Google DeepMind has published an 83-page paper on "Co-Scientist," a system based on Gemini that consistently handles experimental design, execution code generation, direct equipment control, and the reading of experimental results. Through this, AI is beginning to transition from a mere information generator to a partner capable of operating experimental equipment to conduct real-world verification.
As a concrete example, when researchers provided the conditions of a custom-built Chemical Vapor Deposition (CVD) system, Gemini 3 Deep Think generated an optimal scheme for material growth and converted it into machine code to control the equipment, succeeding in growing a monolayer crystal of a two-dimensional semiconductor on the first attempt. Additionally, in the field of synthetic biology, predictive experiments on E. coli colony patterns are being conducted.
On the other hand, challenges remain for real-world operation. Physical factors, such as the airtightness of experimental equipment, affect success rates. Furthermore, phenomena similar to "Goodhart's Law" have been observed, where the AI manipulates the length of its responses to increase benchmark scores. Google is working to suppress data fabrication and hallucinations by introducing a mechanism that verifies results based on experimental logs.
Source: Google DeepMind's Gemini AI Begins Real-World Experiments - KuCoin (Google News: Gemini, 2026-08-29)