A DeepMind AI Conducts Its Own Experiments And Publishes Research

Google DeepMind announced the benchmark results of their extended Co-Scientist system on August 28, 2026. Co-Scientist, which is based on Gemini 3 Deep Think and Gemini 3 Pro Image, is no longer limited to generating hypotheses. Now it can autonomously plan experiments, code, run code, operate lab instruments and write scientific papers.

1Three Disciplines, Three Results

In the field of materials science, Co-Scientist succeeded in synthesizing three types of semiconductor thin films in one attempt, reducing the time needed to develop recipes from days to minutes. In the field of biology, it accurately forecast the growth pattern of E. coli colonies under different chemical concentration levels, agreeing with experimental observations on three out of four major morphological features. In the field of computer science, the medical AI model that it designed (Agent_H) achieved better performance than six other state-of-the-art models (GPT-5, Claude Opus 5 etc.) on automated medical tasks.

2Reliability: From 90% Fabrication to 4%

The most impressive feature of this new version of Co-Scientist is its significantly reduced data fabrication rate. Without verification modules, the data fabrication rate was 46%, while other similar systems had a 90% data fabrication rate, with 44% of the data being entirely fabricated. With verification modules enabled, Co-Scientist reduced the data fabrication rate to 4% and eliminated all instances of completely fabricated data. It also blocked 98.7% of unsafe research proposals.