AI Is Moving From Scientific Assistant to Scientist
What Google I/O 2026 really said
Every year, Google I/O has so many announcements that the truly important ones can get buried. 2026 was no different. Gemini 3.5 Flash and Pro improvements, a new multimodal model called Gemini Omni, an upgrade to the agent IDE Antigravity 2.0, and a list of 100 announcements. Inside all of that, the quietest but deepest announcement was this:
Gemini for Science.
The question it raises is simple. Will AI become a better tool that helps scientists, or will it become an actor that performs science itself? With this announcement, Google clearly took one step toward the latter.
---
The announcement has four components
(1) Literature Insights — Built on NotebookLM. AI handles literature search and analysis, extracting and comparing core findings across thousands of papers.
(2) Hypothesis Generation — Built on Co-Scientist. It reads data and generates hypotheses. Google announced it alongside a validation paper published in Nature on the same day.
(3) Computational Discovery — Built on AlphaEvolve and Empirical Research Assistance (ERA). It mutates code and models to automate computational experiments. ERA has produced results that exceed existing ensemble baselines in epidemiological forecasting tasks such as COVID-19 hospitalization prediction.
(4) Science Skills — The real infrastructure in this announcement. It connects more than 30 life-science databases and tools, including UniProt, AlphaFold Database, AlphaGenome API, and InterPro, so agentic platforms like Antigravity can access them through one workflow. It became available on GitHub and Google Antigravity on May 19, and Antigravity users can try it now.
The first three are experimental tools being rolled out gradually. The fourth, Science Skills, is already usable. That is the most dangerous and most interesting part.
---
The real meaning: the biggest inflection point since AlphaFold
AlphaFold 1 in 2018, AlphaFold 2 in 2020, AlphaFold 3 in 2024. That line of progress was about building specialized tools where AI could solve specific scientific problems extremely well, such as protein structure prediction. AlphaFold's success was so overwhelming that the 2024 Nobel Prize in Chemistry went to John Jumper and Demis Hassabis for the work behind it.
But the signal in 2026 is different. Reports say John Jumper has also been pulled into Google's AI coding efforts, and Google's official framing now points beyond "narrow, specialized models" toward general agents that support researchers across scientific fields.
The next battlefield is not simply building another single-purpose model like AlphaFold. It is building agents that autonomously combine models, databases, and tools to perform science. Gemini for Science is an implementation of that judgment. AlphaFold becomes a component. The new game is how an agent combines those components.
That is why this is different from a normal tooling upgrade. The paradigm itself is changing.
---
What becomes possible in the near future
If this paradigm takes hold, several things become possible.
Drug candidate discovery accelerates sharply. Docking 1,000 molecules against a single protein and automatically hypothesizing the binding mechanisms of top candidates becomes a minutes-level workflow. Google's public AK2 rare-disease example already points in that direction. An analysis that normally takes hours was completed in minutes and produced new possible disease-mechanism insights.
Rare-disease mechanisms become easier to investigate. A single researcher can combine genomic data, structural data, and literature analysis. Work that previously required a large consortium to hold onto for a long time becomes more accessible.
The expansion goes beyond pharma. This is already spreading into other fields. BASF is using AlphaEvolve for supply-chain optimization, and Bayer Crop Science has adopted Co-Scientist for agricultural research. Daiichi Sankyo in Japan and U.S. Department of Energy national laboratories are also private-preview partners.
The researcher's role splits. Humans own problem definition and interpretation. Agents take on data processing, literature exploration, and candidate generation. The labor structure of doctoral research itself may be reorganized.
---
What is missing
This is where the important part begins. Gemini for Science is a tool for making analyses that used to take a long time happen faster. It is not yet a tool for making previously invisible things visible. There is a difference between higher throughput and deeper insight.
Most AI-for-science tools in the industry look similar right now: more data, faster processing, larger models. But almost nobody is addressing the meta-principle behind why a certain molecule fits a certain site. Tools are good at producing binding scores and ipTM values. They are not yet providing a conceptual lens for why that site is empty, and why that molecule fits there in particular.
That lens may be the next inflection point. Just as AlphaFold offered the lens of "predicting structure," the next tool needs to offer the lens of "why there is an empty place there." That is the space Google has not yet filled.
---
Closing
The real announcement at Google I/O 2026 was not a model upgrade. It was a redefinition of how science is conducted. AI is no longer staying only in the role of assistant to scientists. It is moving toward the role of an actor that performs science. AlphaFold becomes one component inside that system.
Two questions remain. First, if agents perform science, where does the scientist's role move? Second, if every tool enters a throughput race, where will truly new discovery come from?
Time will answer the first. The second will probably come not from a tool, but from a lens.
---
Sources: [Google Gemini for Science announcement](https://blog.google/innovation-and-ai/technology/research/gemini-for-science-io-2026/), [Google Research ERA announcement](https://research.google/blog/empirical-research-assistance-era-from-nature-publication-to-catalyzing-computational-discovery/), [Google DeepMind AlphaEvolve update](https://deepmind.google/blog/alphaevolve-impact/), Google I/O 2026 materials, Nature Co-Scientist and ERA papers.