Nature papers on AI in science

Nature and Google this week put new detail around a fast-moving corner of AI research: systems built not just to answer questions, but to help run parts of the scientific process. Two Nature papers published on May 19 described multi-agent systems aimed at hypothesis generation, experiment planning and data analysis, while Google used its I/O announcements to package related prototypes under a “Gemini for Science” banner. (nature.com) Nature’s own news coverage on May 19 said teams of AI agents are being built to generate hypotheses, interpret data and suggest ways to develop medicines. (blog.google) A separate May 18 arXiv paper, “AI for Auto-Research: Roadmap & User Guide,” argued that the technology is improving quickly but remains more reliable in structured, tool-mediated tasks than in fully autonomous scientific judgment. (nature.com) ### What exactly did the new Nature papers report? Nature on May 19 published “Accelerating scientific discovery with Co-Scientist,” which introduced Co-Scientist as “a multi-agent AI system built on Gemini for structured scientific thinking and hypothesis generation.” The paper said the system is meant to augment scientists by helping generate and refine testable ideas. (nature.com) Nature the same day also published “A multi-agent system for automating scientific discovery,” which introduced Robin as “the first multi-agent system capable of fully automating both hypothesis generation and data analysis for experimental biology.” That paper placed the work closer to an end-to-end lab workflow than a literature assistant or chat interface. (nature.com) ### How does Google describe Co-Scientist and the related tools? Google’s May 19 post on “Gemini for Science” said the company was introducing “a collection of science tools and experiments” designed to expand “the scale and precision of scientific exploration.” The post said three prototypes were being shown on Google Labs: Hypothesis Generation, Computational Discovery and Literature Insights. (nature.com) Google said Hypothesis Generation is built with Co-Scientist and uses a multi-agent “idea tournament” to generate, debate and evaluate hypotheses, with claims backed by clickable citations. The company said Computational Discovery is built with AlphaEvolve and ERA, short for Empirical Research Assistance, and is designed to generate and score thousands of code variations in parallel for modeling tasks in areas including solar forecasting and epidemiology. (nature.com) ### Where does the “autonomous research workflow” claim come from? Nature’s news article framed the broader development as “teams of AI agents” that can speed research by dividing work across specialized systems. (blog.google) That report said such systems can generate hypotheses, interpret data and suggest paths for medicine development. The May 18 arXiv roadmap used even broader language, saying long-horizon agents can execute experiments, draft manuscripts and simulate critique with minimal human input. But the same paper said end-to-end autonomous systems had not “consistently reached major-venue acceptance standards” and argued for “human-governed collaboration” as the most credible deployment model. (blog.google) ### What are researchers and journals warning about? Nature Machine Intelligence published a commentary in February saying multi-agent AI systems need transparency. The article said these platforms can take an open-ended research question and run a full cycle from literature search and data analysis to hypothesis generation and paper writing, but argued that their operation must be inspectable. (nature.com) Nature also published a March news article on end-to-end automation of AI research, reporting that an AI system could produce papers with minimal human involvement and pass a first round of peer review at a workshop of a major machine learning conference. That article added to the case that capability is moving faster than norms for validation. (arxiv.org) ### What should readers watch next? May 19 is now the key date in this cycle: that is when Nature published the Co-Scientist and Robin papers and when Google published its Gemini for Science overview. The next concrete checkpoints are likely to come from follow-up evaluations, external replication and any additional technical releases tied to Google’s Co-Scientist, ERA and AlphaEvolve prototypes. (nature.com 1) (nature.com 2) (nature.com 3)