More papers, fewer breakthroughs: Can AI reshape scientific discovery?
A deepening divergence
Scientific progress unfolds at two distinct levels: general scientific discovery and disruptive innovation. The former refers to new knowledge, data, or patterns generated within existing paradigms—incremental or developmental science that expands knowledge but remains confined within established theoretical frameworks. The latter encompasses research breakthroughs that not only generate new knowledge but fundamentally challenge, rewrite, or overturn existing theories, creating entirely new research paradigms and problem spaces. The impact and significance of disruptive innovation far exceed those of general discoveries; they mark scientific revolutions and represent leaps in human understanding.
Yet a troubling trend has emerged. Research by Park et al., published in Nature, revealed an alarming pattern: over the past half-century, the disruptiveness of scientific papers and patents has systematically declined.1 By constructing a “disruptiveness index” (CD index), they found that an increasing proportion of research consolidates and develops within existing knowledge frameworks rather than proposing revolutionary theories or establishing new fields. This finding has sparked widespread concern about whether science is entering an era of “incremental innovation.”
In recent years, as this decline has continued, artificial intelligence (AI) technologies have swept across the scientific landscape, rapidly becoming an indispensable research tool. From literature reviews and code development to data analysis and manuscript editing, AI applications pervade nearly every research activity, significantly enhancing scientific productivity.2 Nevertheless, enhanced efficiency does not necessarily translate into higher-quality or more disruptive innovation. AI’s contribution to science must be evaluated across two dimensions: for general scientific discovery, AI has demonstrably accelerated progress—enabling faster data processing, pattern recognition, and hypothesis testing, allowing researchers to produce new scientific knowledge at unprecedented speed. For disruptive innovation, however, the picture is more nuanced and concerning. This tension suggests that efficiency gains do not automatically lead to better science; in fact, prioritizing speed over depth—a tendency often amplified by powerful tools—could paradoxically suppress disruptive thinking.3 This raises our core research question: as scientific disruption declines, is AI part of the problem or part of the solution? Specifically, can AI transcend its role as an efficiency tool to become a driver of transformative breakthroughs?
Data insights
Our quantitative analysis, based on a large-scale academic paper dataset drawn from the Scopus database, reveals a complex and illuminating picture. Among various AI technologies applied in scientific research, machine learning and deep learning have emerged as the most widely used and frequently cited technologies (Figure 1A). Meanwhile, the overall integration of AI into scientific research has expanded substantially, with publication volume rising sharply across almost all disciplines (Figure 1B). While this expansion in publication volume illustrates AI's immense power as an efficiency tool, it also raises a profound concern: when publishing becomes so effortless, will scientists remain motivated to pursue the high-risk, long-term research that yields truly disruptive breakthroughs?
