AI may narrow scientific exploration, authors argue

The authors highlight that AI has drastically reduced the cost of generating predictions from existing data. For example, Google’s GNoME identified 381,000 candidate stable inorganic crystals, and DeepMind’s AlphaFold generated over 214 million protein structures. In contrast, building new observational infrastructure, such as longitudinal cohort studies, remains expensive and slow. The cost of making predictions with AI has fallen roughly 100-fold in the past two years, widening the gap between cheap exploitation of existing data and expensive exploration of new data sources.

To address this, the authors recommend that funders subsidize new data infrastructure, particularly in neglected domains. They also urge universities and funding agencies to stop penalizing researchers who pivot into new fields, as AI tools lower the informational cost of entering unfamiliar terrain. Without these changes, the authors warn that AI will accelerate the trend toward less disruptive, more narrowly focused scientific work.