How AI Can Advance Science if Institutions Shift Away from Rewarding the Status Quo

The challenges associated with utilizing artificial intelligence in science are becoming more prevalent. While individual researchers are experiencing increased productivity, the scientific community as a whole is delving into a more restricted range of ideas.

For instance, researchers utilizing AI tools are publishing about three times as many papers and receiving close to five times as many citations compared to their counterparts. Despite this, an examination of scientific literature reveals that AI-based research covers 4.6% fewer topics than non-AI research, a trend observed in over 70% of the subfields analyzed.¹.

The issue does not lie in the technical aspect but rather in the institutional framework. While AI tools enhance scientists’ exploratory abilities, the existing incentive structures lead researchers to concentrate more heavily on problems that institutions easily recognize, evaluate, and reward, rather than venturing into uncharted territories.

These developments did not originate with AI. Studies indicate that both papers and patents have showcased reduced disruptive tendencies over the past six decades.². However, AI is hastening these transformations. The emphasis on speed and large-scale pattern identification narrows the focus of existing systems. The academic career progression further reinforces this pattern. AI facilitates the swift and cost-effective expansion of familiar research areas, while current incentive mechanisms reward this rapid pace.

To address these concerns, we have put forth three key recommendations.

Enhancing Research Diversification

The prevailing funding systems predominantly promote the downstream utilization of existing data rather than the upstream generation of novel datasets and measurement capabilities. By employing advanced AI models on publicly available datasets, researchers can often achieve publishable outcomes at a relatively low cost.

For example, Google’s Graph Networks for Materials Exploration (GNoME) is a deep learning tool that has identified 381,000 stable inorganic crystal candidates, expanding the known materials landscape by a significant margin.³. Similarly, AlphaFold, a protein structure prediction system developed by DeepMind in London, has generated over 214 million potential protein structures, enabling a profound examination of biological interactions on an unprecedented scale.⁴.

In contrast, establishing a longitudinal cohort study or initiating a biodiversity monitoring program can entail substantial ongoing investments over extended periods before yielding publishable findings.

This growing asymmetry is noteworthy. While the cost of AI model predictions has decreased approximately 100-fold in the past couple of years, the establishment of new observational infrastructure dictates high long-term operational expenses.^5–7. Consequently, the disparity between low-cost and high-cost development options is widening with each passing month.

We advocate for deliberate funding support toward data infrastructure, especially in overlooked areas like diseases excluded from major cohort studies. Although time-consuming and perhaps unglamorous, these investments are crucial for enabling AI to leverage observations in previously unexplored domains.

Embracing Research Diversification

Universities and funding organizations must cease penalizing researchers who utilize AI to venture into unfamiliar fields. AI tools mitigate the informational barriers associated with domain transitions. For instance, ecologists delving into genomics can now navigate unknown research literature more efficiently. However, the current evaluation system still penalizes shifts to alternative subfields⁸. Hiring panels tend to assess candidates based on a continual publication track record within a single realm, and funding bodies often require preliminary data from an applicant’s previous work as a prerequisite for support.

The Balancing Act of AI in Scientific Exploration