Google unveils Gemini for Science

Google is turning the scientific method into an autonomous workflow.

At I/O, the tech giant launched Gemini for Science, a new suite of AI tools and prototypes on Google Labs designed to take on ideation, empirical coding, and literature review. Alongside the launch, Google DeepMind and Google Research published two related milestone papers directly in Nature.

Instead of launching another narrow model, Google is betting on general-purpose agentic systems to accelerate research.

Multi-agent tournaments and thousands of code iterations

The most ambitious piece of the rollout is Hypothesis Generation, built on a system called Co-Scientist.

Instead of prompting a chatbot and hoping it doesn't hallucinate, Co-Scientist sets up an "idea tournament." Specialized Gemini agents propose concepts, cluster them for diversity, act as virtual peer reviewers to critique flaws, and iteratively evolve the best ideas through simulated debates. Crucially, most of the system's compute goes into verifying claims against databases like UniProt and literature, backing every output with clickable citations.

The second tool, Computational Discovery, tackles the experimental bottleneck. Powered by AlphaEvolve and ERA (Empirical Research Assistance), it generates and tests thousands of code variants in parallel using tree search. Google tested it across genomics, epidemiology, and solar forecasting, matching expert human coding performance.

Rounding out the suite is Literature Insights, built on NotebookLM, which organizes research papers into custom, side-by-side comparison tables and generates slides, reports, and audio summaries.

Google also rolled out Science Skills, bundling over 30 databases into platforms like Google Antigravity. In early internal tests, researchers compressed structural bioinformatics workflows from hours to minutes, uncovering novel insights into a rare genetic disease tied to the AK2 gene.

Enterprise players like BASF, Bayer Crop Science, Daiichi Sankyo, and Klarna are already testing these capabilities via Google Cloud.

Why it matters

Google is doing something very smart here: it is moving past point solutions and building an operating system for scientific inquiry.

Tools like AlphaFold proved AI could solve deeply isolated problems, but the day-to-day bottleneck for scientists has always been the friction between reading literature, formulating testable hypotheses, and writing simulation code. By chaining NotebookLM, multi-agent debate tournaments, and automated code generation together, Google is attacking the entire research lifecycle.

For independent researchers and smaller labs, this is a massive equalizer. Hypotheses that once required an army of postdocs to screen and script can now be generated and stress-tested in an afternoon. If you have an institutional Google account, registering on Google Labs is an absolute no-brainer.

However, there is an inherent risk worth watching. Multi-agent tournaments score hypotheses based on consensus and existing literature. Scientific revolutions often start with ideas that sound completely absurd to current consensus. We have to wonder whether an AI review board might prematurely filter out the next radical breakthrough because it does not fit the historical training data.

Still, by embedding these agents directly into enterprise pipelines and academic conferences, Google is setting a high bar that raw chat interfaces simply cannot match.