Google launches Gemini for Science discovery tools

At Google I/O (May 19, 2026), Google introduced Gemini for Science, an experimental suite of AI tools that aim to support stages of the scientific method, according to Google's I/O keynote and the company blog. The public-facing prototypes are Hypothesis Generation, Computational Discovery, and Literature Insights, each built around Gemini-class multi-agent reasoning and linked systems described on Google's research pages. Google Research and product posts identify Co-Scientist, Alpha Evolve, Empirical Research Assistance (ERA), and NotebookLM as components or underlying projects used to assemble the prototypes, per the Google and DeepMind blogs and the Research.Google post.

Per the Research.Google post, ERA is documented in a publication in Nature and is presented as an expert-level scientific coding assistant that can search literature, generate and optimize code, and evaluate computational experiments against defined metrics. The Google and DeepMind posts include applied examples: DeepMind's blog highlights that Co-Scientist helped identify a drug-repurposing candidate that, in reported lab tests, blocked 91% of a scarring-linked response, and documents collaborative use cases across liver fibrosis, ALS research, and cellular aging.

Google frames access to the prototypes as a gradual rollout: the company is making the tools available through Google Labs with a separate enterprise path via Google Cloud, per the Google blog and Digital Trends reporting. Digital Trends also reports that a feature called Science Skills pulls insights from more than 30 major life-science databases and research tools.

The announced components combine several technical patterns visible across recent AI-for-science work: multi-agent idea generation, automated code synthesis and optimization, and literature-grounded retrieval with citation linking. Systems like ERA that use tree search and large-model guidance to propose and evaluate code variants mirror patterns in automated machine learning and program synthesis research. Industry practitioners will recognise the same core building blocks used elsewhere: retrieval-augmented grounding for claims, agent orchestration for iterative exploration, and automated experiment search to scale combinatorial testing.

Companies and academic projects pursuing AI-assisted discovery have increasingly combined large language or generalist models with domain toolchains and experiment automation. For practitioners, the Google suite both reflects this trend and consolidates it in a single branded workbench that stitches model reasoning, domain connectors, and computational engines. Reports of published validation (the ERA Nature paper) and real-lab case studies (Co-Scientist examples in DeepMind's blog) strengthen the narrative that such systems are moving from lab demos toward reproducible, documented workflows.

For practitioners and teams evaluating these tools, relevant indicators include the scope and format of the Google Labs trusted-tester program, API and data-connectivity limits for Science Skills, reproducibility artifacts tied to the ERA Nature paper, and metadata quality for the clickable citations in literature outputs. Observers should also watch for independent reproductions of the lab results showcased in DeepMind's writeups and for published benchmarks comparing ERA-style automated coding to expert human performance across domains.

Industry teams building research tooling should treat this announcement as a consolidation of emerging patterns rather than a single breakthrough; the offering bundles multi-agent orchestration, program synthesis, and retrieval-grounded outputs into a cohesive user path. Adoption decisions will hinge on access, provenance guarantees for literature and code, and the ability to reproduce published experiment results.