Gemini for Science is Google’s latest bet that AI should not just write emails and summarize meetings, but actually help make real scientific discoveries. Announced around Google I/O 2026, it is less a single product and more a workbench of tools, agents, and data pipelines that try to plug AI into every step of the scientific method, from brainstorming hypotheses to crunching code to decoding dense literature.

If you zoom out, the timing makes sense. Science is drowning in its own success: millions of papers a year, increasingly complex simulations, and datasets that laugh at traditional analysis tools. Google’s argument is simple but ambitious – if large models can reason, write code, and navigate information at scale, why not enlist them as “co-scientists” and build an AI layer that sits directly on top of the scientific process. Gemini for Science is that layer, stitched together from DeepMind’s research systems, Google Research’s tooling, and the company’s growing “agentic” Gemini stack.

At the heart of the announcement is a trio of experimental tools in Google Labs, each aimed at a different bottleneck in research. First is Hypothesis Generation, built on a system appropriately named Co-Scientist, a multi-agent Gemini setup that doesn’t just answer questions but runs a kind of internal “idea tournament.” A researcher starts with a problem – say, why a particular genetic mutation drives a disease – and Co-Scientist responds by breaking the challenge down, scouring the literature, proposing alternative explanations, and having its own agents argue over which hypotheses actually make sense. The system leans heavily on citations and cross-checks, which matters a lot in a field where a confident but wrong answer can derail months of lab work.

Google has already pushed Co-Scientist through the gauntlet of peer review: a Nature paper published in May 2026 details how multi-agent AI can generate and refine novel hypotheses for complex problems, particularly in life sciences. In that work, the system helped design experiments and suggest mechanisms for disease pathways, with human scientists still firmly in charge of validation. That mix – AI as an engine for ideas, humans as the filter of reality – is very much the philosophy behind Gemini for Science.

The second Labs prototype, Computational Discovery, tackles a different pain point: the brutal, often tedious loop of writing and refining scientific code. Built with AlphaEvolve and an AI system called Empirical Research Assistance (ERA), it acts like an always-on collaborator that generates, tests, and scores thousands of code variations in parallel. Give it a well-defined problem and an evaluation metric – forecasting solar output, modeling epidemics, optimizing a physical simulation – and it starts exploring the space of approaches far faster than a single human coder could.

ERA itself has its own Nature-backed credentials. Google describes it as an “AI system designed to help scientists write expert-level empirical software,” and reports expert-level performance across six diverse, challenging problems, from physics to climate modeling. That research underpins Computational Discovery: instead of manually trying one model tweak at a time, scientists can use an AI agentic engine to propose and evaluate whole families of algorithms, surfacing directions that are genuinely novel rather than just incremental. It is, in effect, automated scientific exploration of the code space.

The third Labs tool, Literature Insights, is probably the most relatable for anyone who has ever opened a PDF and felt their soul leave their body. Built on NotebookLM, Google’s long-form “AI note-taking” system, it is designed to ingest a curated corpus of scientific papers and turn them into something researchers can actually navigate. Instead of a pile of PDFs and browser tabs, you get structured tables with custom attributes, side-by-side comparisons of methods or datasets, and a chat interface that answers questions grounded strictly in the uploaded literature.

NotebookLM has already shown it can help people synthesize information and generate reports in other domains, and here it is pointed squarely at scientific workflows. Google says Literature Insights can generate artifacts that normally take multiple tools – think reports, slide decks, even audio and video overviews – and, crucially, help uncover gaps and opportunities across papers, not just summarize them one by one. In a world where staying current in a single subfield is a full-time job, that kind of synthesis is not a luxury; it is survival.

These three tools sit under the Gemini for Science umbrella, with early access coming via a waitlist at labs.google/science. That framing – as “experiments” in Google Labs rather than polished, mass-market products – is intentional. Google wants researchers to stress-test them, push on their limits, and effectively co-design how AI should behave in the lab and in the literature stack.

But Gemini for Science does not stop at individual researchers signing up for a Labs experiment. The same underlying systems are already being productized for enterprise and institutional use through Google Cloud, where AI-powered R&D is quickly becoming its own category. Google says companies like BASF are tapping AlphaEvolve to optimize complicated supply chains, using algorithmic exploration to juggle thousands of possible decisions and constraints. Payments company Klarna, meanwhile, has talked publicly about using AlphaEvolve-style algorithmic evolution to double training speed for some of its models, essentially treating ML architectures themselves as a search space for an AI to explore.

On the more traditional “science” side, major organizations such as Daiichi Sankyo, Bayer Crop Science, and US National Labs participating in the Department of Energy’s Genesis Mission are trialing Co-Scientist to accelerate fundamental research. The pitch there is less about marginal efficiency and more about surfacing unexpected hypotheses – molecular targets you wouldn’t think to test, or mechanistic explanations that pull from multiple subfields. All of this is still framed as “private preview” and “pilots,” but it is clear Google sees a sizable business in AI-augmented scientific and industrial R&D.

If the Labs tools are about how scientists think, another pillar of Gemini for Science is about the data they think with, especially in biology. Alongside the Labs announcement, Google is launching what it calls Science Skills, a specialized bundle that pipes insights from more than 30 major life science databases directly into agentic Gemini platforms like Google Antigravity. Think of it as domain-specific instruments bolted onto a general-purpose AI agent: UniProt for protein sequences, the AlphaFold database for predicted 3D structures, AlphaGenome APIs for genomic context, InterPro for protein families, and dozens more.

In practical terms, this means a researcher can stay inside an environment like Antigravity and still perform workflows that once required a patchwork of websites, command-line tools, and custom scripts. Structural bioinformatics tasks, like analyzing how a mutation might affect a protein’s folding or interactions, can be run in minutes instead of hours, according to Google’s own internal testing. In one early case, the company says its teams used Science Skills to quickly analyze a rare genetic disease driven by mutations in the AK2 gene, generating new hypotheses about the underlying mechanisms far faster than with their previous manual pipelines.

Antigravity itself is part of Google’s broader “agentic Gemini” push – an environment where AI agents can orchestrate tools, call APIs, and manage long-running workflows rather than just respond in a chat window. Plugging Science Skills into that environment essentially turns it into a kind of scientific desktop: an agent that knows both the general language of science and the specifics of the databases and tooling that define modern biology.

To its credit, Google is not pretending it can build this future in isolation. The Gemini for Science announcement leans heavily on a long list of collaborators: over 100 institutions, including Stanford University on liver fibrosis research, Imperial College London on antimicrobial resistance, and a multi-year collaboration with the Francis Crick Institute. The company has also set up a “trusted tester” community that ranges from PhD students to industry experts and Nobel laureates, tasked with stress-testing these agents on real problems and spotting both blind spots and failure modes.

Google is also experimenting with AI in the meta-layer of science: the academic workflow itself. Working with conferences like ICML, STOC, and NeurIPS, it is piloting agentic tools for peer review and paper quality, including a Paper Assistant Tool and something called ScholarPeer aimed at improving figures and review feedback. The idea of AI-assisted reviewing will almost certainly be controversial in parts of the community, but it fits the same thesis: if there is a highly structured, information-heavy task in science, Gemini is supposed to be able to help.

Of course, Gemini for Science does not appear out of thin air. It sits on top of a decade of Google-led AI work that has already rewritten parts of the scientific playbook. AlphaFold, perhaps the most famous example, opened up access to predicted 3D structures for hundreds of millions of proteins, and Google says it has been used by more than 3 million researchers to attack problems from malaria vaccines to enzymes that break down plastics. AlphaGenome extends that approach deeper into genomics, using AI to better map genetic variation to disease drivers.

Layered into that are more familiar tools that scientists touch daily: Google Scholar for literature discovery, Earth Engine and Earth AI for geospatial and climate analysis, Colab for running notebooks in the cloud, MedGemma for medical imaging, Gemini Deep Research for long-context analysis, and the new Gemini Deep Think models aimed at more rigorous, step-by-step reasoning on complex tasks. Gemini for Science essentially reorganizes all of these pieces – plus new agents like Co-Scientist, ERA, and AlphaEvolve – into a coherent story: AI as infrastructure for science, not just an add-on.

There are obvious caveats. Even Google’s own messaging emphasizes that AI systems like Co-Scientist can propose promising hypotheses and design experiments, but they cannot validate the underlying biology or physics on their own. Independent commentators have pointed out that, in early Nature-reported case studies of AI “co-scientists,” human oversight remains critical and some suggested molecules or mechanisms do not pan out on closer inspection. The risk is not just that an AI gets something wrong, but that the scientific community over-indexes on what is easy for agents to explore, potentially skewing attention away from questions that are harder to formalize.

Then there is the question of accessibility and power. Right now, many of these tools are gated – private previews, trusted testers, conferences and large enterprises, all connected through Google Cloud. For big pharma, national labs, or tech giants, that is a manageable on-ramp. For a small academic lab or a researcher in a resource-constrained setting, it might feel more like a wall. As with any infrastructure shift, who gets access first, and on what terms, could shape which problems get tackled and which are left waiting.

Still, taken on its own terms, Gemini for Science is a notable inflection point. We have had AI paper summarizers, code assistants, and data tools for years, but they have generally lived in separate silos, each solving a narrow pain point. This is one of the first major attempts by a hyperscaler to treat the entire scientific method – hypothesis, experiment, analysis, iteration, communication – as something an integrated AI stack can plug into end-to-end.

If Google and the wider community can get the guardrails, validation, and access story right, the upside is hard to ignore: faster idea generation, more systematic exploration of models and code, literature that feels less like a black hole and more like a map. The company is explicit about the endgame – using AI agents not just to make researchers more productive, but to accelerate the rate of scientific progress itself, targeting everything from rare diseases to climate resilience.

For now, Gemini for Science is both a product line and a statement of intent. It signals that in Google’s view, the next wave of AI is not just about conversational agents and productivity suites, but about embedding AI directly into the machinery of discovery. Whether that vision delivers the breakthroughs its creators hope for will depend on what thousands of human scientists do with these tools – and just as importantly, what they refuse to let the AI decide on its own.