On May 19, 2026, at Google I/O, Google introduced Gemini for Science — a suite of experimental AI tools designed to assist researchers at every major stage of scientific work. The announcement came from James Manyika, Senior Vice President at Google-Alphabet, and Pushmeet Kohli, Vice President at Google DeepMind and Chief Scientist at Google Cloud.
The premise is not subtle: Google believes the same agentic stack that powers software development agents can be pointed at the scientific method itself. Literature review, hypothesis generation, experimental design, and algorithm discovery are all, in some sense, search-and-synthesis problems. Gemini for Science is the attempt to find out how much of that work an agent can carry.
This review covers all three experimental tools, the Science Skills bundle for Antigravity, access and pricing status, and what researchers should actually expect from this release.
What Gemini for Science Is
Gemini for Science is not a single product. It is a collection of three experimental AI tools — each built on an existing Google research system — plus Science Skills, a separate component aimed at professional life science workflows.
The three experimental tools are:
- Literature Insights — built on NotebookLM
- Hypothesis Generation — built on Co-Scientist
- Computational Discovery — built on AlphaEvolve and ERA
Access to the experimental tools opens gradually through Google Labs (labs.google/science). Enterprise access is available via Google Cloud. Science Skills launched May 19 on GitHub and for Google Antigravity users.
Literature Insights
Literature Insights extends NotebookLM’s document synthesis capability into structured scientific research workflows. The tool accepts a research query and a corpus of scientific papers, then does several things a researcher would otherwise do manually:
- Searches and filters relevant literature
- Structures findings into data tables — not just summaries, but extracted, comparable facts across papers
- Generates output artifacts: reports, slide decks, infographics, audio overviews, and video overviews
The output formats are the distinctive part. Standard AI literature tools tend to produce text summaries. Literature Insights produces objects — a researcher can get a presentation-ready slide deck or a narrated audio brief directly, without reformatting.
The underlying technology is NotebookLM, which Google has refined over the past two years. NotebookLM’s strength has always been synthesis across a defined source set rather than open web search. Literature Insights inherits that architecture: it works best when the researcher defines a specific corpus rather than asking it to survey everything. The structured table output is a direct response to one of the most common research bottlenecks — extracting specific, comparable data points from dozens of papers written in different formats.
What it does not yet do: the tool cannot validate the accuracy of extracted claims against original methods sections, and the confidence calibration on table data has not been independently evaluated. For researchers using the output to make experimental decisions, verification against source documents remains essential.
Hypothesis Generation
Hypothesis Generation is built on Co-Scientist, a multi-agent research collaboration system Google has been developing internally. The tool operationalizes a specific part of the scientific method: generating novel, testable hypotheses from an existing body of evidence.
The process works as follows:
- The researcher defines a research challenge — a problem statement, a gap in the literature, an unexplained experimental result
- Co-Scientist runs a multi-agent idea tournament — multiple agent perspectives generate candidate hypotheses independently
- The candidates are debated and evaluated — agents argue for and against each hypothesis, with the evaluation criteria including internal consistency, novelty, and testability
- The tournament produces a ranked set of hypotheses with supporting rationale
The “tournament” framing is meaningful. Unlike a single-agent system that produces one hypothesis, the debate structure surfaces the strongest ideas through competition rather than averaging. It also produces structured rationale for each hypothesis — not just the idea, but the argument for it and the counterarguments against it.
This is the most intellectually ambitious of the three tools, and the hardest to evaluate empirically. Scientific hypothesis quality is domain-dependent, slow to assess, and often not verifiable without running experiments. Google has not published independent evaluation data on hypothesis quality versus researcher-generated hypotheses.
The honest read: this tool is interesting as an idea sparring partner rather than a reliable hypothesis engine. Researchers with strong domain expertise can use it to surface approaches they hadn’t considered and to stress-test their own hypotheses. Researchers who lack domain context to evaluate the output are not in a good position to use it for primary hypothesis generation.
Computational Discovery
Computational Discovery is built on two systems: AlphaEvolve and ERA (Empirical Research Assistance).
AlphaEvolve
AlphaEvolve is a Gemini-powered evolutionary algorithm agent originally developed by Google DeepMind for algorithm discovery. The core mechanism: it generates candidate algorithms, evaluates them against verifiable criteria, iterates through evolutionary selection, and produces optimized solutions that were not in the original search space.
AlphaEvolve’s published results are concrete:
- Discovered a more efficient scheduling algorithm for Google’s data centers, reducing compute waste
- Found a functionally equivalent simplification in hardware accelerator circuit design
- Identified improvements that accelerated the training of the LLM powering AlphaEvolve itself — a recursive improvement result worth noting
- In research benchmarks, improvement ranges from 0.39% to 666.02% depending on problem type and how mature the initial algorithm was
Early enterprise adopters include BASF and Klarna, using AlphaEvolve for materials science and financial modeling problems respectively.
ERA (Empirical Research Assistance)
ERA handles the more standard empirical analysis part of Computational Discovery: statistical analysis, data pipeline automation, experiment parameter search, and result interpretation. ERA is less novel than AlphaEvolve — it’s essentially an agent-assisted data science workflow — but it fills the practical gap between algorithm discovery and running actual experiments.
The combination of AlphaEvolve and ERA is designed to handle the full loop: discover a better algorithm, run empirical tests against it, refine. Whether that loop closes reliably in a research context (as opposed to a well-defined engineering optimization problem) is something early access users will be the first to find out.
Science Skills
Science Skills is the most immediately practical component of the announcement, and the one most likely to see rapid adoption in professional life science research.
Science Skills is a specialized bundle for Google Antigravity (the agentic development platform) that integrates more than 30 major life science databases and tools, including:
- AlphaFold Database — protein structure predictions
- AlphaGenome API — genome interpretation
- UniProt — protein sequence and functional annotation database
- InterPro — protein family and domain classification
Science Skills launched May 19, available both on GitHub and directly for Google Antigravity users.
The practical effect: workflows in structural bioinformatics and genomic analysis that previously required manual pipeline construction — querying multiple databases, normalizing formats, cross-referencing annotations — can be executed as agent tasks in minutes rather than hours. The databases are pre-integrated; the researcher does not need to write the connector code or manage API authentication for each service.
This is infrastructure work, not conceptual novelty. But infrastructure work is often where research teams lose most of their time.
Science Skills is the only component of Gemini for Science that is clearly production-ready. It does not require a research hypothesis, does not generate content that needs independent verification, and has a verifiable output: it either retrieved and integrated data correctly or it did not.
Access and Availability
| Component | Access Path | Status |
|---|---|---|
| Literature Insights | Google Labs (labs.google/science) | Gradual rollout, waitlist |
| Hypothesis Generation | Google Labs | Gradual rollout, waitlist |
| Computational Discovery | Google Labs + Google Cloud (enterprise) | Preview, application required |
| Science Skills | GitHub + Google Antigravity | Available now |
Pricing for the experimental tools has not been announced. Science Skills pricing follows Antigravity’s standard token-based billing.
Who This Is For
Literature Insights is useful for any researcher who spends significant time in literature synthesis — clinical researchers, systematic review authors, policy analysts working with scientific evidence. The structured table output format is the specific value-add over existing tools.
Hypothesis Generation is useful for experienced researchers who want an adversarial brainstorming partner, not for researchers who lack the domain knowledge to evaluate hypothesis quality.
Computational Discovery / AlphaEvolve is currently best suited for well-defined algorithmic optimization problems with verifiable criteria. Materials science, logistics optimization, bioinformatics sequence analysis — problems where the evaluation function is clear. It is less suited to problems where “better” requires human judgment.
Science Skills is useful for any life science team using Google Antigravity. It is the most straightforward value proposition in the suite: pre-integrated access to 30+ authoritative databases through an agent interface.
The Honest Limitations
All three experimental tools are previews. Gradual access via Google Labs means most researchers cannot use Literature Insights, Hypothesis Generation, or Computational Discovery today. The waitlist-and-invite pattern means the early feedback loop is controlled, which is reasonable for research tools but also means real-world performance data outside Google-managed conditions is not yet available.
No pricing for experimental tools. This makes planning difficult for research teams considering it for grant-funded workflows. “No pricing” typically becomes “enterprise negotiated pricing” at scale.
Hypothesis quality has not been independently evaluated. The multi-agent tournament structure is interesting, but scientific hypothesis quality is hard to measure without domain expertise and time. Google has not published comparative data against researcher-generated hypotheses.
AlphaEvolve’s range of “0.39% to 666%" is not a confidence interval — it reflects the enormous variance across problem types. The 666% improvement is likely on a problem where the original algorithm was poorly designed, not typical research conditions. Researchers should expect modest, specific improvements in well-scoped problem domains rather than order-of-magnitude discoveries on average.
Science Skills is Antigravity-dependent. Teams not already using Google Antigravity will need to adopt that platform to access the database bundle. For teams with existing bioinformatics pipelines on other infrastructure, the switching cost is real.
Verdict
Gemini for Science is the most ambitious application of Google’s agentic stack yet — pointing a system built for software development at the scientific method itself. Science Skills is the clearest win: practical, available now, and directly reduces integration overhead for life science teams. Literature Insights has genuine potential for synthesis-heavy research workflows, with the structured table output being a real differentiator. Hypothesis Generation and Computational Discovery / AlphaEvolve are interesting and technically credible, but remain experimental tools that will require domain expert evaluation before drawing conclusions about research value.
The honest summary: Google is serious about this direction, the infrastructure is real, and Science Skills is worth a look for Antigravity users today. The rest is worth watching closely as access opens and independent evaluations emerge.
Rating: 3.5/5 — Science Skills is production-ready; the experimental tools need time and independent data.
ChatForest researches AI tools and platforms. We did not run the experiments described here. All technical details are drawn from Google’s announcements, Google DeepMind publications, and independent reporting from the sources listed below.
Sources: Google I/O 2026 Gemini for Science announcement; Engadget on Gemini for Science; HPCwire on Google AI for Science; AlphaEvolve technical paper; SiliconANGLE on Subquadratic/SubQ launch