AI Is Making Scientists More Productive—but Narrowing the Research Frontier

Artificial intelligence is helping individual researchers work faster and publish more. But across science as a whole, AI may be driving researchers toward a narrower range of ideas.

Academics who use AI tools publish approximately three times as many papers and receive nearly five times as many citations as other academics. However, an analysis of the scientific literature found that AI-based research has 4.6% less topical coverage than non-AI research. This pattern appears in more than 70% of the subfields surveyed.¹

The problem is not technical, but institutional. AI tools expand scientists’ ability to explore, yet systemic incentives encourage researchers to focus on problems that institutions can easily recognize, evaluate and reward instead of breaking new ground.

These dynamics did not start with AI. Research shows that papers and patents have become less disruptive over the past 60 years.² But AI is accelerating the trend. Speed and large-scale pattern recognition are precisely the capabilities that intensify existing research priorities. Academic career structures reinforce the effect: AI makes it faster and cheaper to expand familiar lines of research than to enter unfamiliar territory, while existing incentive systems reward that speed.

To address these concerns, we propose three recommendations.

1. Make new scientific terrain measurable

Current funding systems overwhelmingly reward the downstream exploitation of existing data rather than the upstream creation of new datasets and measurement capabilities. By applying increasingly powerful AI models to publicly available datasets, researchers can often produce publishable results at relatively low cost.

For example, Google’s Graph Networks for Materials Exploration (GNoME) is a deep-learning tool that identified 381,000 stable inorganic crystal candidates, expanding the known materials landscape by an order of magnitude.³ Meanwhile, AlphaFold, a protein-structure prediction system created by DeepMind in London, has generated more than 214 million potential protein structures, allowing biological interactions to be examined at an unprecedented scale.⁴

By contrast, building a longitudinal cohort study or establishing a biodiversity-monitoring programme can require years of ongoing investment before producing publishable findings.

This asymmetry is growing. Although the cost of making predictions with AI models has fallen by about 100 times over the past two years, building new observation infrastructure still involves high long-term operating costs.^5–7 As a result, the gap between what can be developed cheaply and what requires substantial investment widens every few months.

Funders should intentionally subsidize data infrastructure, particularly in neglected areas such as diseases excluded from major cohort studies. These investments are time-consuming and unremarkable, but they are essential if AI is to leverage observations from previously overlooked areas.

2. Stop penalizing researchers who change fields

Universities and funding agencies must stop penalizing researchers who use AI to enter new fields. AI tools reduce the information cost of changing direction. For example, ecologists entering genomics can now explore unfamiliar literature more quickly. Yet evaluation systems still penalize researchers who move between subfields.⁸

Hiring committees often evaluate candidates based on a continuous publication record in a single domain. Funding agencies also frequently treat preliminary data from an applicant’s previous work as a prerequisite for support. These practices make it harder for researchers to use AI to pursue genuinely new areas of science.