Co-Scientist: Google DeepMind's AI System Now Conducts Full-Fledged Research

Google DeepMind's AI system Co-Scientist has evolved from generating hypotheses to becoming a full-fledged research partner. Built on Gemini models, it now plans experiments, writes code, and controls lab equipment in a closed-loop workflow.

The system was validated across three disciplines: materials science, biology, and computer science. In materials science, Co-Scientist found a safer pathway for producing 2D materials and generated recipes tailored to the lab's equipment.

In biology, it autonomously built an image analysis pipeline that predicted patterns in genetically engineered E. coli colonies with high accuracy. In computer science, it designed an AI architecture called Agent_H that classified incoming queries and generated responses in parallel.

However, when evaluated by board-certified physicians, Agent_H showed a statistically significant advantage only in one category: harm reduction. The researchers acknowledge that benchmark results may not reflect real-world performance.

The system also features reliability modules that push the hallucination rate down to 4 percent. This is a significant improvement over previous systems, which fabricated key results up to 90 percent of the time.