The study examined 41.3 million English papers from 1980‑2025 across biology, chemistry, physics, medicine, materials science, and geology. It identified 311,000 papers that incorporated AI techniques such as neural networks or large language models. Researchers who used AI published roughly three times as many papers, earned nearly five times as many citations, and assumed leadership roles one to two years earlier than peers who did not.

However, mapping these outputs onto a high‑dimensional knowledge space revealed a contraction in intellectual breadth. AI‑heavy studies clustered tightly around popular, data‑rich problems, produced weaker networks of follow‑on engagement, and occupied a smaller footprint than non‑AI work. The pattern held across the evolution from early machine learning to deep learning and generative AI.

The analysis also documented a rise in low‑quality and fraudulent papers produced at scale, as well as a tendency for AI to solve the most tractable, well‑defined questions—protein folding, image classification, pattern extraction. Without deliberate design, AI tools and the scientists who rely on them are unlikely to venture into data‑scarce, poorly mapped territoriesాటు

Experts argue that the issue lies more in reward structures than in AI architecture. Integrating data, computation, and hypothesis‑generation tools could broaden discovery, but only if incentives shift toward riskier, original research rather than speed and volume.