Google's AI system generates and debates novel scientific hypotheses

Who: Posted on the Google DeepMind blog, authored by the Co-Scientist team — a joint effort across Google DeepMind, Google Research, Google Cloud, and Google Labs. The accompanying research was published in Nature.

What's new: Google DeepMind has released Co-Scientist, a built on top of that acts as a research partner for scientists. Rather than answering a single question, it generates, debates, and ranks scientific hypotheses in a continuous loop. A new public-facing tool called Hypothesis Generation, accessible at labs.google/science, lets individual researchers try it.

How it works: The system runs six specialized agents in parallel. Some propose ideas by reading scientific literature, some critique those ideas like a peer reviewer, and others run an "idea tournament" — borrowing ranking logic from and — to score competing hypotheses against each other. A supervisor agent coordinates the whole coalition, breaking big research goals into parallel workstreams. The system checks every claim against live web search and specialized databases like and , and can call on for protein-structure problems.

The numbers: In one collaboration with Stanford's Gary Peltz, a Co-Scientist-suggested drug candidate blocked 91% of a scarring-linked biological response in lab tests for liver fibrosis, results published in Advanced Science. Researchers at MIT reported that tasks requiring months of literature analysis were compressed to days. Enterprise previews are already running with Daiichi Sankyo, Bayer Crop Science, and the US National Laboratories under a program called the Genesis Mission.

Caveats: All lab validations so far are in life sciences, so the system's usefulness in other fields remains unproven. The blog does not report head-to-head comparisons against a baseline of human researchers working alone, and the published results reflect only cases where Co-Scientist succeeded — negative or inconclusive outcomes are not discussed.