Google researchers published a real-world evaluation of Co-Scientist, a Gemini-based system of cooperating AI agents that proposes experiments, works alongside lab equipment, and writes up results. Trials covered materials science, biology and a computer science task where it designed a method that scored above six leading models on a health benchmark. A double-blind review by 30 domain experts across 450 reviews found its safeguards reduced fabricated content and copied text, and the authors note further experiments are needed to confirm the materials results.
What changed
Co-Scientist had been tested mainly on generating hypotheses in software simulation only.
What it unlocks
Running an AI research assistant that designs experiments, works with lab equipment in the loop, and drafts the resulting paper.
- 30 experts, 450 reviews
- outperformed 6 frontier models on HealthBench
What you need to act on it
- research collaboration with Google
- access to lab equipment
Sources