Google Wants AI to Do Science. The Track Record Is Starting to Matter.
At Google I/O 2026, Demis Hassabis told the audience they are "standing in the foothills of the singularity." That kind of statement usually ends up in a highlight reel and nowhere else. This time, Google came with receipts.
The Gemini for Science Package
Google launched Gemini for Science at I/O, bundling several AI research systems under one brand. The package includes Co-Scientist, a hypothesis-generating AI, and AlphaEvolve, which optimizes algorithms. Neither is publicly available yet. Google is accepting researcher applications for access.
Stanford geneticist Gary Peltz, writing in Nature Medicine, compared using Co-Scientist to "consulting the oracle of Delphi." That is either very high praise or a warning about treating AI output as prophecy. Probably both.
Google Cloud chief scientist Pushmeet Kohli published a piece in Daedalus stating that AI is "moving toward AI that doesn't just facilitate science but begins to do science." The distinction matters. Facilitation means faster literature review and better data visualization. Doing science means forming and testing hypotheses with limited human guidance. The latter is what Gemini for Science is positioned toward.
The AlphaFold Foundation
Google's credibility on AI science does not rest on I/O announcements alone. AlphaFold protein structure predictions have been used by over 3 million researchers worldwide. Isomorphic Labs, the Google subsidiary applying AlphaFold to drug development, closed a $2 billion Series B. Last summer, Google released AlphaGenome for genetics and AlphaEarth Foundations for Earth science.
John Jumper, the AlphaFold Nobel laureate, is now working on AI coding tools rather than science-specific systems. That is either a signal that the science AI problem is solved enough to move on, or that coding tools have better near-term returns. The facts do not say which.
WeatherNext and the Applied Case
WeatherNext provided advance hurricane alerts for Hurricane Melissa's landfall in Jamaica. A new version shipped in November 2025. Weather prediction is a cleaner test bed than drug discovery: the ground truth comes quickly, and the stakes are measurable. The hurricane result is the kind of specific, falsifiable outcome that "AI accelerates science" claims often lack.
OpenAI's Math Result
OpenAI announced this week that one of their models disproved an important mathematics conjecture. The model is described as a general-purpose reasoning system in the vein of GPT-5.5, not a mathematics-specialized tool. A general-purpose model producing novel mathematical results is a different category of claim than domain-specific benchmarks. Details on which conjecture and the verification process have not been provided.
The Shift Being Claimed
Agentic AI research systems are making real research contributions with limited human guidance as of 2026. That is what the facts support. The gap between "limited human guidance" and "doing science" is where the interesting disagreements will happen.
Hassabis's singularity framing, Kohli's Daedalus piece, and the Gemini for Science launch all point in the same direction: Google is positioning itself as the infrastructure layer for AI-driven research. Whether Co-Scientist generates hypotheses that hold up under experimental scrutiny is a question the access program will eventually answer.
The foothills comment might be accurate. Foothills are still pretty far from the summit.
Source: Technologyreview