While Artificial Intelligence (AI) is frequently hailed as a revolutionary force in modern science, quantifying its actual impact across diverse disciplines has remained a challenge. By analyzing a massive dataset of 26 million research articles published between 2000 and 2019, researchers have provided a data-driven perspective on AI's penetration into the scientific record. The study reveals that while AI is a rapidly growing field, it currently accounts for approximately 1.6% of total scientific output, characterized by a unique "double-boom" growth cycle.
Methodology: Research on Research
To move beyond anecdotal evidence, the researchers employed a large-scale bibliometric approach. They utilized the Papers with Code platform to identify a taxonomic backbone of 2,060 specific AI methods. This list was rigorously filtered to remove ambiguous terms and generic statistical methods (such as standard linear regression) unless they were explicitly linked to AI contexts.
Using Python-based keyword matching, the team scanned the abstracts of 26,408,350 articles from the Web of Science (WoS) database. The process involved three rounds of searching and a final manual verification step to ensure the identified articles were genuinely discussing the application of AI techniques rather than merely mentioning them in passing.
The "Double-Boom" Growth Cycle
The study maps the trajectory of AI adoption onto Ulrich Schmoch’s "double-boom" theory. The first boom, occurring around 2005, was largely driven by theoretical potential and high initial expectations. This was followed by a plateau as researchers encountered practical implementation hurdles.
The "second boom," which began post-2017, represents the current era of explosive, application-driven growth. During this period, technical hurdles began to clear, leading to widespread practical application across various domains. By 2019, the growth rate of AI-related articles reached a staggering 30% annual increase.
The Landscape of AI Methods
Out of the 26 million articles analyzed, 435,994 were identified as utilizing AI methods. A key finding is the extreme concentration of techniques; while 449 different AI methods were identified in the literature, the vast majority of papers rely on a small handful of "classics."
Three methods—Bayesian analysis, neural networks, and Principal Component Analysis (PCA)—dominate the landscape. These "Big Three" are favored because they possess well-established theoretical foundations, making them "safer" and more accessible for researchers in non-computer science fields to adopt. Conversely, the study found a "long tail" of methods, where the majority of identified AI terms appeared fewer than 10 times across the entire 20-year corpus.
Conclusion and Future Outlook
Despite the significant growth and the "second boom" of application, AI remains a niche powerhouse, representing only 1.6% of the total scientific record by 2019. This suggests that while AI is transformative, it has not yet become a universal standard in most scientific workflows.
It is important to note that this study's timeframe ends in 2019, prior to the massive explosion of Generative AI and Large Language Models (LLMs). The "second boom" identified by the authors has likely accelerated significantly in the years since. Future research will focus on how specific disciplines differ in their adoption rates and how AI techniques diffuse from computer science into other specialized fields like biology and physics.
Summary updated