Google Deepmind's AI Co-Scientist now plans experiments, runs lab equipment, and writes scientific papers

Details

Google DeepMind first launched Co-Scientist in February 2025 as a Gemini 2.0-based hypothesis generator with documented weaknesses in fact-checking and literature review. The new system, described in a paper by Schmidgall, Zhu et al. (2026), is a closed-loop research agent that carries a scientific question through three phases: ideation, experimentation, and manuscript generation. Verification modules cross-check numerical claims in written outputs against code execution logs to reduce fabricated results. The system was validated across three applications with increasing autonomy. In materials science, Co-Scientist paired with a semi-automated high-temperature furnace to find a safer synthesis route for a 2D material previously produced via hazardous etching, and synthesized three semiconductor thin films on the first try — using Gemini 3 Deep Think to cut recipe development from days to minutes. In biology, it built an image analysis pipeline predicting genetically engineered E. coli colony patterns, with Gemini 3 Pro Image matching unpublished lab results for three out of four scenarios. The computer science track operated fully autonomously to design a medical AI architecture. Caveats are significant: humans still loaded physical samples and precursor materials, fast-mode operation produced smaller and less uniform crystals than optimized recipes, definitive atomic structure confirmation in the materials work is still pending, and the team has not yet verified whether synthesis recipes transfer to other labs.