AI widens scientific productivity but narrows the scope of discovery, study finds

Artificial intelligence is creating a paradox in modern science: individual researchers are publishing and citing work at higher rates than ever, yet the collective scientific enterprise is exploring a increasingly narrow range of topics. According to a recent analysis, academics who integrate AI tools produce roughly three times as many papers and receive nearly five times the citations of their peers. However, the same research indicates that AI-assisted studies cover 4.6% less topical ground than non-AI work, a trend observed across more than 70% of the subfields examined.

The authors argue that this issue stems not from technological limitations, but from institutional structures. While AI expands scientists’ capacity to investigate complex questions, systemic rewards encourage researchers to concentrate on problems that institutions can easily recognize, evaluate, and fund, rather than venturing into uncharted territory. This narrowing effect is an acceleration of a long-standing trend; research has shown that disruptive papers and patents have declined over the past six decades. AI amplifies this by prioritizing speed and pattern recognition, qualities that align with existing, risk-averse academic reward systems.

To reverse this trend, the authors propose three key reforms for funders, journals, and universities. First, they urge a shift in how scientific terrain is valued. Current funding models heavily favor the downstream exploitation of existing datasets using powerful AI models, which can yield publishable results at low cost. Examples such as Google’s GNoME, which identified 381,000 stable inorganic crystals, and DeepMind’s AlphaFold, which mapped over 214 million protein structures, demonstrate the efficiency of this approach. In contrast, building new observational infrastructure, such as longitudinal cohort studies or biodiversity programs, requires years of investment before yielding results. With the cost of AI predictions falling approximately 100-fold in two years, the gap between cheap data exploitation and expensive data creation is widening. The authors call for funders to subsidize data infrastructure, particularly in neglected areas like underrepresented diseases.

Second, the article suggests that institutions must stop penalizing researchers who pivot to new fields using AI. Tools like AlphaFold lower the informational barriers for scientists entering unfamiliar domains, such as an ecologist moving into genomics. Despite this, hiring committees and funding agencies often require a continuous publication record in a single domain or preliminary data from previous work, effectively discouraging interdisciplinary shifts.

By reforming these incentive structures, the scientific community can ensure that AI serves as a tool for broadening discovery rather than reinforcing existing paradigms.