Generating novel scientific hypotheses with Co-Scientist

The Nature publication lends legitimacy, but the critical omission in coverage so far is the false-positive rate: how many of Co-Scientist's autonomously generated hypotheses were experimentally tested and failed, and whether that figure appears anywhere in the paper or was quietly left out of the announcement materials.

Google I/O 2026, covered here on May 19th, framed Google's current AI posture as optimizing deployment over raw model innovation. Co-Scientist fits that read precisely: it is Gemini infrastructure being repositioned into a high-prestige vertical (scientific research) rather than a demonstration of new underlying capability. The multi-agent critique-and-refine loop is an application architecture built on existing Gemini models, not a new model class. That distinction matters when evaluating whether this is a research contribution or a product announcement dressed in academic clothing.

Watch whether independent wet-lab groups outside Google publish replication attempts within the next six months. If Co-Scientist's hypotheses show meaningful experimental hit rates in third-party studies, the capability claim holds; if external validation is absent by end of 2026, the Nature paper may reflect curated showcases rather than reliable scientific utility.