Choosing an AI approach for scientific research is less about picking the most powerful system and more about matching a tool to a stage of the scientific process. Large language models excel at synthesizing existing knowledge and drafting hypotheses; generative design models create novel molecules, proteins, and materials; and multi-agent systems act as virtual collaborators that generate, rank, and refine hypotheses through iterative debate. Each carries distinct limitations, and none removes the scientist from the loop.

Introduction: Navigating AI in Scientific Discovery

AI is now woven through the scientific process, from data collection to interpretation. A Nature review by Wang and colleagues examines a decade of breakthroughs, including self-supervised learning on vast amounts of unlabelled data and geometric deep learning that exploits the structure of scientific data to improve accuracy and efficiency. The same review highlights generative methods that create designs such as small-molecule drugs and proteins by analyzing diverse data modalities, including images and sequences.

What often gets lost in the enthusiasm is that these are different tools doing different jobs. A model that drafts a literature review is not the model that proposes a new compound, and neither is the system that runs an iterative hypothesis-testing loop. Understanding the boundaries between approaches is the first step toward using any of them well.