AI for Science Needs Reasoning, Not Just Data

AlphaFold, developed by Google DeepMind, solved protein structure prediction after half a century of effort, but the conditions enabling its success—the massive Protein Data Bank and reliable experimental data—are rare and costly to replicate in most fields. Instead, AI agents, powered by large language models, are emerging as a more general tool for science. Agents like Google's AI Co-Scientist can reason, plan, and test hypotheses; one such agent correctly hypothesized how antibiotic resistance spreads, a conclusion that took Imperial College researchers a decade to reach. These agents can log every action, addressing the reproducibility crisis, preserve institutional knowledge, and dramatically speed up research by reducing the cost of experimentation. While challenges like hallucination and memory limits remain, agents represent a fundamental shift akin to calculus or the computer, transforming science across all fields.