi-database coverage, provenance transparency, API and detailed metadata.They note it had not yet featured in comparative analyses with mainstream databases.Source: A. The Lens Database of Scholarly… · p. 3

B. The All-Science Journal Classification (ASJC)

• The ASJC taxonomy contains 333 unique research fields at its most detailed level, grouped under 26 subject fields and four broad areas.The broad areas are physical sciences, life sciences, health sciences, and social sciences and the arts and humanities.Source: B. The All-Science Journal… · p. 3
• The Lens assigns publications to one or more detailed ASJC fields using ISSN descriptions in Crossref metadata.Source: B. The All-Science Journal… · p. 3
• The authors consider ASJC suitable for this analysis while noting that other classification systems are possible.Source: B. The All-Science Journal… · p. 3

C. Defining Artificial Intelligence and Identifying Publications for Bibliometric Analysis

• AI was operationalized using 214 OECD phrases developed through bibliometric analysis, text mining, and validation by AI experts.Source: C. Defining Artificial Intelligence… · p. 3
• The authors chose the OECD phrase list because it includes more granular AI subfields than the Liu et al. list and offers a more comprehensive definition.Of the 214 OECD phrases, 113 had no Liu et al. match, 76 had possible matches, and 25 had exact matches.Source: C. Defining Artificial Intelligence… · p. 3
• Publications qualified as AI-related when one or more OECD phrases appeared in the title, abstract, or keywords.The analysis covered peer-reviewed books, book chapters, journal articles, and conference papers or proceedings.Source: C. Defining Artificial Intelligence… · p. 3
• 3,126,436 records formed the final analysis pool after 87 million records were excluded by document type and 137 million records were screened for AI-related terms.The initial Lens dataset contained 224 million scholarly publication records from 1960–2021.Source: C. Defining Artificial Intelligence… · pp. 3–4

IV. RESULTS

• 14% of 333 fields had AI-related publications in 1960; coverage exceeded half by 1972, 80% by 1986, and 98% today.AI began in the physical sciences before spreading into life sciences, social sciences, arts, and humanities.Source: §IV · p. 4
• The Gini coefficient fell from 0.91 in 1960 to 0.72 by 1980, then remained between 0.71 and 0.76 as computer science maintained a high concentration of AI-related output.The authors attribute the limited further decline to computer science’s faster increase in AI publishing intensity and volume than any other field.Source: §IV · p. 4
• 5.3% of total research output was AI-related in 2017–2021, after rising from 0.02% in 1960 and remaining below 1% until 1995.Over 50% of all AI research volume was published in the past five years; annual AI publishing growth averaged 26% in that period versus 17% earlier.Source: §IV · p. 4, Figure 4 · p. 4
• Physical sciences were the largest overall adopter and developer of AI research, while practically all fields showed substantial increases in the past 3–5 years.From 2020 to 2021, AI publishing intensity rose 1.9 times in dentistry, 1.4 times in arts and humanities, and 1.3 times in economics, econometrics and finance, health professions, and social sciences.Source: §IV · pp. 4–5, Figure 5 · p. 5
• Arts-and-humanities AI publishing intensity rose only recently, while geography is among the social-science fields driving recent increases.AI represented a quarter of total research output in computer science.Source: §IV · p. 5

V. DISCUSSION

• AI adoption has risen sharply in recent years across specialized fields, consistent with the study’s broader cross-field pattern.Prior studies covered areas including human resources, engineering, healthcare, supply chains, renewable energy, and education.Source: §V · p. 5
• Earlier AI springs were followed by winters associated with hype, inflated expectations, and funding reductions when AI failed to deliver perceived promises.The authors describe two spring-winter cycles, while noting that some researchers see conditions for another winter today.Source: §V · p. 5
• The current AI surge exceeds earlier surges in depth and breadth and coincides with hardware, software, and cloud advances that make historical barriers less significant.The authors therefore suggest another AI winter is not imminent and may not occur in the same form as earlier winters.Source: §V · pp. 5–6
• AI could raise research productivity, but gains are not yet well demonstrated in science, research, innovation, and technology.The authors call for more research on productivity effects in these sectors.Source: §V · p. 6
• AI successes are more visible than failures, and an evaluation found none of 62 published COVID-19 diagnosis and prognosis models suitable for clinical use because of flaws or bias.Reviews also reported prevalent bias in COVID-19 diagnostic applications, limiting their clinical potential.Source: §V · p. 6
• AI may either help researchers do existing work faster, safer, cheaper, and better or fundamentally reshape knowledge discovery.The latter possibility is illustrated by proposals for highly autonomous AI capable of top-level science.Source: §V · pp. 6–7

VI. CONCLUSION

• AI development and application for research are surging worldwide across the physical, natural, life, and social sciences and the arts and humanities.The authors say the current surge exceeds earlier episodes in magnitude and breadth, though some component may reflect hype or trend-following.Source: §VI · p. 7
• AI is likely to keep improving research speed, cost-efficiency, safety, and productivity, and might fundamentally change scientific methods and knowledge discovery.Source: §VI · p. 7
• The study’s implication for researchers and research organisations is to invest in the many dimensions of AI capability uplift.Source: §VI · p. 7