AI Co-Scientists Are Changing How Scientific Research Is Done

Summary

AI systems built from multiple autonomous agents are beginning to take on parts of scientific discovery. Google’s Co-Scientist can search and synthesize research papers, compare competing explanations, critique hypotheses and pursue several lines of reasoning before proposing an approach. In one example, biochemist Anna Pertl and colleagues asked it to find practical ways to target the cancer-associated protein MYC through molecular condensates. After nearly an hour of clarification, the system reviewed more than 700 papers, generated 108 strategies and rejected all but one. Its remaining proposal reversed the laboratory’s initial direction: instead of dissolving the condensates, it suggested using click chemistry to glue activating and repressing proteins together, potentially preventing MYC DNA from being read. The idea is conceptually compelling but remains an uncharted research direction, not a demonstrated treatment. Other reported uses of Co-Scientist include identifying a drug combination that kills leukaemia cells in a dish and a laboratory treatment that regenerates damaged liver tissue. In a separate Imperial College London test, researchers withheld their unpublished findings; Co-Scientist independently reached a theory that closely matched years of work on how bacteria transfer DNA between distant hosts, doing so in about two days. Similar multi-agent platforms from FutureHouse, Anthropic, OpenAI, Huawei and Sakana AI can divide questions into parallel tasks and connect perspectives across fields. However, the systems can misunderstand terminology, data labels or the intended target, as happened in the MYC and molecular-contact examples. Researchers therefore still have to define worthwhile questions, correct faulty assumptions and verify outputs against experiments and independent data. The article argues that AI may move scientists toward higher-value decisions, while creating a training dilemma: deeper expertise is needed to supervise AI even as students may get fewer opportunities to learn through hands-on research.