Gemini for Science Review: What Google Actually Built for Researchers
Back to the archive
Try ThesisAI

Gemini for Science Review: What Google Actually Built for Researchers

Google's science offering is not one product. It is three experiments, a database bundle and an access form, and the difference between those things decides whether any of it is useful to you this term.

Gemini for Science was announced at Google I/O on 20 May 2026. The framing was ambitious: AI as "a force multiplier for human ingenuity," aimed at biology, chemistry, physics and materials science. The substance is narrower and more interesting than the framing, and it is worth separating what shipped from what was announced.

This post covers the three tools, the Science Skills bundle, how access actually works, what researchers have flagged as weak, and the specific reasons none of it writes a thesis.

The document, not the discovery

ThesisAI drafts a full academic document from one prompt, with inline citations verified against papers retrieved from indexed databases, and exports to LaTeX, Word and BibTeX.

書き始める

What Gemini for Science Actually Is

It is not an app you download and it is not a single subscription. It is a collection of experimental tools in Google Labs, plus a skills bundle that plugs into Google's agentic coding environment, plus a separate API credits programme. Each piece has its own access path.

PieceBuilt onWhat it does
Hypothesis GenerationCo-ScientistDefines a research challenge with you, then runs a multi-agent "idea tournament" where agents generate, debate and evaluate candidate hypotheses
Computational DiscoveryAlphaEvolve and ERAGenerates and scores thousands of code variations in parallel against an optimisation metric, for things like solar forecasting or epidemiological models
Literature InsightsNotebookLMOrganises a corpus of papers into queryable tables with source mapping, and produces reports, infographics and audio or video summaries
Science SkillsGoogle AntigravityA bundle wiring 30+ life science databases (AlphaFold DB, AlphaGenome API, UniProt, InterPro) into agentic workflows for structural bioinformatics and genomics

Google says it worked with more than 100 institutions to validate these systems, including Stanford, Imperial College London and the Francis Crick Institute, with a trusted tester community running from PhD students to Nobel laureates. The published case studies cover antimicrobial resistance work at Cambridge, 2D semiconductor fabrication at Duke, and a theoretical physics group at Rutgers using it to find mathematical errors.

The Part That Is Genuinely Novel

Two of these three tools do something that was not previously available in a packaged form.

The idea tournament

Hypothesis Generation is the most conceptually interesting piece. Rather than asking a model for ideas and getting the statistically obvious ones, it runs multiple agents that propose hypotheses and then argue against each other, with surviving candidates ranked. Google's claim is that outputs come with clickable citations and are "deeply verified."

Treat the verification claim carefully. Adversarial debate between agents filters for hypotheses that are defensible against other instances of the same model. That is a real filter, and it is not the same as a filter for hypotheses that are true, novel, or testable with the equipment in your building.

An idea tournament selects for what survives argument with a model. A literature search selects for what survives argument with reality. What multi-agent debate does and does not buy you

Evolutionary code search

Computational Discovery is the piece with the clearest track record, because AlphaEvolve underneath it has a public history of finding better algorithms. Generating thousands of scored variants in parallel is a genuinely different activity from asking a chatbot to write a function, and for anyone whose research involves optimising a modelling approach, this is the tool worth queueing for.

Literature Insights is the third tool, and it is the one most students will reach for first. It is NotebookLM applied to a scientific corpus: you supply the papers, it builds tables and summaries grounded in them. Useful, and considerably less novel than the other two.

How You Actually Get Access

This is where expectations and reality diverge most, so it is worth being precise.

  • The three Labs tools are a gradual rollout through a request form at labs.google/science. You are joining a queue, not signing up for a product.
  • Science Skills runs inside Google Antigravity, so it assumes you are comfortable in an agentic coding environment.
  • Enterprise access is through Google Cloud, in private preview with organisations like BASF, Bayer Crop Science and the US Department of Energy.
  • Gemini for Research is a separate programme offering Gemini API credits and higher rate limits. Faculty, staff and PhD students at valid academic institutions in supported countries can apply, and applications are reviewed monthly. Google states plainly that credits are granted and removed at its discretion.

Note what is missing from that list: a price, a launch date for general availability, and any path at all for a master's student. If you are writing a taught-degree thesis, the realistic answer today is that you get Literature Insights if your request is approved, and nothing else.

Where It Falls Down

The criticisms of Gemini for Science are not about the ambition. They are about the gap between what a demo shows and what a thesis chapter requires.

Citations still need manual checking. This is the recurring finding across independent reviews of Gemini in research settings: hallucinated citations and factual errors appear often enough that every source has to be verified by hand. "Clickable citation" means the model produced a link, not that a human confirmed the link supports the sentence attached to it. Our post on citing ChatGPT and AI tools covers why a fabricated reference is treated as fabrication rather than a formatting slip.

Transparency is the stated bar and it is not met yet. The considered critique of tools in this category is that suggesting hypotheses, designing tests and summarising papers requires clear sourcing, reproducible outputs and enough visibility for a researcher to trust what they are looking at. An idea tournament that surfaces a ranked list without an auditable trail of why does not clear that bar.

The domain weighting is life sciences. Thirty-plus databases in Science Skills are protein, genome and molecular resources. If your field is education, economics, law or history, the bundle that makes this compelling is not aimed at you.

It is not a writing tool. None of the three tools drafts a structured academic document. Literature Insights produces summaries, tables, infographics and audio overviews. Those are study aids and briefing formats. A thesis chapter is neither.

Gemini for Science vs ThesisAI

These sit at different stages of the research process, which is why the comparison is a division of labour rather than a contest.

Gemini for ScienceThesisAI
Core jobGenerate and test research ideasDraft and structure the document
Stage of workBefore you have resultsWhen you have to write it up
OutputRanked hypotheses, scored code variants, literature tables and summariesA full document with inline citations, up to 80 pages
CitationsModel-generated links, verify each one yourselfRetrieved from indexed databases and verified against the source paper
Reference managersNot supportedZotero and Mendeley import
AccessRequest form, gradual rollout, or enterprise via Google CloudOpen, browser, no waitlist
DisciplinesWeighted to life sciences and computational fieldsAny discipline

Our full Gemini for Science vs ThesisAI comparison goes through this feature by feature. It is also worth reading alongside our Claude Science review, since Anthropic's answer to the same question took a noticeably different shape: a desktop app that runs your analyses and keeps a reproducibility record, rather than a set of hosted experiments.

Who Should Apply

Worth requesting access if: your research involves optimising algorithms or modelling approaches; you work in structural biology, genomics or materials science; you are stuck at the hypothesis stage on a well-defined problem; you are a PhD student or faculty member who can also apply for API credits through Gemini for Research.

Not worth waiting for if: your bottleneck is writing rather than ideas; you need verified citations in a specific style; you are outside the life sciences and computational fields; you are on a taught master's programme; you need something that works this week.

FAQs About Gemini for Science

Is Gemini for Science free?

No price has been published. The Labs tools are experimental and access-gated rather than sold, enterprise access runs through Google Cloud, and the separate Gemini for Research programme grants API credits at Google's discretion to qualified academics.

How is it different from just using Gemini?

The consumer Gemini app gives you Deep Research and Deep Think. Gemini for Science adds the three specialised tools and the Science Skills database bundle on top, with different access paths. If your request has not been approved, you have the consumer product.

Can I use it for my literature review?

Literature Insights can organise a corpus you supply into tables and summaries, which helps you read faster. It does not conduct a systematic review, produce a screening audit trail, or manage a reference library, and its citations need checking one by one.

Do the hypotheses it generates count as my own work?

Your institution decides, and this is worth asking your supervisor in writing before you rely on it. The safe position is that AI-suggested directions are disclosed the same way any other research assistance is, and that you can defend the reasoning behind your hypothesis independently of the tool that surfaced it.

Is Science Skills usable without coding?

Realistically, no. It is designed to work through Google Antigravity, which is an agentic development environment. If you are not writing and running code, this part is not for you.

Which is better, Gemini for Science or Claude Science?

They solve different problems. Claude Science is a desktop app for running and documenting analyses on your own data, with a reproducibility model at its centre. Gemini for Science is a hosted set of experiments for generating and testing research ideas. If you have data and need a defensible record, Claude Science. If you have a question and need directions to explore, Gemini for Science.

Gemini for Science is a serious research programme wrapped in a launch that oversells its availability. The idea tournament and evolutionary code search are real contributions to how research gets done, and neither is something you can rely on for a deadline this term. Apply for access if your work fits. Do not plan a chapter around it, and verify every citation it hands you.

§ End · September 16, 2026
Was this useful?

続行 reading.

All field notes
AI Tools2026

Claude Science Review: Anthropic's Research Workbench

Reproducible artifacts, a reviewer agent, and where it stops.

Read
AI Tools2026

Claude Research vs ThesisAI for Literature Reviews

Both tested on the same research question, side by side.

Read
Methodology2026

How to Write a Research Hypothesis

Formulating hypotheses that are actually testable.

Read

A short note every other Tuesday.

Research writing, citation practice, and the tools we're building. No marketing - unsubscribe with one click.

By subscribing, you accept the 利用規約とポリシー. NO SPAM · 2× MONTHLY