Gemini for Science: Google's AI tools move into research labs

Google has launched Gemini for Science, a suite of experimental AI tools that pushes its Gemini platform beyond answering questions and into the scientific method itself. Announced at Google I/O 2026, the package spans hypothesis generation, computational discovery, and literature synthesis — three steps of research that, until now, have been stubbornly manual. The company is betting that AI’s next value unlock is not better chat but faster science.

The launch, backed by two Nature papers published on 19 May, covers three distinct tools. Co-Scientist, built on Gemini 3.5, runs an “idea tournament” that generates novel hypotheses from a researcher’s prompt, ranks them by plausibility, and surfaces the most promising for wet-lab testing. AlphaEvolve, a successor to the AlphaFold lineage, automates the design and optimisation of protein binders and enzymes for therapeutic and industrial applications. And Literature Insights, powered by NotebookLM, synthesises findings across thousands of papers into a structured evidence map with citation-verified outputs.

AI can help eliminate this bottleneck and serve as a force multiplier for scientific work by handling complex tasks. This allows researchers to focus on identifying and tackling the most impactful scientific problems.
— Pushmeet Kohli & Yossi Matias, Google Research

Agents, not chatbots

The more consequential shift here is architectural. These tools do not run as chat interfaces. They run on Antigravity 2.0, Google’s agentic development platform, which also underpins Gemini 3.5 Flash and the company’s broader coding-agent push. TechCrunch described the I/O announcements as Google betting its next AI wave on agents rather than chatbots — and Gemini for Science is the most specialised expression of that bet yet.

Science Skills — a curated bundle of 30-plus life-science databases including UniProt, the AlphaFold database, and InterPro — plugs directly into Antigravity’s tool-calling layer. A researcher’s natural-language query triggers a chain of database lookups, not a probabilistic text completion. The Google Research team ran a genomic analysis of the AK2 gene through this pipeline and compressed what normally takes hours of complex work into minutes, surfacing a novel disease-mechanism insight in the process.

The architecture matters because it changes the failure mode. A hallucinated fact in a chatbot is noise a reader might miss; a hallucinated protein interaction in a drug-discovery pipeline is a wasted six-month wet-lab run.

Enterprise first, researchers second

Google is routing enterprise access through Google Cloud, and the early adopter list tells its own story. BASF, Daiichi Sankyo, Klarna, and several US National Laboratories are already in private preview. More than 100 institutions are collaborating on validation, including Stanford, Imperial College London, and the Francis Crick Institute. A Labs waitlist form exists for individual academic researchers, but the Cloud distribution path makes the pricing model the critical variable for Australian buyers — a national research team cannot sign a US-dollar enterprise licence on a lab-by-lab discretionary budget.

I see AI as an amplifier of human ingenuity because it empowers the research scientists to ask bigger questions, pursue bigger impact, and to do it at a much earlier stage in their scientific career.
— Yossi Matias, Business Insider

Matias, who oversees Google’s Co-Scientist and ERA programs, told Business Insider the vision is to give researchers “a polymath in your pocket.” But the go-to-market — enterprise Cloud deals, not consumer subscriptions — tells a different story about who Google expects to pay for it. Corporate R&D divisions with existing Google Cloud commitments are the natural first cohort. University research offices, with their grant-cycle budgeting and institutional procurement, are a harder sell.

The trust gap

Victoria Song, writing in The Verge, flagged the gap that matters most to working researchers: the distance between “solve all diseases” keynote framing and the 20-plus-year timeline of real drug development. Demis Hassabis, Google DeepMind’s CEO, closed his I/O keynote by describing the current moment as “the foothills of the singularity.” It is a phrase designed to inspire. It also asks a lot from an audience of bench scientists who measure progress in years per target, not AGI timelines.

It takes time and effort to incorporate AI-generated findings into healthcare systems that are appropriately cautious.
— Yossi Matias, Business Insider

The critique is not that the tools are useless. Co-Scientist identified drug repurposing candidates for acute myeloid leukaemia that the Nature paper validated independently. The AK2 gene analysis delivered a novel mechanism insight that had not been surfaced through conventional methods. These are real results. The risk is that the messaging gap — “foothills of the singularity” versus a Nature paper that is one step in a decades-long pipeline — trains researcher scepticism at exactly the moment Google needs trust to convert pilot users into paying customers.

Singapore is moving. Is Australia?

Singapore has already committed. Under its US$1 billion-plus national AI strategy, the city-state signed a deal with Google to train government researchers on agentic AI tools for science, creating a regional adoption model that Australian institutions are not yet matching. CSIRO, which shed 500 jobs in its 2024 restructure, has no publicly announced Gemini for Science partnership. The contrast is sharp: Singapore is training its research workforce on the tools; Australia’s peak science agency is still navigating budget constraints that make enterprise Cloud licences hard to justify on a line-item basis.

For Australian research buyers — university pro-vice-chancellors of research, CSIRO division chiefs, enterprise R&D heads inside the Atlassians and Cochlears — the decision tree is straightforward. Gemini for Science is the most structured attempt yet to put AI inside the scientific workflow rather than alongside it. It is also a Google Cloud product with an undisclosed price tag, running on a platform whose enterprise data-sovereignty story for sensitive genomic and health data has not been publicly tested in the Australian jurisdiction. The tools are real; the cost, the integration friction, and the validation gap between a Nature paper and a reproducible institutional finding are also real.

The watchword is not “wait and see” — Singapore is not waiting. It is “pilot small, verify independently, budget for the Cloud bill.”